Salgado R. (1988) PH D Thesis Vol I

UNIVERSITY OF NEWCASTLE-UPON-TYNE
CIVIL ENGINEERING DEPARTMENT
COMPUTER MODELLING OF WATER SUPPLY DISTRIBUTION NETWORKS
USING THE GRADIENT METHOD.
by
Ruben 0. Salgado-Castro
VOLUME ONE
**********
NEWCASTLE UNIVERSITY LIBRARY
COB 22970 8
-4..12S-i s L.2k
Thesis submitted in fulfilment of the requirements for the
degree of Doctor of Philosophy in Civil Engineering.
November, 1988.
ACKNOWLEDGEMENTS
I would like to acknowledge the constant and enthusiastic
support given to this work by Professor P. E. O'Connell, Head of
the Water Resources Division of the Civil Engineering Department,
University of Newcastle-upon-Tyne (United Kingdom). He supervised
this thesis and contributed with many discussions and ideas; his
permanent encouragement has been invaluable through all this
time.
I am indebted to Professor E. Todini, Head of the Institute of
Hydraulic Constructions of the University of Bologna (Italy). He
created the gradient method and contributed with his advice for
its further improvement.
It has been a privilege and a pleasure to work with Professors
O'Connell and Todini.
Some institutions made possible my stay in Newcastle
University and I am most grateful to them: The Overseas
Development Administration provided the funds for a three years
scholarship; The British Council managed this scholarship and
also provided extra support for attending Conferences and
courses. I am particularly grateful for the support provided by
the chilean office of The British Council and of that received
from the staff of Newcastle Regional Office. The University of La
Serena (Chile), granted me with a leave of absence which I fully
appreciate and my colleagues at the Civil Engineering Department
have shared the extra burden due to my absence.

The Civil Engineering Department of the University of
Newcastle-upon-Tyne provided a stimulating environment and
the facilities for carrying out this work and I am grateful for
that too. Mr. R. Mackay, lecturer in Hydrology, kindly allowed me
to use his Kriging suite of programs.
My famil y and friends from Chile constantly supported me and I
truly appreciate their encouragement.

ABSTRACT
A water distribution network analysis method known as the
gradient method, due to Todini (1979), has been generalised and
subjected to an extensive program of testing and evaluation. The
method has been extended to include pumps and some pressure
regulating valves, and an original physically-based method has
been proposed for modelling the latter devices. Also, a
generalised version of the algorithm which considers the nodal
demands as a linear function of the pressures has been
introduced. The gradient method has been tested with numerous
examples, showing remarkable robustness and convergence speed
when compared with the most efficient traditional methods. The
gradient algorithm does not break down with disconnected

networks.
The performance of the gradient algorithm when using seven
different linear solvers, including direct and iterative methods,
has been investigated. A multifrontal linear solver has been
identified as the most efficient method when enough computer
memory is available (routine MA27 of the Harwell Library); if
storage is limited, a preconditioned (modified) conjugate
gradient method is the recommended linear solver. A good
compromise between memory and speed is represented by the one-way

dissection method of George and Liu (1981).
An automatic calibration algorithm has been proposed which
estimates the true pipe resistance parameters, based on the
estimation of the unmeasured piezometric heads and unmeasured
flows. For the piezometric head estimation, three different

methods have been proposed and compared: one based on Kriging,
another based on bi-cubic splines and a third based on an
original deterministic one-dimensional interpolation procedure.
The latter producing the closest estimates with respect to the
true values. For the estimation of the unmeasured flows, the raw
(un-calibrated) network model itself is used, based on initial
estimates of the pipe roughnesses, leading to an iterative

procedure. The results of using the proposed calibration
algorithm with a set of test examples show that the unmeasured
flow estimation needs further work and an alternative approach
has been suggested, which, hopefully, would lead to improvements

both in the flow estimation and in the estimation of the true
roughnesses.
TABLE OF CONTENTS
List of Figures viii

List of Tables xii
List of main Symbols xvi
VOLUME ONE
**********
CHAPTER ONE
INTRODUCTION 1
1.1. Computer modelling of water distribution systems. 1

1.2. Network analysis: need for a reliable algorithm
for design and operational purposes 10
1.3. Need for an automatic calibration algorithm for
the network model 12
1.4. Simulation of water distribution networks 13
1.5. Thesis structure 14
1.6. Summary of main achievements 16
CHAPTER TWO
REVIEW OF THE MAIN EXISTING METHODS
OF NETWORK ANALYSIS 18 4,n'-`
2.1. Introduction 18
2.2. Hardy Cross (1936) methods 20
2.2.1. Method of balancing heads 22
2.2.2. Method of balancing flows 26
2.3. Newton-Raphson-based methods 33
2.3.1. Introduction 33
2.3.2. The mathematical background of the Newton-
Raphson method 34
2.3.3. Martin and Peters formulation (1963) 39
2.3.4. Shamir and Howard nodal formulation (1968) 44
2.3.5. The Newton-Cross method of Liu (1969) 49
2.3.6. Epp and Fowler loop formulation (1970) 51
2.3.7. The mesh-nodal approach of Hamam and Brameller
(1971) 56
2.3.8. Other formulations 65
2.4. The linear theory method 68
2.4.1. Wood and Charles formulation (1972) 68
2.4.2. Nodal formulation of Isaacs and Mills (1980) 72
2.4.3. The gradient-based formulation of Wood (1981) 75
2.5. Other approaches 78
2.5.1. Optimisation-based methods 78
2.5.2. Methods based on unsteady state analysis 79
2.5.3. A method based on a finite element approach 80
2.5.4. The gradient method of Todini (1979) 80
2.6. Comparison of the performance of some of the
existing methods: need for a more reliable
algorithm 88
CHAPTER THREE
GENERALISATION OF THE GRADIENT METHOD
TO INCLUDE PUMPS 99
3.2. Traditional formulation of the network analysis
problem 99
3.3. Minimum power dissipation formulation: necessary
conditions 100
3.4. Todini's gradient method 102
3.5. A general head loss/flow model to include pumps
and valves in the gradient method 104
3.6. Extension of the gradient method to include pumps:
derivation of the recursive scheme for the
extended method 107
3.7. On the existence and uniqueness of the network
analysis solution 113
3.8. Other gradient formulations 121
3.9. Comparison of the gradient method with some of
the existing methods of network analysis 130
3.9.2. The computer programs used in the comparison 130
3.9.3. The set of test examples 132
3.9.4. Comparison of the algorithms for the
simultaneous path adjustment, linear theory
and gradient methods 137
3.10. Concluding remarks 141
CHAPTER FOUR
EXTENSION OF THE GRADIENT METHOD TO INCLUDE
REGULATING VALVES:
A NEW PHYSICALLY-BASED APPROACH 143
4.2. Description of some control valves 145

4.2.1. Check valves 145
4.2.2. Pressure controlling devices: main features 147
4.3. Existin g models for pressure controlling devices . 161

4•4. Proposed model for pressure controlling devices 172
4.4.1. Needfbr an alternative model 172
4.4.2. Development of the computer program 173
4.5. Comparison between the results given by the

proposed model and some examples found in the
literature 179
4.5.1. Example JEPPO, from Jeppson and Davis (1976) 179
4.5.2. Example JEPP1, from Jeppson (1976) 183
4.5.5. Example COLLINS, from Collins (1980) 189
4.6. Concluding remarks 193
CHAPTER FIVE
EFFICIENT COMPUTER PROGRAM IMPLEMENTATION
OF THE GRADIENT METHOD FOR
WATER SUPPLY DISTRIBUTION NETWORKS 196
5.1. Automatic generation of initial flow distribution. 196
5.2. Solution of the linear system of equations

generated by the gradient method 199
5.2.1. General overview of direct sparse linear
solvers 199
5.2.2. General overview of iterative methods for
solving linear systems of equations 203
5.2.3. Comparison of the performance of the
gradient method with different linear
solvers 205
5.2.4. Selection of the most efficient method 214
5.3. Concludin g remarks 216

CHAPTER SIX
CALIBRATION OF WATER SUPPLY DISTRIBUTION SYSTEMS:
A NEW COMPUTER-BASED EXPLICIT METHOD 217
6.2. Uncertainties in water distribution modelling 219

6.2.1. Mathematical model for the flow/head loss
phenomena 219
6.2.2. Water consumption 223
6.2.3. Effect of network reduction 224
6.2.4. Influence of network pressure 225
6.2.5. Other factors 225
6.3. Measurements available in a real network and their

accuracy 227
6.4. Problem formulation 229
6.5. Review of existing methods for water distribution
network calibration 232
6.6. Deterministic versus stochastic approach to the
solution of the calibration problem 245
6.7. The rationale of the proposed calibration method 247
6.8. An iterative approach for the solution of the
static calibration problem 252
6.9. Estimating the piezometric heads 255
6.9.2. Estimating the piezometric heads using
Kriging 257
6.9.3. Estimating the piezometric heads using a
deterministic one-dimensional interpolation
method 265
6.9.4. Estimating the piezometric heads using
bi-cubic splines 272
6.10. Examples of applications of the proposed
calibration method 277
6.10.1. Description of the networks used as examples 277
6.10.2. Defining the "true" network characteristics 284
6.10.3. Defining the measurements for the network
examples 284
6.10.4. Defining the initially assumed roughnesses
for the network examples 287
6.10.5. Defining a perturbed set of nodal demands 292
6.11. Comparison of the calibration results using
different head estimation techniques 294
6.11.1. Main objectives 294
6.11.2. Study cases 294
6.11.3. Calibration exercise 297
6.11.4. Results of the calibration 299
6.11.5. Summary of the results 341
6.12. Concluding remarks 346 t,—
CHAPTER SEVEN
FURTHER EXTENSIONS OF THE GRADIENT METHOD 350

7.2. Extended period simulation version of the gradient
method 350
7.3. Extending the gradient method for pressure-
dependent nodal demands 356
CHAPTER EIGHT
SUMMARY, CONCLUSIONS AND FURTHER WORK 366
8.1. Summary 366
8.2. Conclusions 367
8.3. Further work 3704---

8.2.1. Gradient method for network analysis 371
8.2.2. Calibration 372—
REFERENCES
REFERENCES 375
VOLUME TWO
**********
APPENDIX A
EFFICIENT SOLUTION OF LINEAR SYSTEMS OF EQUATIONS
IN THE CONTEXT OF THE GRADIENT METHOD FOR THE
ANALYSIS OF WATER SUPPLY DISTRIBUTION NETWORKS
A.1. Introduction ' A-1

A.1.1. Need for efficient linear solvers in water
distribution network analysis A-1
A.1.2. Different classes of linear problems A-2
A.1.3. Direct methods and iterative methods for
solving determined systems of linear
equations A-10
A.2. Review of direct methods for the solution of

general dense linear systems of equations A-13
A.2.1. Gaussian elimination A-13
A.2.2. Gauss-Jordan elimination A-28
A.2.3. Matrix factorization methods A-31
A.2.4. Comparison of the different algorithms for
the direct solution of general dense linear
systems A-45
A.2.5. Error analysis in the solution of linear
systems A-48
A.3. Review of sparse direct methods for the solution of

linear systems of equations A-57
A.3.1. Data structures A-57
A.3.2. Fill-in A-62
A.3.3. Sparse methods for banded matrices A-69
A.3.4. The envelope ("skyline") method A-71
A.3.5. Minimum degree algorithms A-77
A.3.6. Quotient tree algorithms A-89
A.3.7. One-way dissection methods A-104
A.3.8. Nested dissection methods A-108
A.3.9. Frontal and multifrontal methods A-110
A 4. Review of iterative methods for the solution of

linear systems of equations A-115
A.4.1. Introduction A-115
A.4.2. Jacobi's method (method of simultaneous
displacements) A-117
A.4.3. Gauss-Seidel method (method of successive
displacements) A-119
A.4.4. Successive over (or under) relaxation method A-120
A.4.5. Relationship between the previous methods A-120
A.4.6. Convergence conditions for the previous
methods A-124
vi
A.4.7. Standard conjugate gradient method for the
solution of linear systems of equations.
(Hestenes and Stiefel, 1952) A-127
A.4.8. Preconditioning A-133
A.4.9. Preconditioned (modified) conjugate gradient
method for the solution of linear systems A-142
APPENDIX B
DERIVATION OF THE KRIGING ESTIMATOR EQUATIONS
B.1. Introduction B-1
B.2. Statistical inference B-5
B.3. Estimation via Kriging B-16

B.3.1. Kriging under stationary conditions B-16
B.3.2. Relaxing the second-order stationarity
condition: the intrinsic hypothesis B-25
B.3.3. Introducing uncertainty in the measurement
data B-30
B.3.4. Universal Kriging and k-th order random
functions (k-IRF) B-32
APPENDIX C
DERIVATION OF THE BI-CUBIC SPLINES APPROXIMATION EQUATIONS
C.1. Introducing cubic splines C-1
C.3. Data fitting using one-dimensional cubic splines C-8
C.4. Data fitting using bi-cubic B-splines C-13
C.5. The statistical problem in splines fitting C-19
APPENDIX D
NUMERICAL RESULTS OF THE CALIBRATION EXERCISE D-1

LIST OF FIGURES
VOLUME ONE
**********
FIGURE PAGE
Fig. 1.1. The water distribution network model and

its main components 9
Fig. 1.2. Structure of the extended period simulation
program 14
Fig. 2.1. Representation of a network graph and its
tree and co-tree 60
Fig. 3.1. Flowchart of main steps in gradient method 112
Fig. 3.2. Basic network for examples A, B, C and D,
from Chin et al. (1978) 134
Fig. 3.3. Network for example E, from de Neufville
and Hester (1969) 136
Fig. 3.4. Network for example F, from Barlow and
Markland (1969) 136
Fig. 4.1. Pressure Reducing Valve. Design and
operational features, from Glenfield &
Kennedy Ltd., Pub. 215/R3 148
Fig 4.2. Pressure Sustaining Valve. Design and
operational features, from Glenfield &
Kennedy Ltd., Pub. 215/R3 149
Fig. 4.3. Feasible head loss/flow region for a
pressure reducing valve. (For a variable
flow) 154
Fig. 4.4. Relationship between inlet and outlet heads
in a pressure reducing valve. (For a constant
flow) 154
Fig. 4.5. Relationship between inlet and outlet head,
and aij in a pressure reducing valve. (For
constant flow) 157
in a pressure reducing valve. (For a
variable flow) 157
Fig. 4.7. Response of the network at the downstream
end of a PRV, to variations in the
resistance characteristic coefficient aij . 159
in a pressure sustaining valve. (For
constant flow) 162
Fig. 4.9. Response of the network at the upstream end
of a PSV, to variations in the resistance
characteristic coefficient a i i 162
Fig. 4.10. Modelling PRV's as pseudo-reservoirs 164
Fig. 4.11. Correction of the resistance characteristic
parameter (a), for a PRV, based on the
lumped curve 177
Fig. 4.12. Interaction between pressure regulating
device model and the standard network
analysis algorithm 178
Fig. 4.13. Example JEPPO, from Jeppson and Davis (1976) 181
Fig. 4.14. Example JEPP1, from Jeppson (1976), page 86. 183
viii
FIGURE PAGE
Fig. 4.15. Example JEPP12, from Jeppson (1976), page

110 187
Fig. 4.16. Exam p le JEPP13, from Jeppson (1976), page
112 187
Fig. 4.17. Example COLLINS, from Collins (1980) 190
Fig. 4.18. Other pressure regulating valves, from
Glenfield & Kennedy Ltd., Pub. 215/R3 195
Fig. 5.1. Target "area" for band and envelope methods 201
Fig. 5.2. Comparison of execution times for the
gradient method, with different linear
solvers 211
Fig 5.3. Comparison of storage requirements of the
different linear solvers 212
Fig. 6.1. Basic iterative static calibration procedure,
computing corrected Hazen-Williams roughness
vector 254
Fig. 6.2. Estimation of the piezometric heads using
Kriging 264
Fig. 6.3. Estimation of piezometric heads using the
one-dimensional interpolation scheme 268
Fig. 6.4. Minimum head loss path of nodes, before the
head interpolation has been carried out 270
Fig. 6.5. Minimum head loss path of nodes after the
head interpolation has been carried out 270
Fig. 6.6. Estimation of piezometric heads using bi-
cubic splines 276
Fig. 6.7. Network for Examples A, B and C 278
Fig. 6.8. Network for Example D 281
Fig. 6.9. Network for Example E 282
Fig. 6.10. Network for Example F 283
Fig. 6.11. The performance index Rx 2 301
Fig. 6.12. Comparison between the initial residuals and
those obtained after the one-dimensional
interpolation procedure has been applied,
for example A 303
Fig. 6.13. Comparison of head estimates for 4 transverse
sections of Network E 307
Fig. 6.14. Comparison of calibrated piezometric heads
for 4 transverse sections of Network E,
Case I 312
Fig. 6.15. Comparison of calibrated flows for 4
transverse sections of Network E, Case I 321
Fig. 6.16. Comparison of calibrated C's. Network E,
Case I 331
Fig. 7.1. Main flow chart of extended period
simulation program 352
Fig. 7.2. The effect of pressure on leakages. From:
National Water Council (1980) 357
Fig. 7.3. Relationship between actual demand,
pressure, adopted linear model and
constant demand 357
VOLUME TWO
**********
FIGURE PAGE
Fig. A.1. Different linear systems of equations A-4

Fig. A.2. Arrow shaped linear system A-63
Fig. A.3. Cholesky factor of arrow shaped matrix
(without reordering) A-63
Fig. A.4. Reordered arrow shaped linear system A-63
Fig. A.5. Cholesky factor of reordered arrow-shaped
linear system A-64
Fig. A.6. Undirected graph corresponding to symmetric
matrix given by equation (98) A-65
Fig. A.7. Permutation matrix for interchanging rows 3
and 4 A-66
Fig. A.8. Reordered matrix [eq.(98)] and its
associated graph A-66
Fig. A.9. Diagonal storage scheme for banded matrices A-69
Fig. A.10 Minimum band ordering for arrow shaped
linear system A-70
Fig. A.11. Two examples of Cuthill-McKee orderings for
reducing bandwidth, from George (1981) A-71
Fig. A.12. Skyline representation A-72
Fig. A.13. Graph for level structure example, from
George and Liu (1981) A-74
Fig. A.14. Level structure rooted at x5, for graph
shown in Fig. A.13, from George and Liu
(1981) A-74
Fig. A.15. Algorithm for finding a pseudo-peripheral
node, from George and Liu (1981) A-76
Fig. A.16. Example of application of minimum degree
algorithm, from George (1981) A-81
Fig. A.17. Elimination graphs representing the process
of creation of fill-in A-82
Fig. A.18. The filled graph of F = L + L T A-83
Fi g . A.19. Filled (reordered) matrix A-83
Fig. A.20. Sequence of quotient graphs r i used for
finding the reachable set of nodes in the
elimination process A-87
Fig. A.21. An example of a monotonely ordered tree
(rooted at node 8) A-96
Fig. A.22. Matrix corresponding to the monotonely
ordered tree shown in Fig. A.21 A-97
Fig. A.23. Network example for quotient tree methods A-98
Fig. A.24. Quotient tree corresponding to the tree of
Fig. A.23 A-100
Fig. A.25. Matrix corresponding to the quotient tree
of Fig. A.24 A-100
Fig. A.26. Level structure rooted at node "a" and its
corresponding quotient tree A-101
Fig. A.27. Matrix corresponding to refined quotient
tree of Fig. A.26. b) A-102
Fig. A.28. Level structure rooted at node "t" and its
corresponding quotient tree A-103
Fi g . A.29. Matrix corresponding to refined quotient
tree of Fig. A.28. b) A-103
Fig. A.30. Rectangular grid partitioned with 2
dissectors A-105
FIGURE PAGE
Fig. A.31. Example of a 40 nodes rectangular shaped
graph, partitioned using one-way dissection A-106
Fi g . A.32. Matrix structure corresponding to graph of
Fig. A.31 A-107
Fig. A.33. Example of a 40 nodes rectangular-shaped
graph, partitioned using nested dissection A-109
Fig. A.34. Matrix structure corresponding to graph of
Fig. A.33 A-109
Fi g . A.35. Finite element definition and ordering for a
frontal solution scheme, from Livesley
(1983) A-112
Fi g . A.36. The elimination sequence in a frontal
solution scheme, from Livesley (1983) A-113
Fi g . A.37. Geometric interpretation of the convergence
of the conjugate gradient algorithm for n=2
and n=3, from Gambolatti and Perdon (1984),
solving A h = b A-134
Fi g . B.1. Estimated semi-variogram B-9
Fi g . B.2. Nugget effect and corresponding true semi-
variogram B-13
Fi g . B.3. Relationship between covariance and semi-
variogram B-17
Fi g . C.1. Representation of a B-spline C-4
Fi g . C.2. B-splines involved in the spline function
for the interval ( Xi_i,Xi) C-5
Fi g . C.3. Extended set of B-splines: interior and
exterior knots C-6
Fi g . C.4. Rectangular subspace R for the independent
variables x and y, divided into panels Rij
by knots Xi i1,2,.. .h+1 and Pi j=1,2,...k+1 C-15
LIST OF TABLES
VOLUME ONE
**********
TABLE PAGE
Table 2.1. Results of the comparison of 60 small water
distribution networks (less than 100 pipes)
from Wood (1981a) 89
Table 2.2. Comparison of network analysis methods in
terms of global and relative CPU times, and
number of iterations, from Carpentier et
al. (1985) 92
Table 3.1. Summary of the performance of the programs
LOOP, LT and GRAD with test examples 138
Table 4.1. Summary of effects of PRV's in different
formulations of network analysis equations
[following proposition of Jeppson (1976)
and Jeppson and Davis (1976)] 165
Table 4.2. Comparison of results reported in the
literature and our results for networks
containing pressure control devices 180
Table 4.3. Comparison of results presented by Jeppson
and Davis (1976) and our results, for
example JEPPO 182
Table 4.4. Comparison of the results given by Jeppson
(1976) and our results for example JEPP1 184
Table 4 .5. Comparison of the results given by Jeppson
(1976) and our results, for example JEPP12. 186
Table 4.6. Comparison of the results given by Jeppson
(1976) and our results, for example JEPP13 188
Table 4 .7. Network solutions for example COLLINS, from
Collins (1980) 190
Table 4.8. Operational cases for example COLLINS,
based on all the possible combinations of
pump operating modes 191
Table 4.9. Results of running our program for the
operational modes possible in example
COLLINS 192
Table 5.1. Main data corresponding to examples used
in the comparison between different linear
solvers in the gradient method for pipe
network analysis 207
Table 5.2. Comparison of execution times and number of
iterations required by different linear
solvers in the gradient method for pipe
network analysis 208
Table 5.3. Comparison of storage requirements (Bytes)
for different linear solvers used with the
gradient method for pipe network analysis 210
Table 6.1. Network data for Examples A, B and C 279
Table 6.2. Measurement data for Examples A, B and C 285
Table 6.3. Measurement data for Examples D, E and F 285
TABLE PAGE
Table 6.4. Initial values of C (Hazen-Williams) used
for testing the calibration algorithm.For
Examples A and B and low and high
uncertainty 288
for testing the calibration algorithm in
Example C 289
for testing the calibration algorithm. For
Examples D, E and F, and for low and high
uncertainty 290
Table 6.7. Modified nodal demands (us), used to study
the impact of bad demand estimation on the
calibration algorithm. Examples A, B and C 292
Table 6.8. Modified nodal demands (us), used to study
the impact of bad demand estimation on the
calibration algorithm. Examples D, E and F. 293
Table 6.9. Comparison between estimated piezometric
heads and its residuals, for example A 303
Table 6.10. Summary of the comparison between different
piezometric head estimation procedures 304
Table 6.11. Estimated piezometric heads for Cases I, II,
III, IV and V, using the one-dimensional
interpolation procedure 305
Table 6.12. Performance indices for the different head
estimation procedures, for examples A, B,
C, D, E and F 308
Table 6.13 Summary of the comparison for the
calibrated piezometric heads: average
heads 313
calibrated piezometric heads: variance of
the heads 314
calibrated piezometric heads: maximum
residual 315
Table 6.16. Summary of the comparison for the
calibrated piezometric heads: variance of
the absolute value of the residuals 316
calibrated flows: average flows 322
calibrated flows: variance of the flows 323
calibrated flows: maximum residual 324
calibrated flows: variance of the absolute
value of the residuals 325
calibrated C's: average C's 332
calibrated C's: variance of the C's 333
calibrated C's: maximum residual 334
TABLE PAGE

calibrated C's: variance of the absolute
value of the residuals 335
Table 6.25. Summary of the ratios between average
calibrated roughnesses and average true
roughnesses 338
Table 6.26. Summary of the "global success" index 341
VOLUME TWO
**********
TABLE PAGE
Table A.1. Comparison of the algorithms used in the

direct solution of dense linear systems of
equations A-46
Table B.1. Computation of the experimental semi-
variogram B-8
Table B.2. Basic analytical models for the experimental
semi-variogram B-14
Table B.3. Possible polynomial models for generalised
covariances, from Delhomme (1978) B-37
Table D.1. Summary of the comparison between true,
initial and calibrated piezometric heads.
Example A D-1
Example B D-2
Example C D-3
Example C D-4
Example E D-5
Example F D-6
Table D.7. Summary of the performance of the
calibrated heads. Example A D-7
calibrated heads. Example B D-8
calibrated heads. Example C D-9
calibrated heads. Example D D-10
Table D.11 Summary of the performance of the
calibrated heads. Example E D-11
calibrated heads. Example F D-12
xiv
TABLE PAGE

initial and calibrated flows. Example A • pp D-13
initial and calibrated flows. Example B • D-14
initial and calibrated flows. Example C • PO D-15
Table D.16 Summary of the comparison between true,
initial and calibrated flows. Example D D-16
Table D.17 Summary of the comparison between true,
initial and calibrated flows. Example E • PO D-17
initial and calibrated flows. Example F D-18
calibrated flows. Example A D-19
calibrated flows. Example B D-20
calibrated flows. Example C D-21
calibrated flows. Example D D-22
calibrated flows. Example E D-23
calibrated flows. Example F D-24
initial and calibrated C's. Example A D-25
initial and calibrated C's. Example B D-26
initial and calibrated C's. Example C D-27
initial and calibrated C's. Example D D-28
initial and calibrated C's. Example E D-29
initial and calibrated C's. Example F D-30
calibrated C's. Example A D-31
calibrated C's. Example B D-32
calibrated C's. Example C D-33
calibrated C's. Example D D-34
calibrated C's. Example E D-35
calibrated C's. Example F D-36
LIST OF MAIN SYMBOLS
NETWORK ANALYSIS
Symbol Description
NN Total number of nodes in the network.
NP Number of links in the network (pipes, pumps,
valves, etc.).
NL Number of loops in the network (natural loops).
NS Number of sources in the network (known piezometric
head nodes), typically reservoirs.
Flowrate, (NP*1) column vector.
Qi Flow corresponding to link "i", i-th component of
vector Q.
Flow in link joining nodes "i" and "j".

6Q iiFlow correction in pipe joining nodes "i" and "j".
eQ Flow correction in loop "i". Also used as flow (mass)
imbalance at node "i".

dg Pipe flow differential, (NP*1) column vector.
Nodal piezometric head, ((NN-NS)*1) column vector.

Hi Piezometric head corresponding to node "i", i-th
component of vector H.
&Hi Piezometric head correction at node "i" (in nodal
approach).
HO (NS*1) column vector of fixed (known) piezometric heads
for the source nodes.

dH Nodal piezometric head differential, ((NN-NS)*1) column
vector.
Head loss, (NP*1) column vector.

Symbol Description
hi Head loss corresponding to link "i", i-th component of
vector h.
Head loss in link joining nodes "i" and "j":
ij = H i - H j I
Ih
Nodal consumption (demand), ((NN-NS)*1) column vector.
qi Nodal consumption corresponding to node "i", i-th
component of vector g.
Sa Flow (mass) nodal imbalance, tNN-NS)*1) column vector.
Flow exponent in head loss/flow relationship.
Pipe resistance (NP*1) column vector.

Kij Resistance corresponding to pipe joining nodes and
"j".
I h-IJ- = K-13- Q-1.3

-n
Pipe conductance, (NP*1) column vector.

R 1J
- Conductance corresponding to pipe joining nodes "i" and
Qij = R ij nijl/n
Refers to the value of a variable corresponding to the
k-th step within an iterative procedure.

a Resistance parameter in head loss/flow
relationship, (NP*1) column vector.
Resistance parameter corresponding to link "i", i-th
component of vector a.
Resistance parameter corresponding to link joining

nodes "i" and "j".
Symbol Description
Resistance characteristic parameter in head loss/flow
relationship, (NP*1) column vector (it has head units).
Resistance parameter correspondin g to link "i", i-th
component of vector 13.
Resistance parameter correspondin g to link joining
nodes "i" and "j".
Note that 131j has units

- = a- - Q-ij- n + 131j_i
h-13 of piezometric head.
-I
Al2 Branch to node (NP*(NN-NS)) connectivity (incidence)
matrix, corresponding to unknown head nodes (non-
source nodes).
A 10 Branch to node (NP*NS) connectivity matrix,
corresponding to source nodes.
A21 Transpose of matrix Al2: A21 = Al2T.

A01 Transpose of matrix A10.
A 11 Diagonal (NP*NP) matrix, defined as:
A 11 = diag.( c'ijl Q ij1 11-1)
Diagonal (NP*NP) matrix with the flow exponents "n"

corres p onding to each link. N = diag.(ni), where ni is
the i-th component of a vector n of flow exponents (if
the exponents change from one link to another).
A11* Diagonal (NP*NP) matrix defined as:
A 11 * = diag.( c< i j l Q ii1 n-1 )'
A 11 * = A 11 diag.(13/Q)
(NP*NP) diagonal matrix of the derivatives of the head

loss vector with respect to the flows: G = N A11*
Symbol Description
sE Head loss imbalance per pipe, a (NP*1) column vector.
.1
B 11 Block partitioned matrix used to compute the inverse
B12 of the system matrix:
B21 1
_, F 13 11_L B 12 1
F =
B 22 L A 21 : 0 J L B 21 : B 22 -I n•n••••n
A31 Loop to branch (NL*NP) incidence matrix.
Q2 (NN-NS) flow vector, corresponding to the flows in the
tree of the network (dependent flows).

Q3 NP-(NN-NS) flow vector, corresponding to the flows in
the co-tree of the network (independent flows).

MII
Partitioned matrices of the incidence matrix
A 22 - (NN-NS)*(NN-NS) matrix associated
A22 I to the dependent flows.
A 32 A 12 =
(NP-(NN-NS))*(NN-NS) matrix asso_
A32 I
-
.. ciated to the independent flows.
p (NN*1) column vector of nodal pressures.
Pi pressure at node "i", i-th component of vector p.

Z (NN*1) column vector of nodal ground levels.
Ps Service pressure: pressure at which nodal demands are

assumed to be satisfied.
Pmin. Minimum allowable pressure within the network.
Pmax. Maximum allowable pressure within the network.
A22 Has also been used to describe the sensitivity of nodal

consumption with respect to the pressures: diagonal
(NN-NS)*(NN-NS) matrix defined as:
xix
Symbol Description
I
A22 = diag.(-6q-/Op-).
3. 3.
Nodal consumption at service pressure.
auxiliary nodal consumption variable:
g o = a s + A 22 -a + A 22 Ps
Refers to the estimated value of the variable
(piezometric head, resistance parameter, flow, etc.).
Refers to the true value of the variable.
Refers to the mathematical expectation of a random
variable.
LINEAR SYSTEMS OF EQUATIONS
Symbol Description
A, B A capital letter denotes a matrix, particularly:
L Denotes a lower triangular matrix.
U Denotes an upper triangular matrix.
D Denotes a diagonal matrix.
a, A, A A lower case letter, or a capital (or lower case)
underlined letter denotes a vector.

a, p A greek letter denotes a scalar.
ai A lower case subindexed letter denotes a scalar,
normally the i-th component of a vector.
aii A lower case doubly subindexed letter denotes a scalar,
normally the coefficient of a matrix A, located at the
intersection of row ' . i." and column "j".
Ai A capital subindexed letter denotes either a matrix in
a sequence of matrices or a block-partitioned matrix.

(.)T Denotes transposition, either of a vector or matrix.
A -1 Denotes matrix inversion.
(.)(k) Refers to the k-th value of the variable (scalar,
vector or matrix) within an iterative procedure.
I I xlI Denotes the norm of a variable (vector or matrix).

lx1 Denotes the absolute value of a scalar.
Xj.Denotes a particular eigenvalue of a matrix.

f(A) Spectral radius of matrix A.
<x,y> Inner (scalar) product of vectors "x" and "y":
n
<x,y> = xT Y = yT x = E xi yi
i=1
KRIGING
Symbol Description
(X, H) State variable representing the piezometric head at any

point x of the domain:
(x,H) = (xl, x2, H) T in a two-dimensional space.
Z(x, H) Random function representing the value of the state
variable H at a location x (in general).

Z(xo, H) Random variable (at the particular location x=x0.
Simplified notation:
Z(x i , H) <=> Zx i <=> Zi
E[Z(xl, H)] expected value of the random variable at x=x1.

r(x l , x2) Semi-variogram:
r(x l , x 2 ) a k E( {Z(x 11 H)Z(x 2 ,H)1 2 ]
Cov(Z(x1,H),Z(x2,H)) Covariance:
Cov(Z(x 1 ,H),Z(x2 ,H)) = EE Z(x 11 H)Z(x 23 H) ]
Var(Z(x,H)) Variance of the random variable:
Var(Z(x,H)) = EC Z(x, H) 2 ] - {E[ Z(x, H)]1 2 = 0.2
xoi Unknown weight of the Kriging estimator, where"i"

refers to the measurement Yi and "o" to the point being
estimated Xo , i.e.:
n I
Yo* = Y0 * (X0 ) = E
%o i Y i I
i=1
SPLINES
Symbol Description
One dimensional splines:
s(x) General spline polynomial function.
Xi Set of knots associated to the spline.
Number of knots (Xi , i=1,2,. .,h)
(xr, f(xr)) , r=1,2,. .,m set of data points.
M(x) Cubic Basic (or fundamental) spline (B-spline).
Mn , i(x) General B-spline of degree (n-1). For a cubic spline
n=4, i.e.: M 4, i(x) = M(x) (simplified notation).
r- Linear (unknown) weights, combining the cubic B-splines
into a general cubic spline polynomial:
h+4
s(x) = E r i M(x)
i=1
(h+4)*1) column vector of linear weights ri.

A Matrix of observation equations (observation matrix),
of order m*(h+4) in a cubic spline: An = Mi(x).
A r = f Observation equations, overdetermined system.

(m*1) column vector with the value of the spline
polynomial at each data point:
fi = f(xi) = s(xi) i=1,2,. ..,m
ATAr=ATf Normal equations. Linear system of (h+4)*(h+4)
equations (determined system).
Bi-cubic splines:
s(x,y) General spline polynomial function.
Xi Set of knots associated to the independent variable x.

Pj Set of knots associated to the dependent variable y.
Symbol Description
h Number of knots in the x-axis (Xi , i=1,2,. .,h)
k Number of knots in the y-axis (Pj , j=1,2,. .,k)
Set of (h+1)*(k+1) panels in which the x-y space is

sub-divided.
(xr,yr,f(xr,yr)) , r=1,2,..,m set of data points.

M(x) Cubic B-spline related to the independent variable x.
Nj (y) Cubic B-spline related to the dependent variable y.
r• •
1J Linear (unknown) weights, combining the cubic B-splines
into a general cubic spline polynomial:
h+4 k+4
s(x,y) = E E r ij Mi(x)Nj(Y)
i=1 j=1
r ((h+4)(k+4)*1) column vector of linear weights rij.

A Matrix of observation equations (observation matrix),
of order m*(h+4)(k+4).
A r = f Observation equations, overdetermined system.
f (m*1) column vector with the value of the spline
polynomial at each data point:
fi = f(xi) = s(xi) i=1,2,.. .,m

A T Ar=A TI Normal e quations.Linear system of (h+4)(k+4)*(h+4)(k+4)
equations (determined system).

CHAPTER ONE
INTRODUCTION
1.1. Computer modelling of water distribution systems.
The increasing complexity of water supply distribution systems
on the one hand, and the pressing need for the efficient
management of the resources involved in those systems on the
other hand, have led to the need for simplified representations
of the real systems, which can make alteasier to take adequate
decisions on the operation and design of such systems. The answer
to this need is a model, which although being a simplified
representation of a more complex reality, encapsulates its main
features and reproduces the response of the reality to a certain
stimulus, within a certain degree of accuracy. A model enables
the predicted behaviour of the real system to be assessed
beforehand, in a simpler, cheaper and quicker way, without having
to involve the real system in that process. Thus, different
alternative operating policies, or different alternative designs
can be evaluated with the help of a model, and only the most
efficient alternative can be implemented in the real system.
Future conditions can be anticipated and simulated in the
model, in order to assess the performance of the real system
under a different scenario. This is typically the design case,
where increased consumptions are fed into the model to study the
need for future reinforcements in the system, or to improve the

quality of the service. Clearly this is something that is
impossible to do in the real system itself.
In general, the process of building and operating a model helps
us to get a better understanding of the reality, something which
is beneficial in itself, apart from the benefits deriving to

design and operation.
Different kinds of models have been used to represent water
supply distribution systems. In the past, analogue models,
exploiting the known analogy between ,electric and hydraulic
networks, have been used for that purpose. The advent of digital
computers, and the availability of powerful numerical methods to
solve complex mathematical problems, has allowed the development
of more versatile, cheaper and easier-to-use computer-based
models, constituting nowadays the main field for model

development.
Even though a model is meant to be a simplified representation
of a reality, this does not mean that the model itself is simple.
A computer-based model of a water supply distribution system
involves a number of separate components, which only when linked
together can be used effectively for practical purposes; some of

these components are described below:
a) Physical component.
Any real water distribution network is made up of a number of
physical devices, joined together in some particular way:
reservoirs, pipes, pumps, valves, etc. are some of the typical
devices which can be found in any distribution network. Each

device has some physical characteristics which must be explicitly
considered: dimensions (pipe lengths, diameters), characteristic
parameters (pipe roughness, pump and valve characteristic
curves), etc. Any network model has to include some or all of
these devices as the basic components of the network, with their

corresponding physical characteristics. Additionally, the
topography, geometry and connectivity of the water distribution
system are also important parts of the physical component.
Depending on the model objectives, existing physical devices
(for operational purposes for example) or future ones (new pipes

or pumps in a design-oriented model) may be represented within
the model.
b) Demand component.
The purpose of any real distribution network is to satisfy the
water consumption requirements of the population being served by
it, at a pre-specified minimum service level. In fact, the
demands can be seen as the driving force in the behaviour of a
water supply distribution system; in general, unless there is
some transfer of water between reservoirs or leakage, the
absence of demands will imply no flows in the network, while
higher demands will produce higher flows.
Due to the fact that the demands are spatially distributed
alongside the pipes, a simplifying assumption is made in order to
render the problem mathematically tractable by concentrating the

demands in the nodes.
Furthermore, depending on the model objectives, interest may be
focused on the consumptions either at a particular time (for
design purposes we are normally interested in the demands at peak
time), or during a certain period of time (for operational
purposes we may be interested in the demands throughout a day, a

week, etc.).
c) Operational component.
Each model, according to its particular objectives, will have a
set of operating rules, establishing the way the system is to be
managed. These rules follow closely the real operation of the
system: pumps switched on and off, valve movements, reservoir
levels, pressure reducing valve settings (P.R.V.'s), etc.
In design applications the operating rules are assumed to be
known, although in optimum operation modelling the set of rules
itself is the main unknown.
d) Measurement component.
Real water distribution networks have data measurement systems
with varying degree of sophistication, ranging from manually
operated systems to fully automated telemetry systems.
As far as the network model is concerned, measurements are a
key factor, since they enable us to determine how close our model
is representing the real network. This information can be fed
back into the model, in order to improve its predictions. The
measurements are then the bridge between the model and reality.
Typical field network measurements are: p ressures, water
levels, and flows, whereas the characteristic parameters of pumps
and valves are normally the subject of laboratory measurements.
Together with the basic data gathering system, bad-data

detection and missing data replacement mechanisms (both normally
computer programs), and parameter estimation routines to estimate

unmeasured data are needed.
e) Hydraulic component.
The hydraulic component in the model is mainly related to the
kind of hydraulic regime in which the real network is operating.
We may be interested either in the steady state behaviour of the
network, or in the unsteady state (slow or fast transient); in
each case different physical processes are involved, and a
different approach is needed for modelling purposes.
Another aspect of the influence of the hydraulic component on
the overall modelling deals with the head loss/flow relationship
for each physical device in the network. In this sense we have to
decide, based on our hydraulics knowledge, on the formula which
best represents the real behaviour of a pipe, or which values
should be assigned to initial estimates of roughness, minor
losses coefficients, maximum/minimum pressures and velocities,
etc.
f) Mathematical component.
Depending on whether the network is under steady state or

unsteady state regimes, the problem of representing
mathematically the behaviour of the real system leads either to a
set of non-linear equations, or to a set of hyperbolic partial

differential equations, respectively; each case demands
completely different numerical techniques to be applied to find a

solution to the problem.
Additionally, different ways of setting up the physical
conditions for the solution lead to different ways of assembling
the sets of equations and, indeed, to different analysis methods.
Each method has its advantages and disadvantages, which should be
recognised by the modeller, since the amount of work demanded by
their solution, or their stability properties, might be quite

different.
Since the systems of equations, either in the steady or
unsteady state case, cannot be solved analytically, we need to
find a suitable numerical scheme to solve the mathematical
problem. Because the amount of computation involved in most of
the solution schemes is generally high, except for relatively
small and simple problems, this leads to the need for a digital
computer.
g) Computational component.
In a computer-based model there are two computer-related

components: hardware and software.
On the hardware side we have the machine itself, with its
central processing unit and a vast variety of p eripherals, which
handle the input/output of information: video display units,

printers, plotters, digitizers, etc. It is indeed the
availability of powerful and cheaper hardware, which has made it
possible to expand the use of computer-based water distribution
models, into the operation and design of medium and small
networks. In particular, the availability of microcomputers and
personal computers is revolutionising this and, indeed, other

sectors of the water industry.
On the software side, apart from the system-related software
(operating system, utilities, compilers, etc.), there are a
number of items of specially dedicated software which work in
connection with the different components already mentioned, for
example:
* Network analysis Programs: for solving the steady (or unsteady)

state of the networks. They incorporate the hydraulic and the
mathematical components of the model.
* Demand forecasting Programs: for the efficient storage,
retrieval and prediction of consumption data.
* Data management Programs: providing input/output facilities
for data and results, data updating, etc.
* Data measurement management Programs: including bad data
detection, missing data replacement routines, storage and
retrieval of measurement data.
* Unmeasured data estimation Programs: estimation of unmeasured
nodal piezometric heads and pipe flows.
* Automatic model calibration programs: responsible for

guaranteeing agreement between model and reality.
* Extended period simulation Programs: for studying the
behaviour of the water distribution system during a certain

period of time and under particular operating conditions.
* Optimum design Programs: to determine the best physical

configuration of the system in order to meet certain
objectives (cost, reliability, service level, etc.)
* Optimum operation Programs: for determining the operational

rules which optimise the performance of the system.
Hence, the computational component interacts with many of the
other components of the model, and it might be better seen as
playing a support role, perhaps at a different level in
comparison with the rest of the components. Figure 1.1 shows
schematically the interaction of the different model components,

where we distinguish between a raw model and a calibrated model;
the latter incorporates the measurement information, thus
resulting in a better representation of the reality. In
principle, only calibrated models should be used for practical
purposes, though there are some cases where a calibrated model is
not possible, for example when studying the behaviour of a system
in meeting future demands.
Another important point illustrated in Fig. 1.1, is that a
computer-based water distribution model incorporates not only
computer-related elements, but also physical, operational,
mathematical, hydraulic and customer-related aspects. The network
analysis program is only one component of the whole model,
although it may be used a posteriori in tandem with simulation
and optimisation programs.

Demand
forecasting
programs
•n•n• n111• •nnnn
r-
I
emand perational Data if

component component management I
programs I
Bad data detection
and missing data
replacement prog.
hysical Data
component Base
Estimation of
unmeasured
data programs
Measurement
component
L_ _
-,, Network Automatic

ydraulid` athematica Analysis Raw Calibration
comone, component program Model program
Ep
Calibrated
Model
z
Usable
results/
Fig. 1.1. The water distribution network model and its main
components.
The work presented in this thesis, contributes to some key
aspects of the computational component of water distribution
systems modelling. Specifically, we shall concentrate on the
followin g problems:
* Network analysis for the steady state condition.
* Automatic calibration of the network and estimation of
unmeasured variables.
* Extended period simulation.
1.2. Network analysis: need for a reliable algorithm for design
and operational purposes.
It is a well-known fact that most of the traditional methods
for water distribution network analysis do have stability
problems [Wood (1981a)]. Indeed, problems related to slow

convergence, or lack of convergence, have been documented since
the earliest stages of the development of network analysis
algorithms.
Because the network analysis program is used intensively in
many simulation and optimisation applications (design and
operation), it becomes clear that an efficient and reliable
network solver is highly desirable. Traditionally, the solution
to many convergence problems has been via manual intervention and
via the application of problem-dependent relaxation coefficients,
tuned in each case to produce a stable trajectory towards the
solution. Since, in simulation and optimisation applications,
convergence problems would certainly lead to the break-down of
10
the main algorithm, an unreliable network analysis routine is not
acceptable. Finally, the possibility of using a network analysis
program in a real-time application is simply unthinkable, without
having a priori a fully reliable and stable method of analysis.
On the other hand, the increasing availability and use of
micro and personal computers in the water industry, has made it
highly desirable that any network analysis program should be able
to be implemented on such machines. This adds storage and speed
as important requirements for the development of network solvers,
apart from the stability problem already mentioned.
Here, a non-traditional method of network analysis has been
used as a starting point, which, in theory should not present
most of the convergence problems reported in the application of
traditional methods. This method of analysis is known as the
gradient method and was introduced by Todini (1979).
Todini's gradient method [Todini (1979), Pilati and Todini
(1984)] was initially formulated to incorporate pipes only, and
the initial aim of this work has been to expand the method to
incorporate pumps, valves and pressure regulating valves.
From the numerical point of view, Todini's original gradient
method was implemented with a standard conjugate gradient scheme
for the solution of the linear systems of equations generated in
the network analysis problem. A further aim of this work has been
to investigate the feasibility of alternative numerical schemes
for the solution of the linear systems of equations, since the

performance of the overall analysis program (and consequently
11
that of its simulation and optimisation applications) relies
heavily on the efficient solution of the linear systems. In
investigating these numerical schemes the need for microcomputer
implementation has been kept in mind.
1.3. Need for an automatic calibration algorithm for the
network model.
As already noted, if a model is to be used for operational
purposes, it is a necessary condition for this model to represent
the reality as closely as possible. To ease this task an
automatic calibration algorithm is highly desirable. At the
present time, no satisfactory explicit automatic model
calibration algorithm is available, and many users are still
calibrating network models by hand, on a trial-and-error basis,
and with considerable manpower expenditure. Some existing
calibration algorithms just automate the trial-and-error
approach, without a proper understanding of the underlying
problem. On top of that, calibration techniques based on non-
linear programming techniques lead to solutions which are
extremely expensive from the computational standpoint.
In this work, a further aim has been to produce a calibration
algorithm which can use the raw model (see Fig. 1.1.), the
limited amount of measurements normally available in water
distribution networks to solve the problem of matching reality
and model in an automatic and computationally efficient way.
12
1.4. Simulation of water distribution networks.
Most practical applications of the calibrated network model can
be regarded as simulations of the water distribution system,
carried out at different time horizons. Thus, for operational
purposes we shall be especially interested in short term
simulations, say 24 or 48 hours in advance, in order to
anticipate the expected behaviour of the system, and take
adequate measures to cope with the forecasted demands in an
efficient way, guaranteeing minimum service levels. Medium term
simulations can be carried out, for example one week in advance,
prior to predicted extreme weather conditions. Short and medium
term simulations basically deal with a fixed physical
system and aiming at the efficient management of the storage
within the system in order to satisfy customers demands.
Design optimisation problems can be considered as long term
simulations (say a few years in advance), and they are geared
towards the determination of major changes in the physical
structure of the system, to comply with predicted future demands
and different operating conditions.
The gradient method for network analysis has been used in the
development of an extended period simulation program, which can
handle short term simulations. The basic structure of the
extended period simulation program is shown in Fig. 1.2.

Data Base F Calibrated model
Reservoir levels
+ Consumptions
Network analysis Extended period
program simulation
Piezom. heads program
+ Flows
Results
•Fig. 1.2. Structure of the extended period simulation program.
1.5. Thesis structure.
In Chapter Two a review of the existing network analysis
algorithms is presented, identifying their advantages and
disadvantages, particularly from the convergence point of view.
In Chapter Three Todini's gradient network analysis algorithm
is extended to include pumps. Also a comparison with other
similar techniques is included in this chapter.
Chapter Four deals with the modelling of pressure regulating
devices introducing a new physically-based algorithm.
In Chapter Five some aspects related to the efficient computer
implementation of the gradient method are dealt with,
particularly, the efficient solution of the symmetric positive
14
definite linear systems of equations generated by the gradient
method.
Chapter Six deals with the development of an automatic
calibration algorithm. A review of the existing calibration
methods is presented and a new computer-based explicit
calibration algorithm is introduced. Because the calibration
algorithm relies on estimates of the piezometric heads, three
alternative methods for estimating unmeasured piezometric heads
are proposed and compared.
In Chapter Seven further extensions of the gradient method are
introduced. An extended period simulation version of the gradient
method is presented. Also, a generalised version of the network
analysis algorithm is introduced, able to handle explicitly the
influence of the pressures in the nodal demands.
Finally, in Chapter Eight some conclusions are drawn and
areas of further work are identified.
Four appendices are also included. Appendix A presents a
complete review of the existing direct and iterative methods for
solving the linear systems of equations generated in the gradient
network analysis method. Appendix B summarises the theory and
derivation of the equations used in the application of Kriging as
an unmeasured piezometric head estimator, while Appendix C serves
the same purpose for the case of the estimation using bi-cubic
splines. Finally, Appendix D presents the numerical results of
using the proposed calibration algorithm for solving some test

examples.
15
1.6. Summary of main achievements.
In pursuing the extension of the gradient method for the
analysis of water distribution systems, and in looking for an
automatic algorithm for the calibration of a network model, the
following original contributions have been made:
a) Todini's original formulation of the gradient method has
been extended using a general head loss/flow model which allows
the inclusion of pumps and valves into the algorithm, and which
provides a general framework for the future incorporation of
additional devices. The corresponding algorithm equations have
been modified accordingly. The implemented computer program has
demonstrated that the extended formulation does work, and that
stable and computationally efficient when compared with some
of the best traditional algorithms available.
b) An exhaustive investigation has been carried out, covering
the widest possible range of sparse linear solvers, and the most
efficient methods have been identified and implemented at
microcomputer level, in the context of the gradient method.
c) An original method for modelling pressure reducing and
sustaining valves has been proposed, implemented and tested. The
method follows closely the physical behaviour of these devices,
and has proved to be robust and able to detect problems in the
solutions given by alternative methods. Due to its physically-
based characteristics, the method can be used as an adequate
framework for the modelling of other regulating devices in the
future.
16
d) A new algorithm for the automatic calibration of network
models has been proposed and tested with synthetic data. The
results show that the algorithm is able to replicate piezometric
head and flow measurements.
e) In the context of the measurement component of a network
model, three alternative nodal piezometric head estimators have
been proposed, implemented and compared. The methods are based on
the application of Kriging, bi-cubic splines and a one-
dimensional deterministic interpolation method. The results of
the tested examples indicate that the deterministic interpolation
method performs best.
f) An extended period simulation program based on the gradient
method has been implemented.
g) Finally, a generalised version of the gradient method, able
to consider explicitly the variation of nodal consumptions with
the pressures has been proposed.

CHAPTER TWO
REVIEW OF THE MAIN EXISTING METHODS OF NETWORK ANALYSIS
2.1. Introduction.
Given the network geometry, its physical properties and the
demands to be satisfied, the network analysis problem consists in
finding the flow in each pipe and the piezometric head in each
node corresponding to the steady-state conditions.
Some assumptions are usually made in order to simplify the
solution to the analysis problem:
a) the kinetic heads (velocity component of the total head) are
much smaller than the friction head losses, and they can be
neglected without loss of accuracy;
b) minor head losses at the junctions of pipes and fittings are
smaller than the friction losses, and they can also be neglected;
c) the consumptions derived from the use of water by the
customers are assumed to be concentrated in the nodes of the
network, and not along the pipes, as happens in the real world;
d) the demands are constant and do not change with pressure
variations in the network. A particular time instant is
considered; consequently the demands are "instantaneous demands";
18
e) all the pipes are working under positive pressure;
f) a non-linear mathematical relationship between head loss and
flowrate is accepted as an adequate representation of the flow of
water through the pipe.
The analysis problem is mostly concerned with looped networks,
since in branched networks with only one source, the solution is
directly obtainable from the nodal demands and the network
connectivity. In looped networks, the flow distribution depends
not only on the resistance of one particular pipe, but on the
resistances of the rest of the pipes in the network as well.

Nevertheless, most of the real networks are a mix of partly
branched and partly looped networks.
It must be pointed out that, under the previous assumptions,
the flow distribution does not depend on the ground level. Only
the pressures are affected by the ground level. To avoid
confusion and misunderstandings, we shall work in terms 01 nzde.1
piezometric heads as state variables, instead of pressure; note

that the flow is always produced from a higher piezometric head
node into a lower one, but the same is not generally true when
dealing with pressures.
We shall review in this chapter most of the network analysis
methods, in chronological order, starting with the early
solutions of Cross (1936), one of which was the standard solution
for hand computation and early computer programs. Next, we shall
review the methods made possible by digital computers, where the
capability of the computer in solving large linear systems of
19
simultaneous equations is fully exploited; most of these methods
are based on the explicit application of the Newton-Raphson
technique for the solution of non-linear systems of equations.
The linear theory method is also reviewed, both in its original
formulation [Wood and Charles (1972)] and its gradient-based
formulation [Wood (1981a)]. We shall refer to a network analysis
method as a "gradient-based" algorithm, if the algorithm itself
produces the linearization, which is usually done applying the
gradient operator to the non-linear functions, without having to
pass the non-linear problem to an independent Newton-Raphson
algorithm. A set of non-traditional methods is reviewed, some of
them based on the explicit application of optimisation theory,
while others are based in the application of the unsteady state
equations. Finally, the gradient method of Todini (1979) is
introduced.
Published numerical comparisons of some of the existing network
analysis methods are reviewed and some preliminary conclusions
are drawn. Particular attention will be paid to the convergence
properties of the methods, as well as their storage requirements
and the amount of work involved in their application.
2.2. Hard Cross (1936) methods.
Hardy Cross (1936) developed two methods based on systematic
successive approximations, or "successive corrections" in Cross'
words, for the solution of the network analysis problem. Mainly

due to their simplicity, Cross' methods became widely used, and
they are invariably considered the starting point for network

analysis.
20
H. Cross stated that two conditions for the steady state
solution of the flow in a network are:
a) Flow continuity: at any node of the network, the total
inflow must be exactly equal to the total outflow.
b) Potential continuity: the total (algebraic) change in
potential (piezometric head) around any closed loop of the network
is zero.
These two conditions, together with the non-linear head
loss/flow relationship for each pipe, must be fulfilled
simultaneously in a network under steady state conditions, thus
generating a system of non-linear equations, which Cross managed
to transform into a succession of (linear) scalar problems.
As will be seen later on, these two "conditions", which have
been elevated to the category of "laws" by some authors
(especially in the case of the second condition), are in fact the
necessary conditions, but not pufficient conditions for the
steady state flow.
There are two H. Cross methods, each one starting from the
fulfilment of a different condition (flow continuity or potential
continuity). In the first method, known as "loop flow method", an
initial flow distribution satisfying flow continuity (but not
potential continuity), is successively improved until potential
continuity is met, within some pre-specified accuracy. In the
second method, an initial solution which fulfils potential
continuity is improved until flow continuity is accomplished,
21
within some accuracy limits. We shall review both methods in the
next paragraphs.
2.2.1. Method of balancing heads.
This is the most widely known Cross' method, sometimes thought
to be the Cross method, and its success can be attributable to
the fact that it is suitable for hand computation. The method is
also known under such names as: the "head balance" method, loop
method or "loop-flow method".
Starting from an initial flow distribution, supplied by the
user, the method allows the calculation of corrective loop flow
equations. Earlier computer programs were based on this method.
This method can only be used for looped networks with a single
source, though the addition of some "imaginary" pipes allowsu3to
include more than one reservoir.
Traditionally, the loop-flow method has been derived in the
following way:
Let the head loss/flow relationship for the pipe joining nodes
"i" and "j" be expressed as:

= Kij Qij 2(1)
where:
.is the resistance characteristic parameter, which depends

K 13
on the roughness, length, diameter and units used.
Qij represents the flow from node "i" to "j", which initially
satisfies the flow continuity (or node balance) condition.
22
The second condition establishes that, under steady state
conditions, these flows must produce pipe head losses which are
balanced round any loop, i.e. the algebraic summation of the head
losses around any loop must be zero:
E h 1 J• = 0 V loop (2)
loop
Suppose that the steady state equilibrium conditions have not
been reached yet, but we are close enough; then a loop-flow
correction SQk can be calculated for each loop "k", so that the
new corrected flows Qij + SQk are a better approximation to the
equilibrium conditions. Note that a flow correction applied over
all the pipes in a loop does maintain flow continuity. The new
head loss in the pipe joining nodes "i" and "j" will be:
= Kij (Qij + S Q0 2(3)

or, expanding the binomial:
= K ij (Q ij 2 + 2 Q ij 8Qk + 8Qk 2 ) (4)
In the vicinity of the equilibrium conditions, the corrective

terms 6Qk 2 will be small enough to be neglected, leading to:
hij z Kij Qij 2 + 2 Kij Qij SQk (5)

and imposing the loop condition E h i j = 0:
E h ij z E K-
1J Qij 2 + 2 SQk E Kij Q i j = 0 (6)
which allows us to compute the corrective term:
- E Kij Qij2
8Qk = (7)
2 E Kij Qij
i.e.
- E hii
eQk = (3)
2 E (hij/Qij)
The process can be easily tabulated to facilitate hand-
computation, and it can be programmed in a computer or even in a
23
programmable calculator.
In general, for a head loss/flow relationship of the form:
= Kij Qijn (9)

the correction formula becomes:
- E hij
8Qk = (10)
n E (hij/Qij)
The original Cross' method was subsequently improved by
different authors, in order to reduce the amount of computation
involved and improve its convergence.
Dubin (1947) developed a slide-rule-like device to ease the

hand computation, and suggested some improvements to the original
Cross loop method. He recognised that the convergence of the
method depended on the way the loops are chosen, suggesting that
the loops should be chosen in such a way that the common pipes
are those having the minimum resistance.
Voyles and Wilke (1962) insisted in choosing the loops

having the minimum resistance in their common pipes, they gave a
correlation between the number of iterations and the "common flow
resistance".
Barlow and Markland (1969) suggested that retaining the
quadratic term in the binomial expansion (4) should improve the
convergence of the method. They also found that an over-
relaxation factor of 1.25 affecting their loop-flow correction
6Qk [computed from the full binomial expansion (4)], reduced the
number of iterations by about a half in some small test networks
24
(10 nodes and 13 pipes).
Williams (1973) demonstrated that taking second order terms (or
higher) for the computation of loop-flow corrections, does not
improve the convergence. He adhered to the use of acceleration
factors, applied directly to the correction formula:
- A E hij
8 Qk =
n E (hij/Qii)
where A is a constant acceleration factor, empirically
determined by Williams, which depends mainly on the number of
pipes in the network. Improvements of as much as 80% were
reported by Williams in networks of 100-250 loops (120-290
nodes).
By the early 1970's, new methods had emerged which proved to
have better convergence properties than the Cross loop method,
they are discussed later.
For the application of the Cross loop-flow method, or any of
its improved versions mentioned so far, a balanced initial flow
distribution is needed as input data.
In the case of mixed (looped and branched) networks, the method
requires the isolation of the looped part, as a pre-requisite for
the application of this algorithm. More than one fixed-head
source can be handled via the introduction of "pseudo-loops",
made of imaginary pipes.
Because loop corrections do not affect flow continuity at the
nodes, the accuracy in the computation of the correction factor
25
is not critical in the initial iterations. This was important in
hand calculations.
From a topological point of view this method requires the
specification of the loops (i.e. pipe belonging to each loop,
following a certain order). Epp and Fowler (1970) described a
method for the automatic generation of the loops within the
computer; these loops are the natural set of loops, i.e. those
which do not contain other loops inside themselves.
The application of the Darcy-Weisbach relationship, with
Colebrook-White formula for the calculation of the friction
factor, requires an iterative procedure, since this formula is
implicit in the friction factor.
As it will be seen later on, the Cross loop method can also be
interpreted as the application of Newton's method for the
solution of the non-linear steady state flow problem.
2.2.2. Method of balancing flows.
This is the second method proposed by H. Cross (1936) and is
usually referred to as the nodal method or "head-equations"
method.
Starting with some assumed heads (any set of heads fulfils
automatically the potential continuity condition), successive
head corrections lead to the fulfilment of the flow continuity.
H. Cross proposed that at any stage when flow continuity has not
been achieved, the imbalance at a node can be computed and spread
to the pipes connected with this node "in inverse proportion to
26
their resistances".
Cross himself (1936), on solving a number of examples, pointed

out that the convergence of this method can be "slow and not very
satisfactory".
Cornish (1939-1940) published a development of Cross' balancing
flows method, which is known as the "quantity balance method".
The derivation of the Cross-Cornish method is as follows.
Starting with an estimated piezometric head at each node (given
by the user), let us assume that the node (flow) balance is not
fulfilled. This means that:
i) there is a difference between the inflows and outflows at a
node that is greater than a maximum tolerance value. Let us
denote this difference by (Rai (for node "i"), i.e.:
8Q i = E Q ij qi (12)
where:
flow through the pipe joining nodes "i" and "j".
q i : consumption at node "i"
h.. = K . • Q..n (13)
is the head loss/flow relationship, and if the situation is
such that the nodal balance is not fulfilled, the previously
assumed piezometric head is either underestimating or
overestimating the head losses in any pipe connected to this
node, by a quantity that is denoted by SHi. It is possible to
calculate that correction (SHi), based on the flow imbalance 6Qi,
27
from the following relationship:
+ SHi = Kij (Qij + SQipn (14)
where eQij is the imbalanced flow contributed by the pipe

connecting nodes "i" and "j". SHi is the error in the assumed
piezometric head, which is exactly the correction needed at node
"i". We can re-write this expression as:
SQi j
h-1J- + SH-1 = K-1 •J Q..
1J [ 1 + --- -I (15)
or,
h•1J
• + SH-1 = 1J (16)
but, as a first approximation:
6 j I 1 +n SQij
1+ die
(17)
this is, if: SQij/Qij is small.
Hence,
hij = hij 1 +n (18)
then,
SQij
1J 1 1J n 1J -- (19)
and the piezometric head correction can be computed as:
SQij
SHi = n hij ---- (20)
28
which is reordered to give:
SH1 = 6Qii (21)

n hij
We do not know the value of SQij for each pipe, but we do know
their summation (EQij) for any node "i", which is exactly the
imbalance existing at this node:

E 8Q ij = E Q ij - q i(22)
j j
then, on applying the summation operator on (21) we get:
1
E SH- = E 8Qij (23)
j n h-lj
• j
By introducing (22) into (23), and factorizin g by SH i , which is
constant for node "i", we obtain:
SHi E = E Qij - qi (24)

j n h-1 J• j
which results in:
E Q ij - qi
i
SH i = (25)
E
j n h•1J-
Note that the correction formula (25) collapses if the head
loss in a pipe is exactly zero, and becomes unstable if this
value is very small.
According to Cornish (1939-1940), the method operates as follows:
1. Set the iteration counter "k" to zero. Estimate some
reasonable values for the initial piezometric heads. Those
29
estimates can be as simple as linearly interpolated values
between fixed head nodes (reservoirs), or each node can be given
a reasonable pressure (say 20-30 m.), which, when added to the
ground level, produces the initial head. Let this initial head be
the vector 11 (1o) , where the superscript refers to the iteration
number.
2. Calculate the flows per pipe, according to the initial
piezometric heads estimated before:

Q ij (k) = R ij ( R1(k)_Hj(k))1/n
with Rij and "n" depending on the head loss/flow relationship
in use.
3. Compute the flow imbalance at each node:
8Qi (k) = E Q(k) _ qi (26)

J
4. Calculate the ratio Qij ( k )/n h 1 (k) for each pipe and add up
the ratios corresponding to each node "i", obtaining:
E {Qij (k) /n h()) (27)

i
5. Calculate the correction, dividin g equations (26) and
(27), for each node:

s Q(k)
5H(k) = (28)
ciij(k)
E
j n hij(k)
6. Calculate the corrected piezometric heads, for every node
"i", as:
R i (k+1) = H i (k) + s1j1(k)
(29)
7. The process can be stopped either if the maximum 6'H(k) is
less than a pre-specified tolerance or, if the maximum nodal
30
imbalance SQ i is less than a maximum limit. Otherwise, increment
the iteration counter: k to k+1 and go back to step 2.
In comparison with the loop-flow method, the nodal method has
the advantage that the initial solution for the heads is much
easier to find than the initial balanced flow distribution
required by the loop method. The main disadvantage is that the
convergence can be very slow, and this is the main reason why
this method was not used initially, when computers were not
available.
The nodal method can handle problems with mixed networks
(partly looped and partly branched) straight away, without the
need to split up the looped from the branched part, as in the
case of the loop method. There is no need to supply as input the
definition of loops, since they are not needed in this nodal
approach.
An additional advantage of the nodal method is related to the
fact that we can use the Darcy-Weisbach relationship in a direct
fashion (rather than the iterative one required by the loop
method), thus savin g some computation time. This is shown by
McCormick and Bellam y (1988) and Featherstone (1983, page 114).
Dillingham (1967) reported some convergence problems with the
nodal method and suggested some remedies, based on intuition and
practical experience, but no guarantee of convergence could be
given.
McCormick and Bellamy (1968) reported some possible sources of
failure in the convergence of the nodal method, with some
31
possible solutions. According to them, a faster conver g ence can
be achieved, if the head corrections (29) are not made in the
sequence given by the nodes numbering, but according to the
magnitude of the correction SHi. Proceeding in that way, a number
of corrections can be made in one node before another node is
considered.
As pointed out by McCormick and Bellamy (1968), there are a
number of situations that can result in no convergence, and it
is quite a formidable task to write a computer program with
convergence guaranteed. On top of that, the computer time
necessary to obtain the steady state solution cannot be predicted
on the basis of the network size; the resistance characteristics
of the pipes and the connectivity of the network seem to have a
significant impact in this respect. Manual intervention and some
sound engineering thought seems to be the solution to some
pathological cases of slow convergence.
Barlow and Markland (1969) reported some important reductions
in the number of iterations when a quadratic expression is used
to determine SHi (instead of equation 17). Using this approach in
a network with 14 nodes and 25 pipes, they reduced the number of
iterations from 121 to 77. In their experience, over-relaxing the
correction SHi does not result in a significant improvement,
sometimes it can even be counterproductive. For chronic ill-
conditioned cases, they recommended a loop-flow method.
An inherent problem with the nodal method, which is recognised
by many authors, seems to be the fact that the head correction
32
are carried out one node at a time, without allowing proper
interaction between the nodes of the network. The loop method
does allow some interaction, through the pipes common to two
different loops.
Thus, this approach suffers from an internal contradiction,
which is solved in the following methods, in the sense that
having been developed when computers were not available, it

became used only with the availability of such machines but then
does not take full advantage of them, because it does not solve
all the head corrections simultaneously.
2.3. Newton-Raphson-based methods.
2.3.1. Introduction.
Under this general heading, a number of different methods for
the solution of the network analysis problem are described, all
of them using the Newton-Raphson method to solve the
corresponding systems of non-linear equations. This leads to some
misunderstandings, since the Newton-Raphson method is a general-
purpose algorithm. The main difference between the different
network analysis methods reviewed under this general heading,
will be shown to be in the way that these methods formulate the
problem and assemble the equations.
For completeness, a brief review of the mathematical background
of the Newton-Raphson method is included . After that, its
application to the solution of the network analysis problem is

reviewed, as proposed by different authors.
33
2.3,2. The mathematical_ background of the Newton-Raphson
method.
Newton's method is widel y accepted as an efficient algorithm

for the solution of one-dimensional non-linear equations.
Let
f(x) = 0 (30)
be a non-linear function, whose solution is sou ght within a
certain interval [a, b]. Let x = x ( ° ) , with x ( ° ) E [a, b], be the
true solution to equation (30). Hence,

f(x ( ° ) ) = 0 (31)
and f(x) X 0, for all x E [a, b], and such that x
Newton's method is based on approximating the non-linear
function locally via a linear function; starting from an
approximate solution, say x (1) , the linearized problem provides a
better approximation, say x (2) , to the true solution of the non-
linear problem. Via successive linearizations, a sequence of
approximations is obtained ( x( 1) , x( 2) , x (3) , ....), which gets
closer and closer to the true solution.
A Taylor series expansion of the original function p rovides the
way to linearize the originally non-linear p roblem. Thus,
expanding the function, centred at an ap p roximation x(k):
f(x(k).1.sx(k))=f(x(k))+f1(x(k))8x(k)+kf" (x(k))(sx(k))2+... =0 (31)
This means that, if x (k) is such that f(x 1 ) X 0, we are
looking for a better approximation, say x( k) + 8x (k) , where the

value of the non-linear function becomes zero.
34
In fact, if x( k) is close to the true solution x( 0 ), the term
(8x( k )) 2 and its higher order terms will be very small, and they
can be rejected, reducing (31) to:

f(x(k).1.sx(k)),,,. f(x(k)) 4. f , (x(k)) 8x(k) (32)
Because we aim at f(x(k)+Sx(k)) = 0 , equation (32) can be used
to compute the value of the correction:

- f(x(k))
(33)
f'(x(k))
and the following recursive formula allows us to compute the
sequence of x-values converging to the true solution:

x(k+1) = x(k) .4. sx(k) (34)
We either stop the iterative procedure when two successive
values are close enough, or when the function is nearly zero.
This implies that some minimum accuracy must be specified.
The method converges fairly well for monotonic functions,
although several cases of non-convergence and oscillatory
behaviour are reported in most of the numerical analysis
textbooks; the user must be aware of them, but they are not
likely to occur in pipe network analysis.
The Newton-Raphson method is the direct extension of the one-
dimensional Newton method to an n-dimensional space. In this
context, the following non-linear system of simultaneous
equations needs to be solved:
f i (x l , x2 , x3 , . . , xn ) = 0
f 2 (x 1 , x2 . x3 , .. • , xn ) = °
(35)
fn (x l , x2 , x3 , • , xn ) = °
35
which can be represented in vector notation as:
f(x) = Q (36)
Now, let us suppose we have an approximation x( k ) to the true

solution and we need to find a better approximation, say x(k+1) =
x (k) ex(k); then, following a similar approach as in the one-
dimensional case, we expand the function (36) using a Taylor
series centred at x( k ) and drop all the terms of second or higher
order. In so doing, we get:
ef l ell
- 1 Sxn ( k ) = 0
fi(xl,...,xn)+ --- ex i ( k )+ --- 8x2(k)+...+ !f
exl Ox2 erxn
6f 2 ef2
f2(xl,...,xn)+ --- ex 1 (k).1. ___ ex2(1).4....4. n 6x(k) = 0
Ox i153(2 Oxn
(37)
Oft) afnex2 (k).4.. . . afn ex(k) = 0

exl(k).1. ___
On. (5x2 Oxn
where all the functions and derivatives are evaluated at x = x(k)
This system can be represented in matrix form as follows:
Of: Of
40t xn)
On. Ox 2 Oxn
ef 2 af 2 6f2
6x2(k)
On. Ox 2 exn
=-1* (38)
Of n Ofn afn
ex ]. Ox2emn
where, again, all the expressions are evaluated at x=x(k)
36
The s y stem (38) is a linear system of equations in the unknown
vector Sx(k) = ( sxl ( k) , 6x 2 (k) , ... ,sxn(k) ) T , the column vector
of corrections.
Let us define the right-hand-side vector f(x( k )) as:

f(x(k)) = ( f i (x l (k), ... , x n (k)) , f 2 (x l (k), ...
... , fn(xl(k),...,xn(k) ))T
On defining the Jacobian matrix as:
af l 6f1 Of1
Ox i Ox2 OXn
6f 2 6f2 Of2
j(k) = Ox i tpx2 dxn (39)
:
Of n Ofn 6fn
1 Oxn
- 6)( 15x2
the system (38) can be written in an even more compact way:

j (k) sx( k ) = - vx(k)) (40)
Algebraically, the unknown corrections are obtained by

inverting the Jacobian matrix:
sx(k) = _[j(k)]-1 f(x(k)) (41)
but the numerical solution can be obtained more efficiently by
Gauss elimination or any other efficient linear solver [see
Appendix A].
The iterative sequence of ap proximations to the true solution

of the non-linear system is obtained by:
x(k+1) = x(k) .1. sx(k) (42)
37
As in the one-dimensional Newton method, the iterative
p rocedure is stopped when two successive solutions are close
enough. In the n-dimensional case, this is when some pre-
specified vector norm is smaller than a given accuracy, say E:
I 8X(k) I < 6
Again, as before, the main assumption in the Newton-Raphson
method is that the successive corrections are small; only in that
case is it acceptable to neglect the second and higher order
terms in the Taylor's expansion. This is the key factor in the
success of the method, since it produces a succession of linear
problems out of a non-linear one; however this may also be the
origin of difficulties, especially when the algorithm is started
with an initial solution too far from the true solution.
The algorithm also relies on the inversion of the Jacobian
matrix [equation (41)], this implies that the Jacobian has to be
a non-singular matrix, which indeed is the case in most
engineering applications. Some reductions in the amount of work
can be obtained, if the inverse of the Jacobian is not updated at
every iteration, but at every two or three iterations.
This is the basic mathematical framework of the Newton-Raphson
method. Some special properties of the method emerge when
applying the method to water distribution networks, which will be
explained in due course.

2.3.3. Martin and Peters formulation (1963).
Although Warga (1954) is recognised as the first author to

suggest the use of the Newton-Raphson method to the solution of
general networks, (gas, petroleum, power networks and also water
supply), the credit for being the first to solve the water
distribution analysis problem with the Newton-Raphson algorithm
is attributed to Martin and Peters (1963).
In general, the application of the Newton-Raphson method for

network analysis, as in the case of the Cross methods, may
involve two different ways for the formulation of the equations:
a loop-flow approach and a nodal (head) approach.
The solution proposed by Martin and Peters (1963) is based on

the head equations (i.e. nodal approach), although in their
original paper they used pressures instead of heads.
The nodal balance condition states that, in every one of the NN

nodes of the network, say the i-th node, the following balance
must hold:
E Qij - q i = 0 (43)
Since, in general, the empirical head loss/flow relationship

for pipes is of the form:
hij = Hi - Hj = Kij Qij n (44)
or
(1-n)/n
IHi-Hil
Qij = (Hi - Hj) (45)
Kii1/n
then, on defining m=(1-n)/n, the nodal balance equations (43)

can be written as:
39
m
1Hi-Hil
E Qij - q i = E (Hi - Hj) - qi = 0 (46)
J j K.13.1/n
Let us call each one of these NN equations as:
m
Illi-Hil
f i (H 1 , H 2, ..., H NN )= E (H i - Hj) - qi = 0 (47)
j 1J
for i= 1,2, ..., NN.
Since in the standard water distribution network analysis
problem, the resistance characteristics Kij are considered known,
as well as the nodal consumptions qi, then the following system
of NN non-linear equations in NN unknowns Hi is obtained:
IH1-Hil
f 1 (H 1 , H 2 , ..., HNN )= E (H 1 - Hi) cil = 0
i Kiil/n
m
IH2-HI
f 2 (H 1' H 2 , ..., H NN )= E. (H 2 - Hi) - q2 = 0
• i K 2 , '/fl (48)
IHNN-HjI
fNN (H 1 , H 2' ...' H NN )= E (H NN - Hj) q-NN = 0
i K NNj l/n
which is in the familiar format of the Newton-Raphson method
(section 2.3.2, equation 35).
Note that for all source nodes (reservoirs for example), the
nodal balance equation (43) is redundant, and their corresponding

heads are known (the reservoir water level); then, each source
node can be eliminated, thus reducing the system (48) to NN-NS
40
equations, where NS is the number of sources. This condition is
not introduced into the formulae, in order to keep the notation
as simple as possible, but we have to bear it in mind.
To apply the Newton-Raphson method, the partial derivatives of
the functions fi(Hi, H2,..., HNN)= f(H) need to be computed:
6f 1 (H) 1 IHi-Hilm
= for iXj (49)
6 Hj n Kij 1/n
and
Ofi(l) Of(H)
= E for i=j (50)
6 Hi j 6 Hj
from equation (49), it is easy to see that:
0f1(fl) Of(H)
= (51)
6 H -J 6 Hi
which implies that the Jacobian is a symmetric matrix.
The Newton-Raphson algorithm can now be applied:
1. Find an initial estimate of the piezometric heads in every

node. Denote the correspondin g vector as:
H(o) = (H 1 (o) , ..., W
"NN(°))T.
2. Initialise the iteration counter k=0.

3. Evaluate the functions f(h) at H=H (k) ; then:
f(H(k)) = f(H i (k) , ... , HNN(k))
4. Calculate the coefficients of the Jacobian matrix using the

formulae (49) and (50) for this purpose, and evaluating them
at the point H = H( k ) i.e.:
Of i ( H)1 1 IHi(k)-iii(k)Im
_ for i7j (52)
6 H J- Klj
11=11(k T
41
Ofi(H)] 1 IHi(k)-Hi(k)im
_ E for i=j (53)
6
Hi n j 1J
H=H(k)
5. Assemble the linear system of equations.

6. Reduce the linear system by eliminating the rows and columns
corresponding to each one of the NS sources. This leads to a
system of NH=NN-NS equations, which is still symmetric.
7. Solve the reduced linear system:
-
af laf l Of
6'H 1 (k)
f 1 (H 1 ,... , HNH)
Ni l6E12 • 61-INH
6f 2612 2 6f2 6H2(k)
.. f 2 ( H 1 ,..., HNH)
Ni l0E12 aHNH * = -1* (54)
MNH MNH MNN

8HNH(k) fNH(H1,..., HNH)
Ni lOH2 aHNH .,
- 4- Mil -
where all the expressions are evaluated at H=H(k), as shown in

equations (52) and (53).
The solution of the linear system, which can be obtained by

Gaussian elimination or any other method (see Appendix A for a
review on linear solvers), produces the unknown vector:
8H(k) = (8H 1 (k) , 8H2(k), ..., eHNH(k))T
8. Calculate the corrected heads:

H(k+1) = H(k) 4. 8H(k) (55)
9. If max.18H( k )I < E, where E is a pre-specified accuracy,
stop; otherwise, increase the iteration counter (k to k+1)
and go to step 3. Other convergence criteria may be used:
42
e.g. the maximum norm of the residuals, the maximum number
of iterations, etc.
As pointed out by Martin and Peters (1963), there are in this
formulation a few things that can go wrong:
a) If Hi (k) -Hj (k) is a very small quantity, which is the case
when two adjacent nodes have almost the same piezometric head,
the coefficients of the Jacobian can result in extremely large
numbers, overflowing the computer capacity. In these cases,
Martin and Peters (1963) recommended the temporary removal of
these pipes from the network model.
b) Although the corrections 611 (k) decrease quickly in
magnitude, cases with oscillations were detected; the
recommendation of Martin and Peters for these cases was to halve
the correction 611(k).
So far, only pipes have been considered in the network
(equation 44), but it is not difficult to include any other
device, provided that its head loss/flow relationship can be
expressed as an analytic function. We shall return to this point
later on, when dealing with more complex devices.
Additionally, Martin and Peters (1963) introduced the problem
of the agreement between the mathematical model, represented by
the Newton-Rap hson solution of the network, and the real network.
They recognised the fact that, normally, agreement will not be
achieved; an adjustment of the resistances was suggested as a
remedy, relying just on some intuitive knowledge of the system

and guesswork.
43
2 3,4. Shamir and Howard nodal formulation (1968).
Shamir and Howard (1968) published a paper on the solution of
the network analysis problem, acknowledging some previous work
done at MIT. Also, some work carried out by Pitchai (1966) was
referenced.
One of the main reasons why this paper became classical, and a
compulsory reference in future works, seems to be the fact that a
generalised method for the solution of the network steady state

flow problem was presented.
The Shamir and Howard method is based on a head-equations
formulation (nodal approach), accepting not only piezometric
heads as unknowns, but also nodal consumptions and pipe
resistances (i.e. either diameters or roughnesses). Furthermore,
a methodology to perform a sensitivity analysis is presented,
allowing the user to study the effect of changes in a single
variable over the rest of the network, at a very low
computational cost.
The Shamir and Howard method is based on node continuity:
fi = EQ ij - q i = 0 (56)
for every node i=1,2, NN.
Following the usual procedure of expanding each one of these
nodal equations, and taking just the first order terms for the
extended set of unknowns:
X: set of unknown piezometric heads (subset of II).

Y: set of unknown pipe resistances (subset of K).
44
Z. : set of unknown nodal consumptions (subset of g).
the following linear system of equations is obtained:
Ofi Ofi
fi(X,Y,Z) + E --- 8Xj + E --- 8Yj + E 8ij 8Zj = 0 (57)
j OXj j elYj j i=1,... NU
where:
NO : total number of unknowns.
Kronecker delta (8ij=1 if i=j and zero otherwise).
8Xj : correction to the unknown piezometric head in node "j".
Let us assume that there are, as before, NH nodes with
unknown heads.
8Yj : correction to the unknown resistance in branch "j". Let
us assume that there are NR branches with unknown
resistances.
8Z j : correction to the unknown nodal consumption in node "j".
Let us assume that there are NC nodes with unknown

consumptions.
Note that, apart from the third and fourth terms of the left-
hand-side of equation (57), the problem can be dealt with in
exactly the same way as Martin and Peters (1963).
A necessary condition for the solvability of the system of

equations represented by (57), is that the total number of
nodes NN, must be equal to
NN = NH + NR + NC (58)
The system of equations (57) can be written as:

af 1af iaf i of1 afl Of1
-fl
axi ax NH i aY NR 0z1 aZNC
-f2
af 2af 261'2o f 2 af2 6f2 exiim(k)
sylvk)
ax i àY1 ayo azi 6ZNH
(59)
eYN?(k)
6Z 1 k)
af NN MNN MNN 6fNN M NN MNN
---
6X 16XNH 617 1OYNH 6Z 1OZNc
ezNc(k) -fNij
As we did in the case of the Newton-Raphson method, the
iterative solution of the non-linear system requires an initial
(estimated) solution, say:
( X1(°),..,XNH(0),Y1(0),-.,YNR(°),Z1(°),--,ZNC(°) )T
This initial solution is improved successively until a
sufficiently close approximation to the true solution is found.
The iterative procedure follows the same steps as Martin and
Peters' (1963) method, and both the Jacobian coefficients and the
right-hand-side vector need to be updated at each iteration.
In order to compute the coefficients of the Jacobian matrix,
each partial derivative can be obtained from the nodal-balance
equation. Thus, if the Hazen-Williams formula is being used, in
which case the node-balance for node "i" becomes:
(Hi-Hi)
fi = E qi = 0 (60)
j Kii 0.54 ifii_Hil0.46
and the corresponding partial derivatives are:
of(H) 0.54 Of(H)

= -------- = for iXj (61)
6X j- Kii0.54 Iii i - H j 10.46 a Hi
46
and
6fi(H) 6f1(H)
= E for i=j (62)
6 Xi j 6 Hj
and
6f1(H) - 0.54 (Hi= Hj)
= ------------------ (63)
Kij1.54 ificHil0.46
a Ilk
where pipe "k" is connecting nodes "i" and "j".
and, finally:
6fi(H)
= - 1 if i=k, and 0 otherwise. (64)
6 Zk
McCormick (1969) provided the corresponding equations for the

partial derivatives when using the Darcy-Weisbach formula with
the Shamir and Howard formulation.
With this proposed generalisation the Jacobian matrix is no

longer symmetric, except for the case when there are neither
resistances nor consumptions as unknowns.
The solution of system (59), at an intermediate step "k", gives

the corrections, allowing us to updated the unknowns:
x 1 (k+1) = x 1 (k) + eXi(k)
x 2 (k+1) = x 2 (k) 4. sx 2 (k) .. x(k+1) = x(k) + sx(k)
x NH (k+1) = Y ( ) XY )
"Nli -k. + .'"NEr(k.
Irl (k+1)= y1(k) + sy1(k)

(k+ 1)
= y (k) + 6y2(k)
2 y(k+1) = y(k) 4. 61(k) (65)
:
YNR(k+1) = y (k) + syNR(k)
NR I
Z i ( k+ 1) = Z 1 (k) + szi(k)
Z 2 ( k+1 ) = Z 2 ( k ) + sz2(k) z(k+1) = z(k) + 6'Z(k)
z Nc (k+1) =
Z NC (k) + SZNC(k)
47
Because of the presence of unknown resistances and nodal
consumptions, the condition that the number of nodes must be
exactly the same as the number of unknowns (equation 58), is not
sufficient to guarantee a solution to the non-linear system, as
formulated by Shamir and Howard (1968). According to the authors,
the existence of the solution will depend on the way the unknowns
are selected, and their distribution within the network, since
clearly if all the unknowns are concentrated in only one region
of the network the set of equations (59) becomes rank-deficient;
a set of rules to select the unknowns was proposed by Shamir
(1973):
a) In any node, at least one of the following variables should be
unknown:
- consumption at the node,
- either head at the node or the head of the adjacent nodes,
- the resistance of one pipe connected to the node .
b) A node having an unknown consumption should be connected to
at least one other node with a known consumption.
c) When solving for an unknown resistance, the subsystem
consisting of an unknown resistance and the two terminal nodes
should not have more than one unknown in addition to the unknown
resistance.
Shamir and Howard (1968) stressed the point that any device can
be included in this formulation, provided that the analytic
expression of the head loss/flow relationship is available.
However, they warned the users, that in the event of functions
with discontinuous derivatives, difficulties with the convergence
of the algorithm may be expected.
48
2 3.5- The Newton-Cross method of Liu (1969)
K. T. Liu (1969) proposed a simplified version of the Newton-
Raphson method for the analysis of water distribution networks.
Liu's approach is based on a decomposition of the Jacobian matrix
into two matrices, one containing the diagonal of the original

Jacobian, and the other containing the off-diagonal coefficients.
Thus, the original matrix of coefficients in the system:
sHi(k) - fl(H)
OH2 °NH
M2 M2 M2
6H 2 (k) - f2 (H)
6H 1 151-12 aHNH
• (66)
MNH MNH MNH .414 (k)

• -LANH fNH(11)
OH2 aHNH
can be decomposed as the sum of:
afl Of Of1
0 0 0
OH]. 6E12 aHNH
Of2 af2 of2

0 0 0 • •
aH2 OH1 OHNH

(67)
MNH MNH MNH

0 0 •
---- • 0
aHNH OH 115H2
•••
The argument of Liu is that, because, for the diagonals,
49
of i (H) mi(fi)
=-E
0 111 j 6 Hi
the diagonal terms are much bigger (in absolute value) than the
off-diagonal terms, then the latter ones can be dropped,
obtaining an approximate linear system which is much simpler to
solve than the original system (66). In other words, Liu is
app l y ing the same concept behind the method of Jacobi (or
simultaneous dis p lacements method) for the iterative solution of
linear systems (see Appendix A).
The new (approximated) linear system becomes:

- EV
6f1
0 •. 0 - flap
6f2
0 ... 0 - f2(H)
OH2
* = (68)
:
6f NH
0 0 ••• 6HNH(k) - fNH(11)
6H NH
_
with all the expressions evaluated at 11=11(k)
Equation (68) requires only scalar o p erations to find the
solution vector 6H(k):
- fi(H(k))
6H i ( k ) = V i=1,2,...,NH (69)
6f i 1
6H-
[ 1 H=H(k)
The rest of the algorithm remains identical to the standard

Newton-Rap[son method. According to Liu (1969), the algorithm may
diverge when starting from a poor initial solution.
50
It is interesting to note that Liu's method is, in essence,
Cross' head-equations method. In fact, if we compare equations
(69) and (25), both are completely equivalent. The point here is
that the methods proposed by Cross (1936), are actually based on
Newton's method, applied locally either to the loops or to the
nodes.
2,3.6. EPP and Fowler loop formulation (1970).
Although the nodal approach has a simple formulation, and the
initial head solution is very easy to produce, especially when
compared with an initial flow solution, a number of convergence
problems have been reported in the literature [Martin and Peters
(1963), Shamir and Howard (1968), McCormick (1969), de- Neufville
and Hester (1969), etc.].
In an attempt to cope with these convergence problems, Epp and
Fowler (1970) explored the possibility of using a loop
formulation instead of a nodal approach. They also had in mind
the fact that, normally, the number of loops is smaller than the
number of nodes, producing a reduced-size system of equations;
however, this point may not be relevant when adequate sparse
techniques are used.
On the negative side, a loop approach needs a definition of the
loops, either provided by the user or produced internally by the
computer (with the associated computational cost).
Epp and Fowler (1970) gave a solution to the problem of
generating automatically the loop definition, based on the
51
standard (minimum) connectivity data. Also, they presented a loop
labelling algorithm, in order to reduce the bandwidth of the
matrix of the linear system, thus reducing the storage

requirements. Because the loop method needs an initial balanced
flow distribution, Epp and Fowler (1970) devised an algorithm to
produce it, using a minimal resistance spanning tree.
To derive the loop formulation in the Newton-Raphson format, and
in order to take account of the relationship between loops, pipes
and nodes, some complementary notation needs to be introduced.
First of all, the establishment for each pipe of an arbitrary
positive reference direction is needed, say i-j if the pipe
starts in node "i" and ends in node "j"; thus a flow or head loss
in the pipe i-j will be positive if it is from node "i" to "j",
negative otherwise. Additionally, a positive direction has to be
arbitrarily defined for moving around a loop, this is normally
the clockwise direction. Thus, when moving around a loop to
verify the condition E hij = 0, we shall be adding the head

losses if their direction coincides with the loop rotation
direction and subtracting the losses when they are in the
opposite direction.
The derivation of the loop formulation is as follows: the
potential continuity condition (see derivation of Cross' loop
method), requires that the summation of head losses around any
loop must be zero, i.e.:
E hij = 0 (70)
loop
where the summation is carried out throughout all the pipes i-j
52
belonging to the loop.
Introducing the usual head loss/flow relationship for pipes,
equation (70) becomes:

E Kij '-' S fl = 0 (71)
loop
Starting with an initial flow distribution which fulfils the
node-balance condition, say Qij (0 ), and assuming that the loop
balance is not fulfilled, i.e. E K ij Qij n / 0, a loop flow
additive correction can be calculated, in order to get a better
approximation to the fulfilment of the loop-balance condition;
this correction can be represented, at the iteration "k", as a

vector 8Q(k)=(6Q 1 (k),6Q 2 (k), ..., SQNL (k))T , where NL is the
number of loops in the network. This is equivalent to what was
done in the Cross' loop method, but now is done for all the NL
loops simultaneously. The corrected loop balance condition can be
written as:
NL
E K ij (S ji Q ij + E S ijh 8Q 1 (k))n = (72)
Vi,jEl h=1
1=1,2,..., NL
where:
+1 if pipe i-j has the same direction as loop "h"

8-•
1Jh = - -1 if pipe i-j is in the opposite direction with
respect to loop "h"
0 otherwise
Note that the second summation in (72) accounts not only for
the loop flow correction in the current loop, but also for the
53
corrections in the neighbouring loops (for these pipes belonging
to more than one loop).
Equation (72) is a system of NL non-linear simultaneous

equations in NL unknowns: sii(k)=(8Q 1 (k) , 8Q2 (k) , ..., eQNL(k))T,
which can be put in the standard Newton-Raphson format:
NL
f 1 (Q,8Q) = E K ij (8 ij1 Q ij .4. E rs ijil eQh (k))n = 0
Vi,jEl h=1
NL
f2 (Q,6Q) = E Ki j (e ij2 Q ij + E 6 ijh 8Q h (k))n = 0
Vi,jE2 h=1
(73)
NL
f NL (§2,6Q) = E K-V Og ijNL Qij + E S
ijh 8Qh ( k )) n = 0
Vi, jEN h=1
The Newton-Raphson solution of the non-linear system of NL
equations, requires that the following linear system must be
solved successively:
-
of i af l afl
... 6'Q1 (k) - f1(42,8Q)
084:4 1 '58 % 68QNL
6f 2 af 2 6f2
... 8Q2(k) - f2(Q,8Q)
06Q 1 68Q2 • 68QNL
* = (74)
:
: :
af NL tlfNL afNL
• • 8QNL(k) - fNL(Q,6Q)
66Q 1 158Q 2 68QNL
where all the functions and partial derivatives need to be
evaluated at the initial conditions, where Q=Q(k) and sa=sb since
54
the initial solution is balanced.
The partial derivatives, evaluated for the initial conditions,
can be computed directly from (73):
afi
= + E nK-
ij Q-
ij-n-1 for 1=1,2, ..., NL (75)
6 8Q 1 Vi, jEl
afl
= - E n Kij Qi3n-1 for 1X1( and 1=1,2,... NL (76)
OeQk Vi, jEk
Vi, jEl
Note that the matrix of coefficients is symmetric.
On solving the linear system (74) for SQ(k), a better
approximation for the pipe flows can be computed by:

NL
Q ij (k+1) Qii(k) E6'ijh 8Qh(k) (77)
h=1
If Qij (k+1) and Q1 (k) are close enough, say their difference
is less than E, a pre-specified accuracy, or if I5( k )1<E, then
the procedure can be stopped, otherwise the iteration counter is
increased to k+1 and a new 6Q(k) is computed.
As pointed out by Epp and Fowler (1970), the introduction of
more than one fixed head source can be performed via the
introduction of "pseudo-loops", with a fixed head loss between
two nodes. They also gave a procedure for including pumps within
this formulation.
I. P. King (1970) compared the results of network analysis
carried out both by the loop and nodal methods, obtaining
evidence that the loop approach performs better than the nodal
approach, in terms of execution time, for small and medium size
55
problems (less than 200 nodes); the nodal approach was apparently
better for larger problems.
Epp and Fowler (1971) recognised the possibility that the loop
method may perform better in some problems, while in others, the

nodal approach can be the best.
2.3.7 The mesh-nodal aPProach of Hamam and Brameller (1971).
A rather different approach, in the context of the application
of the Newton-Raphson method to network analysis, is the so
called "mesh-nodal" approach of Hamam and Brameller (1971), which
can be applied to any kind of network: electrical, gas, water or
any other fluid.
Before describing the mesh-nodal approach, some additional
notation is needed, dealing with the handling of the topology of
the network. Also, some graph-theory definitions are introduced.
Following the notation introduced by Todini (1979) and Todini
and Pilati (1987), three topological matrices for describing the
branch-node and branch-loop connectivity are defined. The branch-
to-node incidence matrix for the non-source nodes A l2 , is a

(NPx(NN-NS)) matrix, with only two non-zero elements per row:
1 if pipe i ends at node j.
A. (i , j) = - -1 if pipe i starts in node j. (78)

0 otherwise
where NP is the number of branches, NN the number of
nodes and NS the number of source nodes. Thus NN-NS represents

the number of unknown head nodes.
56
The branch-to-node incidence matrix for the source nodes A10,
is a (NPxNS) matrix, of similar characteristics as to Al2:

-
1 if pipe i ends at node j.
A w( i, j) =- -1 if pipe i starts in node j. (79)

0 otherwise
The branch-to-loop incidence matrix A13 is a (NPxNL) matrix,

defined as:
_
1 if pipe i flows in the same direction
as loop j.
A 13 (i,j) = -1 if pipe i flows in the opposite direction
as loop j. (80)
0 if pipe i is not in loop j.
where every loop has a pre-defined positive direction, normally

the clockwise direction.
In general, the transpose of a topological matrix is denoted by
interchanging the sub-indices in the matrices, thus:
A21 = A l2 T(81)
A01 = A lO T(82)
A31 = A13 T(83)
With these topological matrices, most of the relationships used
in the network analysis problem can be expressed in a simpler

manner:
Nodal balance:
A21 Q = a (84)
Pressure drop in a pipe:
h = A l2 H (85)
57
Loop balance:
A31 h = (86)
The head loss flow relationship in pipes, can also be
simplified with matrix notation; thus:
= Kij Qij n(87)

can be represented as:
h = A ll Q (88)
where:
A11 = diag(K i lQ i I n-l ) i=1,2,.., NP (89)
with Ki and Qi denoting the resistance and flowrate at pipe
"i", respectively. Clearly, All is a (NPxNP) diagonal matrix, and
is a function of the flow (i.e. it is not a topological matrix).
With this notation, the loop balance can re-written, by

replacing (85) into (86):
A31 Al 2 H = Q (90)
An imp ortant conclusion can be drawn from the last equation: it
is straightforward to prove that equation (90) holds for any
piezometric head vector H, because:
A31 Al2 = 0 (91)

This means, in practice, that the loop balance condition is
essentially a topological result, rather than a physical law.
Similarly, the following also holds:
A31 A 10 = 0
A21 A 13 = ° (92)
A01 A 13 = °
Finally, some graph-theory concepts and notation are necessary.
In an implicit way, some graph-theory notions have already been
58
introduced and used , e.g. the concepts of node (vertex), branch
(or edge), incidence and indeed the network itself, which is a
"graph", i.e. an oriented line diagram, composed of all the
branches and nodes.
A "path", will be a sequence of edges which, when followed
sequentially, allows us to move from one node to another.
A connected graph will be a graph such that a path exists
between every pair of nodes. A subgraph is a subset (edges and
vertices) of a graph, and it can be connected or disconnected.
A circuit (or loop) will formally be a subgraph such that there
are exactly two different paths joining every pair of vertices in
the subgraph.
The "tree" of a connected graph will be any graph fulfilling
the following conditions:
a) It is a connected subgraph.
b) It contains all the vertices of the graph.
c) It has no circuits (loops).
As a result of the previous conditions, a tree has exactly one
path joining every two vertices. Furthermore, a NN vertices graph
defines a tree with only (NN-1) edges on it; this can be proved
by induction.
The "co-tree" of a graph is simply the complement of the
corresponding tree; thus, if NP is the number of edges and NN the
number of vertices, there are NP-(NN-1) edges in the co-tree. In
the graph-theory nomenclature, the edges of the tree are referred
59
to as "branches", and the corresponding edges of the co-tree are
known as "chords". Hence, the addition of a chord to a tree
produces a circuit. In water distribution networks, we may refer
to the links between nodes also as "branches", meaning any kind
of physical device connected between two adjacent nodes: pipe,
pump, valve, etc. Fig. 2.1. helps to visualise some of the graph-
theory concepts already introduced.
NP n 16 edges
NN 11 nodes
65 m
edge
o) Network graph.
(NN — 1) m 10 branches
branch
b) Tree of network shown in a)
NP — (NN — 1) .- 6 chords
c) Co—tree corresponding to network shown in a) and tree shown in b)
Fig. 2.1. Representation of a network graph and its tree and

co-tree.
60
The concepts of tree and co-tree are relevant to the mesh-nodal
approach of Hamam and Brameller (1971), since they allow us to

sep arate the flows in a water network between dependent and
independent flows, the independent flows being all the flows

corres p onding to the co-tree of a network, while the dependent
flows correspond to those in the tree. It can be proved that all

the dependent flows can be computed directly from the independent
ones, via topological considerations only. This means that, in
order to know the flow distribution in a network, we only need to
specify the values of the independent flows. Thus, we can
partition the flow vector Q into two sub-vectors, say Q2 and Q3,
the former corresponding to the dependent flows (i.e. flows in
the tree) and the latter to the independent flows (i.e. flows in
the co-tree). Hence:
Q2<-- tree of dependent (NN-NS) flows

Q = [.... (93)
Q3 <-- co-tree of independent NP-(NN-NS) flows
With the previous definitions, the mesh-nodal approach of Hamam

and Brameller (1971) is now introduced.
First of all, the computation of the dependent flows from
the independent ones is presented. On introducing the partitioned
flow vector (93) into the nodal balance equation (84), we get:
Q2
A21 Q = A21 HI = a (94)
Q3
The matrix A21 can be partitioned, accordingly to the

dep endent/independent flow definition; thus:
61
A 21 = [ A 22 : A23 ] (95)
which, when introduced into (94) gives:
[ A22 : A23 ] [ •Q2

• • •] = 2 (96)
Q3
which, on expansion, gives:
A22 Q2 + A23 Q3 = g
The dependent flows can be computed from the independent ones,
provided that A22 is non-singular, as:
= [ A22] - [A22 ]-1 A23 Q3 (97)
which can be expanded back to:
Q . [_Q2 [1 . A22-I R I2-1[ -A2 A23 1

+ Q3 (98)
3 0 31 3 Q
or, in a more compact way:
Q = Qa + M 13 Q3 (99)
where:
A22 -I 1
Qa = (100)
03
[
and
[ - A 22 -1 A231
M 13 = (101)
13
To compute MI3, it is necessary to go back to the pressure drop
equation (85). Noting that Al2, the transpose of A21, can be

expressed using the partition (95) as:
62
A22 1
A l2 = [ ---- (102)
A32
equation (85) becomes:
A22
A l2 H = [ 7, ---I H = h (103)
A 32
By introducing the corresponding partitioning of the head loss
vector, according to the dependent and independent flows:
h = -h2- (104)
h3
{
equation (103) can be decomposed as:
A22 H = h2 (105)
and
A32 H =h3 (106)
From (105), it is possible to compute El, provided that A22 is

non-singular:
H = A22 -1 h2 (107)
which, when introduced into equation (106) gives:
A32 A 22 -1 h2 = 123 (108)

or, re-arranging:
[ - A32 A 22 -1 : 13 ] . [ --- -2--- = 03 (109)

h3
Equation (109) is indeed the loop e quation (86), with:
A31 = [ - A32 A22 -1 : 13 ] (110)
Transposing the matrix A31 from equation (110):
63
[-- A 22 -1 A 23 1
A 13 =
13
which shows that matrix M13 in equation (99) is exactly the
same as A13, and equation (99) can be re-written as:
Q = Qa + A 13 Q3 (112)
Now, introducing (112) and (88) into the loop equation (86), we
get:
[ A 31 A ll A 13 ] Q 3 + A 31 A ll Q a =0 (113)
This is a non-linear set of equations in Q3, the independent

(co-tree) NP-(NN-NS) flows, with:
[ A22 -1 g ]
Qa = (114)
03
Equation (113) can be solved iteratively with the Newton-Raphson

method by:
Q3(k+1) . Q 3 (k) _ j33 -1 f(Q3(k)) (115)
where:
J 33 = [ A 31 (N A ll ) A13 ] (116)
with:
N = n I : diagonal matrix with the flow exponent in (87).

and:
f(Q3 (k) ) = [ A31 A ll A 13 ]C2 3 ( k) + [ A31 A ll Qa ] (117)

k) Q(k)
Equation (115) can be solved numerically as a linear system of

equations in 8Q 3 (k) = (k+ 1) _ Q3(k):
J 33 (Q3 (k+1) _ Q 3 (k)) . _f(Q3(k)) (118)
64
or
J33 6' Q 3( k) = -f( Q3(k)) (119)
Upon computing 6'Q 3 (k), Q3 (k+1) can be calculated and improved,
if necessary, by re-solving (118) until convergence is achieved.
Having Q3 ( the last Q3(k+1) ) , equation (97) allows us to

compute the dependent flows Q 2 , and the problem is actually
solved, since the head losses can be computed with (88) and the
piezometric heads can be obtained by considering the head losses
accumulated from the sources up to each node.
.3.8. Other formulations.
The main approaches for the solution of the network analysis
problem have been • reviewed, based on the nodal and loop
equations; both lead to a non-linear system of equations which is
solved via the application of the Newton-Raphson algorithm. A
number of variations and improvements to these main approaches
have been made through the years, some of them are reviewed in
the following paragraphs.
Zarghamee (1971), following a nodal approach, presented a
Newton-Raphson-based formulation able to handle pumps and
pressure reducing valves.
Lam and Wolla (1972a and b), based on previous work by Broyden
(1965) on "quasi-Newton methods", proposed an improvement to the
Newton-Raphson nodal solution of the network analysis problem,
where the the Jacobian matrix (and its inverse) is not explicitly
computed, but it is iteratively approximated, thus saving the
65
time spent in the assembly and inversion of the matrix. Another
advantage of this approach, according to Lam and Wolla (1972a),
is that the analytic partial derivatives are not needed, which
makes things easier from the computational standpoint.
This modified Newton-Raphson method was tested by Lam and Wolla

(1972b) although a case (51 branches and 32 nodes) where the
method failed to converge was reported. The authors did not
provide the corresponding full data set in their paper and
attributed the failure to the fact that the initial values of the
heads are too far from the final ones.
For the cases where the algorithm may present a low convergence
rate, Lam and Wolla (1972b) included a multiplicative relaxation
parameter in the correction of the unknown variables, which is
computed explicitly via a one-dimensional search.
Another interesting feature of Lam and Wolla's formulation
(1972a and b), is that they considered the nodal demands as a
general explicit function of the nodal pressures, allowing for a
more realistic modelling in this respect, since in practice the
consumptions are indeed pressure-dependent.
Lemieux (1972), in applying a relaxation parameter similar to
that of Lam and Wolla (1972b), claimed that convergence is always
achieved.
Kesavan and Chandrashekar (1972), based on graph-theory
concepts, re-formulated the nodal and loop approaches, leading to
the corresponding non-linear equations, expressed as a function
of topological incidence and loop matrices. Their derivation is
66
independent of the numerical method used to solve the non-linear
systems. They did not give details on the iterative technique
used to solve the non-linear equations, which is said to be based
on a graph theory approach.
Donachie (1974), working with a nodal-based Newton-Raphson
formulation, found that a simple way of avoiding convergence
problems is to halve the correction step (i.e. a relaxation
factor of 0.5), when slow convergence has been detected.
Additionall y , a very simple way of initialising the piezometric

heads, without the risk of producing two equal heads, is
presented: an arbitrary (reference) head taken as equal for all
the nodes is considered, and the initial piezometric head of each
node is simply this reference head minus the node number.-
Finally, Donachie also introduced a systematic procedure for
calibrating the network model, we shall return to this in Chapter
Six.
By this time, most of the developments of Newton-Raphson-based

network analysis algorithms were concentrating on the search for
computationally efficient numerical methods, particularly in
connection with the solution of the linear system generated by
the Newton-Raphson method, since it was recognised as one of the
most expensive steps in the overall solution. A wide variety of
methods emerged, most of them aimed at finding the numbering
system for the nodes (or loops, depending on the approach in use)
able to minimise the creation of non-zeros (fill-in) in the Gauss
elimination process, thus reducing both the storage and
computation time needed. Chandrashekar and Stewart (1975), Shamir
67
(1974), Chin et al. (1978), Gay et al. (1978), and others,
concentrated on this subject.
In the context of efficient linear solvers, a comparison of
the performance of the most efficient linear solvers is presented
in Chapter Five. Appendix A includes an exhaustive review of
sparse techniques, ranging from the simplest to the most
sophisticated direct and iterative linear solvers, for the

symmetric p ositive-definite case.
2.4. The linear theory method.
Under this heading, a number of network analysis methods are
reviewed. This class of methods was originally devised to solve
the non-linearity of the system of equations describing the
steady state, without using the Newton-Raphson technique.
The linear theory method was proposed by Wood and Charles
(1972), within a loop-based framework, but has evolved recently
(Wood, 1981a) to a gradient-based algorithm, thus becoming closer

to the Newton-Raphson based family of methods.
2.4.1. Wood and Charles formulation (1972).
The linear theory method was developed by Wood and Charles
(1972) as an alternative to a looped Newton-Raphson approach,
where a balanced initial flow distribution is needed and where a
system of as many non-linear equations as loops has to be solved.
Another objective in mind at this time was to overcome the slow
convergence which had been detected in some applications of the

looped Newton-Raphson method.
68
In the loop approach, it has been already noted that the
problem is to solve the non-linear system generated by the

simultaneous fulfilment of flow continuity at the nodes and
potential continuity around the loops (or energy paths). This can
be expressed as
E Qij - qi = 0 i=1,2,..., NN-NS (120)

i
and
E h 1„1
. - .-
= E K 1„1 Q ij
. - n = 0 loop=1,2,.., NL (121)
loop loop
The system (120) is an independent system of equations only
when the source nodes are excluded, and is linear in Q. Equation
(121) is a non-linear system of NL equations in the flows, where
NL is-the number of loops (or energy paths).
Equations (120) and (121) constitute a set of NN-NS+NL non-
linear equations having as unknowns the flows only. Because the
number of unknown flows is exactly NN-NS+NL, the coupled system
(120)+(121) is enough to determine the unknown flows. Due to the
non-linearity, an iterative procedure is needed.
The main idea behind the linear theory method of Wood and
Charles (1972) is remarkably simple, and it can be explained in
the following terms. Let us assume that the iterative process
required to determine the equilibrium flows generates the

se q uence of vectors
Q (0) , Q(1) , . . , Q(k) , Q(k+1) .. . (122)
Then, if the flow sequence is converging, and if Q (k) and

Q (k+1) are the flows at the iterations "k" and "k+1"
69
respectively, it can be assumed that both are close to the
equilibrium flow vector, and that both are very similar. Hence,
the non-linear term in equation (121) can be approximated by:
Q ij n z [Qij(k)r-lQij(k+1) (123)
which is linear in Qij(k+1).
Then, equation (121) can be re-written as:
E K ij [Q ij (k)](n-1)Q ij ( k+ 1 ) = 0 (124)
loop loop=1,2,... NL
which can be simplified to:
E Kij (k) Q ij (k+1) = 0 (125)

loop loop=1,2,..•, NL
where:
K 1 (k) = K ij {Qij(k)ln-1 (126)
Similarly, using the sequence of flow vectors (122), equation

(120) can be approximated by:
E Q ij (k+1) _ qi = 0 i=1,2,..., NN-NS (127)
j
which, when coupled with (125), generates a system of NN-NS+NL

linear equations in .il(1+1): the (NN-NS+NL)xl column vector of
approximate flows. Hence,
E Q ii (k+1) _ qi = 0 i=1,2,..., NN-NS

j
(128)
E . -(k) Q 1. J.(k+1) = 0
K 1J loop=1,2, ..., NL
loop
On solving (128) for Q (k+1) , we can re-compute K 1 (k) via
equation (126), and improve the approximation of Q, until a
certain convergence criterion is met.
70
At the beginning of the iterative process (k=0), any imbalanced
flow solution Q (0) can be used to compute K( 0) , because the upper
part of (128) ensures that the next flow approximation will
always be balanced. Hence, in this method, an initial balanced

flow solution is not needed, though an initial flow estimate
closer to the final solution will require less iterations to

reach the equilibrium conditions.
To cope with oscillations detected in some cases, Wood and
Charles (1972) recommended modifying the flows, prior to the re-
evaluation of the new ji (k+1 ), using an average of the last and
previous flow. Thus, after computing Q (k+1) , and if convergence
has not been achieved, increase the iteration counter k to k+1
and compute:
Q(k) = 1.i, ( Q(k-1)+Q(k-2)) (129)
and, then, re-compute:

K 1 (k) = K ij N 1j (k)1(n-1) (130)
and solve (128) again.
The system of linear equations (128) is not symmetric and, as a
result, it is neither possible to take advantage of that feature
for storage savings nor from the stability point of view.
Pivoting for stability reasons will be necessary when solving the
system (128); see Appendix A for details.
It has been argued and recognised [Fietz (1973), and Wood and
Charles (1973)], that one of the main disadvantages of the linear
theory method is its high storage requirements, in order to store
a matrix of the size NP*NP. Althou g h this point is becoming less
71
important with the advent of bigger and cheaper computer memory,
it is a factor to bear in mind when comparing the linear theory

method with other methods.
Jeppson and Tavallaee (1975), Jeppson and Davis (1976) and
Jeppson (1976), have shown how to handle pumps, multiple
reservoirs and pressure reducing valves within the framework of

the linear theory method.
Collins (1980) demonstrated, with a very simple example, that
the linear theory method of Wood and Charles (1972) is inherently
unstable when the head loss/flow relationship is exactly
quadratic (n=2), where the method oscillates indefinitely between
two values. Unfortunately, the head loss/flow relationship is
close to quadratic for the Hazen-Williams equation (n=1.85), and
exactly quadratic for the Darcy-Weisbach equation. The flow
averaging procedure of equation (129) attempts to solve the
stability problems, but there is no guarantee of success.
2.4.2. Nodal formulation of Isaacs and Mills (1980).
Following the initial success of the original loop-based linear
theory method of Wood and Charles (1972), Isaacs and Mills (1980)
p resented the corresponding nodal-based version of the method.
The motivation for a nodal formulation of the linear theory
method is that, when multiple fixed head sources (reservoirs) are
available, an algorithm with the heads as unknowns should be more

efficient. Also, a symmetrical and more storage-efficient version
is highly desirable.
72
The rationale behind Isaacs and Mills' linear theory method is
similar to that of Wood and Charles (1972). Thus, if the head
loss/flow relationship for a pipe joining nodes "i" and "j" is
expressed as:
Hi - H j = Kij lQ 1j 1 n-1 Qij (131)
and, if the node continuity equation for node "i" is written as:
E Qij - qi = 0 i=1,2,..., NN-NS (132)
and if it is assumed that the iterative procedure produces a
sequence of successive flows converging to the equilibrium
conditions:
Q(0), Q(1) , . . , Q(k), Q(k+1) (133)
then, near the equilibrium conditions, Q( k) and Q (k+1) will be
very similar and, consequently, equation (131) can be
approximated as:
Q ij (k+1) lQij(k)in-1 = (1/K 1j ) [H 1 (k+1) _ Hj(k+1)] (134)
or,
Q ij (k+1) = 0 1j (k) [H 1 ( k+1) _ Hj(k+1)] (135)
with
c.1 J.(k) = 1/ (Kij 1Qij(k)In-1) (136)
Similarly, equation (132) can be approximated as:
E Q ij (k+1) _ q i = 0
1=1,2,.. NN-NS (137)
Equation (135) can be introduced into the nodal balance
(equation 137), obtaining:
E 0 " (k) [H i (k+1) _ H j (k+1)] = qi 1=1,2,... NN-NS (138)

i
which is a system of NN-NS linear equations in the NN-NS

unknown heads H(k+1). The matrix of coefficients is now
73
symmetric, and appropriate storage handling can reduce
considerably the storage requirements; from the stability
viewpoint, the solution of the linear system (138) does not

require pivoting.
For updating the flows at each stage, Isaacs and Mills (1980)
proposed the following method:
a) Compute:
[Hi(k+1) _ Hj(k+1)]
Qij* = (139)
1Qii(k)1n-1
Having computed 11(k+1) from (138), Qij* should be fulfilling
the nodal balance condition (137).
b) Average Qij* with the p revious flow to obtain the new
updated flow:
Q ij (k+1) = (Q ij * Q ij (k)) (140)
According to Isaacs and Mills (1980) this flow updating
procedure is computationally efficient and guarantees that
continuity is always maintained, provided that Q( k) fulfils the
continuity condition .
Once Q (k+1 ) has been obtained, a pre-established convergence
criterion will stop the iterations, or else we shall need to
update the iteration counter k, and go back to equation (136) and
to the solution of the linear system (138). The algorithm does
nut need an initially balanced solution, though a sensible
starting point will reduce the number of iterations required to

reach the final solution.
As it can be easily seen from equations (136) and (139), the
74
presence of nearly zero flows will produce unbounded values, and
the collapse of the algorithm. The removal of the pipes
corresponding to very low flows is recommended by Isaacs and
Mills (1980) as the solution for such a problem.
According to Isaacs and Mills (1980), the nodal version of the
linear theory method requires less iterations than the original
linear theory algorithm of Wood and Charles (1972).
Wood (1981b) points out that, from the convergence rate
viewpoint, the methods based on heads are known to be less
reliable than the flow-based methods, and reports that a new
version of his original linear theory method [Wood and Charles
(1972)], now based on a gradient approach, has much improved
performance.
2,4.3. The gradient-based formulation of Wood (1981).
Wood (1981a), in an extensive comparison between different
network analysis solvers, presented a modified version of his
original linear theory method (Wood and Charles, 1972), the main
difference being the way in which the linearization is carried
out. The non-linear equations are now linearized using a gradient

concept, as follows.
The non-linear head loss/flow relationship for pipes is:
= h(Q1) = Kij Q ••n (141)
In an iterative process, for a flow Q1(k) at step "k", the

s ummation of the head losses through any ener gy Path, must be
e qual to the known piezometric head difference between the
75
initial and final nodes: 8E 1 . A loop is simply a path where
initial and final nodes coincide (i.e. 8E1 = 0).
The fact that the loop balance is not exactly met, means that:
E h(Q(k)) = E K ij [Q ii (k)]n 6E1 (142)
1 1 1=1,2,..., NL
where the summation is carried out over all the NL loops (or
"energy paths") in the network.
The aim is to determine a pipe flow correction such that:

Q ii (k+1) =Q 1 (k) gQii(k) (143)
or simply
Q(k+1) = Q(k) sQ(k) (144)
The condition to find SQ( k) is that:
E h(Q(k+1)) = 6E1 (145)
1 1=1,2,...,NL
This means that to find an approximation to SQ (k) , we must

expand (141) using Taylor's series and drop the second and higher
order terms:
h(Q(k+1))....:,. h(Q (k) ) .1. ____ ( Q(k+1) _ Q(k)) (146)

MI
[ Oh Q(k)
and, on imposing the loop-balance condition:
Oh
E h(Q(k+1)) 2., Eh(Q (k) ) 4- E ---- (Q(k+1) _ Q (k) ) = 6E, (147)
1 1 1 OQ
Q=Q(k)
1=1,2,... NL
Note that in (147) the flow correction 8Q(k) is not explicitly
computed and the unknown flows are c alculated instead. Re-
76
ordering (147) leads to:
Oh Oh
E [____1 Q (k+1) = E 1[____I Q(k) _ h(Q(k))1 + 8E1 (148)
1 OQ 1 OIQ
Q=Q(k) Q=Q(k)
1=1,2,... NL
which represents a linear system of NL equations in NP unknowns.
To determine the whole set of unknown flows, the nodal balance

equations need to be added:
E Q 1i (k+1) _ q i = 0 Vi=1,2,..., NN-NS
(149)
which, when coupled with (148), constitutes the new linear

system of equations in Ei(k+1):
E . Q ii (k+1) qi = 0
i=1,2,..., NN-NS
[ Oh Oh
Q(k+1) = E 1[____I Q(k) -
h(Q(k))1 + 6E1 (150)
1 (5Q 1 6,Q4
Q=Q(k) Q=Q(k)
1=1,2,... NL
The linear system (150) replaces the previous linear system

(128) and, due to its stability characteristics, there is no need
for averaging the new flows with those of the previous stage.
Wood (1981a) pointed out that the new gradient-based method has
a superior convergence rate than the original linear theory
method of Wood and Charles (1972). Basically, this gradient-based
algorithm is applying the Newton-Raphson method to the non-linear
set of NN-NS+NL equations formed by coupling (145) and (149)
and, hence, it is actually closer to the methods reviewed in the

previous section 2.3.
77
By this time (1981), the linear theory method had a reputation
as an efficient network analysis solver, mainly due to its
simplicity, and the move towards the g radient-based algorithm

improved its convergence.
2.5. Other approaches.
So far, we have reviewed most of the water distribution
analysis methods which are relevant from the practical point of
view.
Other methods have been developed as a result of the active
research carried out in this field during the last decade, but
they have not been widely used. Only for the sake of
completeness, we shall mention some of these methods here.
2.5.1. Optimisation-based methods.
Collins et al. (1978) presented a completely different approach
to the solution of the network analysis problem, formulating it
within a non-linear optimisation framework. They presented two

alternative formulations of the non-linear problem, one
minimising a "content" function, which is in essence the
expression of the power dissipation within the network, subjected
to the nodal balance equations as equality constraints. The
second alternative formulation consisted of minimising the "co-
content", which is the dual of the "content".
The work by Collins et al. (1977, 1978 and 1979) is relevant
from the theoretical point of view, since it goes back to the
78
fundamentals of the network analysis problem. From the practical
point of view, the approach has a number of shortcomings when
compared with the traditional methods; for example, it is

difficult to understand for the practitioner, and it is computer
intensive. We believe thatii.is quite a good idea to go back to the
basics, but this should ultimately lead to a set of equations, or
"method", easier to solve and be understood by people working in
the water industry, rather than in the academic environment.
Hall (1976) presents a minimization approach to the network
analysis problem, based on (generalised) geometric programming.
Carpentier et al. (1985) interpret most of the network analysis

methods within an optimisation approach.
2.5.2. Methods based on unsteady state analysis.
The steady state condition can be interpreted as the
equilibrium reached by the unsteady state phenomenon, after a
period of time. Thus, in principle, any unsteady state analysis
method can be used to determine the steady state condition,
provided that sufficient time is given to stabilise the flow.
Nahavandi and Catanzaro (1973), and Fox and Keech (1975),
amongst others, present unsteady state methods and apply them to
solve the steady state problem for pipe networks.
We believe that this is not the best way to approach the steady
state network analysis problem, since the traditional methods are
far more straightforward, demanding in general less computational
resources. This transient approach seems to be only a by-product
of the unsteady state flow analysis. It seems that this kind of
79
technique. should be used only in situations where slow or fast
transient phenomena are the dominant ones.
2.5.3. A method based on a finite element approach.
Collins and Johnson (1975), using an analogy between structural
and pipe systems, and taking advantage of the power of existing
finite element packages to solve large linear systems of
equations, proposed a solution to the network analysis problem

via the finite element approach. In so doing, they used a
linearized flow/head loss relationship, which is updated
iteratively to approximate the real non-linear equation. The
linearized flow/head loss relationship is coupled with the nodal
balance equations, considered as boundary conditions in the
finite element terminology, and a system of linear equations is
assembled and solved. Collins and Johnson (1975) compared the
performance of this method with a Cross nodal algorithm,
reporting a faster convergence for their method.
We are not aware of wider testing and/or use of this approach.
2.5.4. The gradient method of Todini (1979).
Todini (1979) posed the problem of finding the steady state
solution flow of a water distribution network as the minimization
of the power dissipated within the network, subject to the nodal
balance condition as an equality constraint. This can be formally
expressed, with the matrix notation already introduced in section
(2.3.7), as:
min P(Q) = (E)1142) T Q (151)
80
subject to A 21Q = a (152)
where:
allQ1In-1
*21Q2In-1
A 11 = (153)
aNPIQNpin-1
with ai equal to the pipe resistance characteristic parameter,

which depends on the formula and units being used.
Upon transforming the constrained minimization problem into an
unconstrained one, via the Lagrange multipliers technique, Todini
(1979) found that the necessary conditions for the steady state
flow in the network are represented by the simultaneous
fulfilment of the nodal balance and the non-linear head loss/flow
relationship for each pipe, which can be expressed in the compact
form:
[ A li ' A l2 1[Q 1 [Q I
(154)
A21: ° H a
The upper part of (154) represents the head loss/flow

relationship for each one of the pipes in the network, which is
clearly a non-linear function of the flows Q, while the lower
part of (154) represents the nodal flow balance condition. In his
original formulation Todini (1979) considered one source node
only, and in that case the nodal balance condition is redundant
for that particular node (reference node) and its balance

equation can be omitted; then, the lower part of (154) is a
81
system of NN-1 independent linear equations in the flows Q.
Contrar y to what most of the authors in network analysis have

done, Todini (1979) did not attempt to reduce the dimensionality
of the non-linear problem (154), and he kept the augmented vector
(Q 1 H) T as the unknown state vector.
Because of the non-linearity of (154), a direct solution is not
possible and an iterative approach is needed. Todini (1979)
thought of the system (154) as the non-linear system formed by
two sub-systems of imbalance functions:
E(Q11) . Q
(155)
a(Q,H) . Q
where (Q,11) is the head loss imbalance function for each
pipe [the upper part of (154)]:
E(Q,H) = A 11 Q + A l2 H = D (156)
and
a(Q,H) = A21 Q - a = Q (157)
is the lower part of (154) and represents the nodal flow
imbalance function.
In an iterative procedure for the solution of the non-linear
system (154) or (155), at an iteration "i", the functions (156)
and (157) will not be exactly zero, this means that, if Q (i) and
11.• II
a (i) are the flow and piezometric head vectors at iteration 1,
then:
SE(i)= A 11 Q(i) + A l2 H(i) Q (158)

and
sa(i) = A21 Q (i) - a D (159)

where:
82
8E (i) : head loss imbalance per pipe at iteration "i".
nodal flow imbalance at iteration "i".
Because the analytic expressions of the imbalance functions are
known [equation (156) and (157)], it is possible to compute the
head loss and nodal flow imbalances via their respective total
differentials, indeed, on differentiating (156) and (157):
dE = n All dQ -4- A l2 dH (160)

and
da = A21 dQ (161)
then, on approximating dQ and dH as the finite difference between
two successive flows and heads in the iterative process:

d Q = Q( i ) .- Q(i+1) (162)
and
d H = H( i ) - H(i+i) (163)
and approximating the imbalances by the total differentials:

6E(i) z dE (164)
and
sg(i) z da (165)
Then, starting with a certain initial augmented state vector

(Q (0) , 11 (0) ) the head loss and nodal flow imbalances can be
computed through (158) and (159), respectively, and the new

augmented state vector (Q (i+1 ), H (5- 1-1) ) can be updated applying
(160) and (161), followed by (162) and (163), recursively, until
a certain convergence criterion is met.
This is the mathematical foundation of the original Todini's
g radient method. To obtain the recursive numerical algorithm,

Todini (1979) followed a matrix-intensive algebraic approach
83
which is presented below. An alternative derivation is proposed
after Todini's derivation, which may be easier to follow.
a) Todini i s (1979) derivation of the recursive numerical scheme:
On differentiating the non-linear system of equations (154) the
equations (160) and (161) were obtained, these equations can be
represented in a more compact way as:
n A 11 : Al2 d Q d E
(166)
A2 1 : o
[ I * [ -c-1-i I = [ " i
where the right-hand-side vector represents the head loss
and nodal flow imbalances which, in an iterative scheme, can be
approximated by the equations (158) and (159), i.e.:
clE = A 11 Q(i) 4' A l2 H(i) (167)

and
cla = A 21 Q(i) - a (168)
The problem now consists in computing the differentials dQ and
dH [in equation (166)] which, in the iterative scheme, represent
the corrections to be applied to the current approximates Q(i)
and H (i) in order to improve them towards the solution. The
computation of dQ and dH is done analytically by Todini (1979):
1 -1 [dEl
n A 11 1 A 12
= 1
1 (169)
: ° dg
[ A21
where All is computed with the current flow Q=Q (i) .
The inversion of the partitioned matrix in (169) can be carried
out explicitly, as another block-partitioned matrix:
84
-1
n A ll : Al2 [ B 11 1 B12]
,, = : (170)
[ A2 . 0 B 21 : B22
According to Ayres (1974), the partitioned matrices Bij can be

computed as shown below. Note that Al2 and A21 are not square
matrices.
B11 =(1/n)[A11-1 -A1111 A l2 (A21 A 11 -1 A l2 )-1 A21 A11-1]
B22 = - n(A21 A 11 -1 Al2)-1

- (171)
B 12 = A 11 -1 A l2 (A21 A 11 -1 Al2)-1
B21 = (A21 A11-1 Al2)1 A 21 A114
On introducing (171) + (170) and (167) + (168) into (169) the
following is obtained:
dQ=(1/n)(All- 1-A 11 -1A l2 (A21 A 11 - 1A l2 )-4 21 A 11 -11(A 11 Q( i)+Al2H(i))-4-
+ A 11 -1A l2 (A21 A 11 -1A l2 )-1(A21 Q(i)-2) (172)

and
clii = (A21 A 11 -1 A l2 )-1 A 21 A 11 -1 ( A ll Q (i) + A l2 13 (3) ) -
- n (A21 A 11 -1 A l2 )-1 ( A21 Q (i) - 2-) (173)
Assuming that a balanced flow solution Q (i) is available,i.e.:
A21 Q (i) - a (174)

then, the equations (172) and (173) simplify to:
d Q = (1/n)[ Q(i)- A 11 -1A l2 (A21 G-1A l2 ]-1 ) 2 ] (175)
and
d H = H( i) + ( A 21 A 11 -1 A l2) -1 a) (176)
which, when considering (162) and (163), allow us to compute
the flow and head updates:
H (i+1).= -[AZI. A 11 -1 A l2 ]-1 g---] (177)
85
and
Q(1+1)=. f(n-1)/n1 Q( i ) - (1/n) A114 A l2 E(i+1) (178)
Thus, starting from some initial solution (00) , H(o)), the

recursive application of (177) and (178) allows us to update the
augmented state vector, generating a sequence of flow and head
vectors convergin g towards the steady state solution of the
network.
Equation (177) represents a linear system of NN-1 equations in
the unknown piezometric heads H (i+1) , which Todini (1979) solved
using the conjugate gradient method.
Equation (178) allows the automatic generation of an initial
flow solution. To do so, the flow exponent "n" is set to 1 in the
first iteration, which leads to the following balanced flow
Q ( ° )= - A 11 -1 A l2 H(°) (179)
b) Alternative derivation of Todini's gradient algorithm:
Introducin g equations (162) and (163) into (160) and (161), the
head loss and nodal flow imbalances can be computed as:
dE = n A11 Q( i ) - n A ll Q(i+1) 4- A l2 H(i) - A l2 H(i+1) (180)
and
dg = A21 Q (i) - A21 ii(i+1) (181)
Replacing H( i) and Q (i) in (180) and (181) by their

corresponding values computed from the imbalances defined by
(158) and (159), and after reordering:
dE = SE(i) - (1-n) A 11 Q(i) En A11 Q(i+1) + Al2 E (i+1) ] (182)
86
and
dg = 8g(i) +a - A 21 Q(i+1) (183)
Then, on assuming (164) and (165), from (183):
A21 ii(i+1) = g (184)

which implies that the new flow is balanced.
And from (182):
(1-n) A 11 Q(i) + n All Q(i+1) + Al2 H(].+1 ) = 0 (185)
Pre-multi p lying equation (185) by (n A11)-1:
{(1-n)/n} Q (i) + Q (i+1) + (n A11)-1 A l2 11(1+1) = ° (186)

and pre-multiplying (186) by A21:
f(1-n)/n1A21Q ( i) -4- A 21 Q(1-4-1) + A 21( n A 11 )-1A l2 H(14-1) = ° (187)
Upon introducing (184) and assuming that the current flow

vector is balanced (i.e. A21 Q (i) = g), equation (187) simplifies
to:
[-- [ A-
21 A 11 -1 A l2 ] H(i+1) = - g (188)
which is the same as equation (177).
The flow updates are obtained from (186):
[ Q (1- 4-1)= {(n-1)/n} Q( i) - 0.1 A10-1 A l2 H(i+1)

I (189)
which is the same as equation (178).

2 6. Comparison of the Performance of some of the existing
methods: need for a more reliable algorithm.
Despite the fact that a number of convergence problems have
been reported for almost every single algorithm already reviewed,
only a few attempts to compare most of these algorithms in a

systematic way have been reported; these are the study carried
out by Wood (1981a) and a comparison performed by Carpentier et
al. (1985), which are complementary. We do not attempt here to
make a full comparison, but only to review the results already
published to draw some preliminary conclusions, which may allow
us to narrow the search for an efficient and stable network
analysis solver. A comparison between some of the most promising
algorithms and our proposed gradient method is delayed until the
end of the next chapter.
The comparison carried out by Wood (1981a) considered five
different methods: the loop and nodal Cross' methods (single path
and single node adjustment methods in Wood's terminology), the
nodal and loop Newton-Raphson approaches (simultaneous node and
simultaneous path adjustment methods in Wood's terminology) and
the gradient version of the linear theory method.
Wood and his collaborators at the University of Kentucky
prepared FORTRAN programs for all the five methods compared, so
that the computer codes used were similar from the programming
point of view. All the programs were designed to work with the
minimum required amount of data (basic branch to node
connectivity data), assembling the equations and generating the
initial conditions internally within the programs. Sparse
88
techniques for matrix manipulation were used in all the programs.
A set of test examples was assembled, with 30 systems of under
100 pipes and 21 systems with more than 100 pipes. Changes in the
data (diameter, roughness, or length) allowed the researchers to

generate two extended sets of test examples of 60 and 31 cases,
for networks with less and more than 100 pipes, respectively.
A detailed comparison of the performance of the different
methods was carried out for all the systems with less than 100
pipes. All the cases where convergence to a pre-specified
accuracy was not achieved, within a fixed number of iterations,

were labelled as failures and their occurrence is summarised in
Table 2.1.
The only failure found with the simultaneous path method
consisted of a case where a pump operating with a steep
Table 2.1. Results of the comparison of 60 small water

distribution networks (less than 100 pipes),
from Wood (1981a).
Network analysis method Failures
Single path adjustment (looped Cross method) 8
Single node adjustment (nodal Cross method) 51
Simultaneous path adjust. (looped Newton-Raphson) 1
Simultaneous node adjust. (nodal Newton-Raphson) 18
Linear theory method (gradient version) 0
89
characteristic curve was considered. No failure whatsoever was
detected with the linear theory method.
In view of the previous results, the systematic comparison for
larg er networks was limited to the simultaneous path and linear

theory methods only, and in all the 31 cases tested both methods
managed to converge within a pre-established number of
iterations. Althou g h a systematic comparison for the single path,

single node and simultaneous node adjustment methods was not
carried out for the larger networks, a number of failures was
found for these methods, confirming their exclusion from the
systematic comparison.
The main conclusions of the Wood (1981a) study, also published
by Wood and Rayes (1981), can be summarised as follows:
a) The single path, single node and simultaneous node
adjustment methods failed to converge in a number of cases and,
as a result, they are not recommended for water distribution
network analysis, unless they are used for known well-conditioned
problems. Convergence problems related to the nodal Newton-
Raphson method have also been reported by de Neufville and Hester
(1969).
b) Although no guarantee of convergence can be given, the
simultaneous path adjustment (looped Newton-Raphson) and linear
theory method (gradient version) exhibit excellent convergence
behaviour. Only in one very special case did the simultaneous
path adjustment method fail to converge. Both methods are
recommended by Wood (1981a) for the analysis of water
distribution networks.
90
c) The computational costs of the simultaneous path adjustment
and linear theory methods are very similar. Average run-times of
5.4 and 6.9 sec. were found in the case of the larger networks,
for the linear theory and simultaneous path adjustment methods,
respectively. Thus, the linear theory method needs 78% of the

time required by the simultaneous path method.
It should be remembered that the storage requirements of the
simultaneous path adjustment method are smaller than that
required by the linear theory method, since, in the former, a
symmetric matrix is obtained, while, in the latter, we get a non-
symmetric one. This parameter was not included in Wood's (1981a)

comparison.
• Focusing on the difference between the simultaneous path and
linear theory method, we have to bear in mind that both methods
were not starting from exactly the same initial solution. A
balanced initial flow distribution was produced by the
simultaneous path program, while a flow distribution computed on
the basis of a uniform velocity of 4 fps. was used in the linear
theory program. Hence, the results of the relative performances
of the simultaneous path and the linear theory methods may be
different, if both programs are started exactly at the same
initial solution.
Carpentier et al. (1985) presented a numerical comparison of
some of the network analysis methods: looped Cross method, looped
Newton-Raphson, nodal Newton-Raphson and the hybrid mesh-nodal
method of Hamam and Brameller (1971). They also looked at all
these methods from a mathematical programming perspective, in
91
order to establish some theoretical differences.
Although the three examples of Carpentier et al. (1985) are

very limited in comparison with those used in the study carried
out by Wood (1981a), some interesting conclusions can be drawn,
particularly in connection with the mesh-nodal method. Only three
methods were found to exhibit an adequate convergence, the Cross
looped method and a quasi-Newton implementation showing very low
convergence. The results for those best performing methods are
shown in Table 2.2.
Unfortunately, Carpentier et al. (1985) did not give any
information about the initial solution (heads and flows) with
which the different programs were run.
Table 2.2. Comparison of network analysis methods in terms of

global and relative CPU times, and number of
iterations, from Carpentier et al. (1985).
Global and relative CPU times (sec.), No. of iterations
NETWORK CHARACTERISTICS
METHOD 46 pipes 212 pipes 443 pipes

40 nodes 148 nodes 297 nodes
10 loops 70 loops 157 loops
Looped Newton-Raphson 2.00 sec. 18.00 sec. 63.90 sec.

(1.00)* (1.00) (1.00)
5 iter. 7 iter. 8 iter.
Nodal Newton-Raphson 1.15 sec. 5(6712r c. 13.00 sec.

(0.58)* (0.20)
Mesh-nodal 0.76 sec. 3.80 sec. 8.40 sec.

(0.38)* (0.21) (0.13)
(*) In parenthesis is th e relative performance: ratio with

respect to the Looped N ewton-Raphson, for the same example.
92
The main conclusions that can be drawn from the comparison by
Carpentier et al. (1985) are the following:
a) The mesh-nodal method of Hamam and Brameller (1971) appears
to be the most efficient one in terms of global CPU time. It is
also the best in terms of CPU time for initialisation (not shown
in Table 2.2).
It is interesting to note that when the mesh-nodal method is
compared with the looped Newton-Raphson method, as Wood (1981a)
did with his linear theory method, the mesh-nodal method requires
38% of the CPU time needed for the looped Newton-Raphson for
networks with less than 100 pipes (see Table 2.2), whereas the
linear theory method required 78% of CPU time
Although the programs used were not exactly the same, this
app ears to suggest that the mesh-nodal performs faster than the
linear theory method. The results are even more favourable for
the mesh-nodal method when a larger network is considered,
reaching 13% for the 443 pipes example (see Table 2.2.).
b) The second conclusion of Carpentier et al. (1985) is
concerned with the comparison between the nodal and loop
approach. According to these authors, when analysing the
continuity and differentiability of the Hessian matrix produced
by the different methods (in the mathematical programming
framework), the looped (flow) methods seem to be better
conditioned than the nodal (head) methods; this coincides with
what some authors have previously found. This conclusion is in
contradiction with Table 2.2, where the nodal Newton-Raphson
93
clearly requires less time to converge than the looped Newton-
Raphson, although the number of iterations are the other way
around. Our explanation of this is as follows.
We can put the problem in simple terms, by observing the way in
which the head loss/flow relationship and its derivative is
handled in the nodal or looped approach:
Looped methods consider: hij = Hi - Hj = Kij Qij n(190)
Nodal methods consider: = R •• (Hi _ H)l/n (191)
Because all the Newton-Raphson methods consider (in one way or
another) the derivatives of functions including either hij or Qij
(looped or nodal methods, respectively), we should analyse the
derivatives more carefully:
6 hij
Looped methods : = n Kij Q1j n-1(192)
6 Qij
6 Qij 1 (1-n)/n
Nodal methods : - (Hi - Hj) (193)
6 Hi n
which can be re-written as:
6 Qij Rij
= (194)
6 Hi n (Hi _
On comparing (192) and (194), we can easily see that the
derivatives needed in the looped (flow) methods are continuous
for whatever flow we consider, whereas in the case of the nodal
methods, equation (194) "blows up" when Hi is exactly Hj. This
would explain why the nodal methods are found to be more trouble-
prone than the looped or flow-based methods (see Table 2.1).
94
1
The results in terms of global CPU time in Table 2.2 can also
be exp lained in terms of equations (194) and (192). In fact, if
0 [i.e. equation
(194) is pot "blowing up"], the derivatives for the nodal methods
[equation (194)] are, in general, "stronger" than those for the
looped methods. Indeed, if we think of a typical troublesome
case, with a pipe of low head loss (i.e. Hi Hi, but not close
enough to cancel), the derivatives given by (194) will be quite
large and the diagonal terms of the corresponding Jacobian matrix
will be even larger, increasing the well-conditioning of the
linear system of equations; clearly, for such problems, a nodal-
based program will converge faster than a looped program. If we
consider another troublesome case, for example a pipe with high
resistance (e.g. small diameter), it will produce a "small"
derivative in a nodal approach [equation (194)], whilst the
corresponding derivative in a looped approach [equation (192)]
will tend to zero, since the flow is "small", thus spoiling the
conditioning of the loop method. In the latter case, the nodal
approach can still manage to converge. This shows clearly that,
in some problems, one method can perform better than others,
whereas, in other cases, the reverse can be true. This also
suggests that, when the nodal methods do converge, they do it
faster than the looped (flows) methods, as shown in Table 2.2.
This has been corroborated by other authors as well [Collins et
al. (1978)].
Looking at equations (194) and (192) from a strictly numerical
point of view, the structure of the derivative for the nodal
95
methods [equation (194)] is the worst we can expect, because it
is subjected to cancellation errors and also to unbounded values
when the heads Hi and Hi are close enough. Expressions of this
kind are normally avoided in numerical analysis [see for example
Appendix C, section C.1, when preferring a recursive approach for
the evaluation of B-splines, instead of the explicit formula].
In summary, from what we have seen so far in this chapter, it
is clear that the methods derived by Cross (loop and nodal
methods), are not reliable enough and should not be used in
computer-based water distribution network analysis. The same can
be said of the nodal Newton-Raphson approach which, as documented
by Wood (1981a), fails to converge in a number of cases.
The looped Newton-Raphson method failed to converge in one case
in the review carried out by Wood (1981a). Although this method
should still be considered as a possible solution scheme, its
results should be carefully analysed.
The linear theory method of Wood and Charles (1972) should not
be used since, as shown by Collins (1980),its behaviour is
inherently unstable for quadratic or nearly quadratic head loss
functions. The "tricks" to overcome instabilities, like averaging
flows and so on, may work for some cases, but there is no
guarantee that they will work in others. Wood himself (1981a)
recognised that the improved gradient-based linear theory method
has much better convergence properties, although, even in that
case, he insisted that no guarantee of convergence could be
given.
96
In the limited comparison carried out by Carpentier et al.
(1985), the mesh-nodal approach of Hamam and Brameller (1971)
showed some impressive results with respect to the looped and
nodal Newton-Raphson methods. These results suggest that the
mesh-nodal method should be competitive with the gradient-based
linear theory method, although further testing is needed.
As far as the non-traditional methods of network analysis are
concerned, such as those based on explicit non-linear
optimisation methods, unsteady flow analysis and the finite
element method (section 2.5), they do not seem to be relevant
from the practical point of view, particularl y when the network
analysis algorithm has to be linked with design or operational
optimisation . routines. All the traditional methods reviewed in

this chapter are far more straightforward than non-traditional
ones.
As a result, there is no conclusive evidence that any of the
most promising methods will guarantee a solution to the analysis
problem. On top of that, most of the comparisons and systematic
testing of the methods have been made considering networks with
pipes and pumps (at most), without including other devices like
p ressure reducin g valves, pressure sustaining valves, etc., which

will undoubtedly test the different methods to the limit. The
ability of the methods to cope with extreme operational
situations, like those when the network becomes disconnected,
either due to the action of non-return valves or pressure
regulating valves, has not been tested in the open literature
reviewed so far, and we intend to do so in the following
97
chapters.
At this stage, there is sufficient evidence to show that the

search for efficient network analysis algorithms should be
extended even further. Perhaps the issue of the reliability of
the methods should be at least as important as the computational
efficiency, if it is planned to use the network analysis
algorithm for operational or design optimisation routines.

CHAPTER THREE
GENERALISATION OF THE GRADIENT METHOD TO INCLUDE PUMPS
3.1. Introduction.
In the previous chapter most of the traditional methods for
network analysis have been introduced, spanning the early work of
H. Cross (1936), the application of the Newton-Raphson technique
in the sixties and the development of the linear theory method in
the seventies.
In this chapter we intend to review some fundamental aspects of
steady state flow network analysis, and to explore an alternative
formulation, adopted here which is referred to as the gradient
method; we extend this method in order to include pumps, using a
general head loss/flow model. In the final part of this chapter,
we compare the gradient method with other similar techniques and
some conclusions are drawn from that comparison.
3.2. Traditional formulation of the network analysis problem.
Most of the traditional methods for water distribution network
analysis are based on the two "conditions" specified by H. Cross
(1936): flow continuity and potential continuity; the former
establishes mass balance at the node level, while the latter has
been thought of by some authors as a loop balance condition
(algebraic summation of all the head losses round a loop = 0).
Other authors think of potential continuity simply as
establishing that the piezometric heads at each node are unique.
99
Indeed, as shown in Chapter Two (Section 2.3.7., equations 90
and 91), the potential continuity condition (loop balance) can be
achieved for any set of piezometric heads, and is due to the fact
that the product of two topological matrices: the loop to branch
matrix A31 and the branch to node incidence matrix Al2 is zero,
i.e.
A31 A l2 Li = 0 , V H
due to:
A31 Al2 = 0
Most of the methods of network analysis have been based, either
explicitly or implicitly, on the acceptance of the loop balance
condition and, in so doing, some of them do not seem to have been
based on the fulfilment of what is presented as a physical law,
but on a topological property.
As a result, it seems necessary to re-establish which are the
physical laws controlling the steady state flow in a water
distribution network or, in mathematical terms, which are the
necessary and sufficient conditions for the steady state flow in
pipe networks. Only a clear understanding of the fundamental
-principles can lead to the development of a more robust numerical
scheme for the network analysis problem.
3.3. Minimum Power dissipation formulation: necessary conditions.
Several authors have addressed the problem of establishing the
fundamental principles governing the steady state flow in water
distribution networks, agreeing in that, apart from the flow
continuity law at each node (mass conservation law) , it is a
minimum energy dissipation rate principle (minimum dissipated
100
power) which controls the steady state flow in a water
distribution network. Basically, this means that the steady state
flow stabilises in a situation wherein a minimum power is
required. This "minimum effort" law is common to other physical
phenomena of interest in civil engineering. By formalising the
minimum power dissipation approach, we can also study the
mathematical conditions for the existence and uniqueness of the
steady state solution.
Collins et al. (1978, 1979), using previous results by Millar
(1951), formulated the steady state flow analysis, both of pipe
and channel networks, within a minimisation context. They
minimised a "content" function, which is basically the power
dissipated within the system, subjected to the flow continuity at
each node as an equality constraint. This approach allowed
Collins and collaborators to apply well-known minimisation
algorithms to solve the network analysis problem.
Todini (1979) posed the network analysis problem as that of
minimising a general function E(Q) = e 1Q1n+1, subjected to the
nodal flow continuity as a set of linear equality constraints.
Upon transforming the constrained minimisation problem into an
unconstrained problem, via the Lagrange multipliers technique,
and on introducing the head loss/flow relationship for each pipe,
he arrived at the conclusion that the only necessary conditions
for the steady state flow condition are simply the simultaneous
fulfilment of the head loss/flow relationship for each pipe, and
the flow continuity at each node. Based on this, Todini (1979)
introduced a new iterative algorithm for network analysis,
101
obtained via the application of the gradient operator both to the
flows and the heads. We refer to this algorithm here as the
gradient method.
Subsequently, Pilati and Todini (1984) extended the original
formulation of Todini (1979) to the case when laminar and
turbulent flow co-exist in the same network (i.e. the flow
exponent "n" in the head loss/flow relationship is variable from

pipe to pipe).
Yang and Song (1979), Song and Yang (1980) and Song and Yang
(1982), have applied the minimum rate of energy dissipation
principle to other hydromechanical systems.
We have adopted the gradient method introduced by Todini (1979)

and Pilati and Todini (1984) as the basic network analysis
algorithm for the work reported in this thesis and, because it
was originally limited to networks containing pipes only, we have
pursued the objective of incorporating other devices into it. In
the rest of this chapter, we show how to include pumps in
Todini's gradient algorithm and, in the next chapter, we deal
with how to model pressure regulating valves in the context of
the gradient method. At the end of the present chapter, a
comparison with other algorithms is carried out.
3.4. Todini's gradient method.
Todini (1979), and Pilati and Todini (1984), demonstrated
mathematically that the necessary conditions for the steady state
flow in water supply networks are the simultaneous fulfilment of
102
the nodal balance and the non-linear head loss/flow relationship
for each pipe. This can be ex p ressed in a compact format, with
the notation introduced in Chapter Two (Section 2.3.7.), as:
1 A 11 : A l2 11 Q 1 1 -A 10 HO 1
(3)
I_ A91 : 0 i 1 H
[ g
The upper part of the system of equations (3) represents the
head loss/flow relationship, whereas the lower part corresponds
to the flow balance at each node. Ho corresponds to the (NS*1)
vector of known piezometric heads (source nodes).
The matrix All depends on the particular head loss/flow

relationship used, Al2 is the topological matrix containing the
branch to node connectivity information for the non-source nodes,
and Alo is the equivalent branch to node matrix for the source
nodes, as defined in Chapter Two (Section 2.3.7.). A21 is the
transpose of Al2.
Because All depends on the flows, the upper part of (3) is a
set of non-linear equations in the flows. Todini (1979), and
Pilati and Todini (1984) concentrated on the solution of the non-
linear system (3), applying the gradient operator to the extended
vector of unknowns (Q, H), without attempting a reduction in the
dimensionality of the problem, like other authors who dealt
either with the heads or the flows. This leads to a linear system
of equations in the corrections dQ and dH which, after some
algebraic manipulations, can be expressed as a coupled set of
equations, solved iteratively for the unknown heads and flows.
103
The gradient method fully recognises the fact that all the
topological information, concerning the structure of the network,
is contained in the branch to node incidence matrix Al2 and,
consequently, no loop information is necessary.
Because the upper part of equation (3) incorporates the head
loss/flow relationship, we need to specify a function as general
as possible, which is able to accommodate any kind of device
normally present in a water network.
3.5. A general head loss/flow model to include pumps and
valves in the gradient method.
We have already used in Chapter Two a general model to describe
the non-linear behaviour of pipe flow; this model was:
= Hi - Hj = Kij Qij n(4)
where
: head loss in pipe joining nodes "i" and "j"; a component
of the (NPx1) head loss column vector h
NP : total number of pipes (branches, in general) in the
network.
nodal piezometric head; a component of the (NNx1)
piezometric head column vector H.
NN : total number of nodes in the network.
resistance of pipe joining nodes "i" and "j", dependent
on the particular formula in use.
Qij : flow in pipe joining nodes "i" and "i"; a component of
the (NPx1) flow column vector Q.
flow exponent, dependent not only on the formula being
used, but also on the hydraulic regime in which the pipe
104
is operating (laminar or turbulent).
Equation (4) represents a nearly quadratic function, which
passes through the origin in the flow/head loss plane. The model
(4) allows us to represent other kind of devices, like gate
valves, butterfly valves, etc., with the condition that their
characteristics remain known and fixed during the modelling.
Other situations, like the introduction of minor losses due to
fittings (bends, tees, reductions, etc.) can also be handled with
equation (4). In these cases, all we need to know, for any
particular device, is the value of"K-1 J- " and "n", which are
normally supplied by the manufacturer or obtained from laboratory
tests.
To introduce centrifugal pumps, other non-linear models have
been proposed in the literature, relating the head gain with the
flow, through quadratic or higher order polynomial functions as:
= Etc, + al Qij + a2 Q i j 2(5)

or
hij = etc, + al Qij + a2 ci ii 2 4. a3 Qii3 (6)
or
= etc, + al Qij + a2 ciij 2 4. a3 Q ij 3 .1.a4 Qij4 (7)
All these models allow us to take into account different types
of pump characteristic curves. Also, piecewise linear
representations have been used in computer models, joining known
head/flow points of the characteristic curve with straight lines.
It has been argued that the higher order models (third or
fourth order), would represent the characteristic curve of a pump
in a more accurate way. However, other authors have put the
emphasis on using accurate data on the pump characteristics,
105
rather than concentrating on a more mathematicall y sophisticated
model.
Due to its simplicity, and because of other considerations
which have to do with the shape of the characteristic curve of a
pump, we have adopted the following model:
h•ij •n +
• = a- • Q• li 15-7-1
li (8)
where the parameters aij, 13ii and "n" have to be determined
from data on the real performance of the puthp.
Upon adopting (8), we see that the model of a pipe (equation 4)
is just a particular case of the model (8) (taking Bij=0); hence,
equation (8) can be considered a general head loss (gain)/flow
model, applicable to most of the devices normall y found in water
distribution systems.
Another advantage considered when adoptin g the model (8)
relates to the fact that, when modellin g centrifugal pumps,
equation (8) automatically avoids the inclusion in the network
model, of pumps with an ascending limb in their characteristic
curves. As we shall see later on, such pumps should not be
included in the model for stability reasons. The computer program
checks that the pump is always working with positive flows and
positive head gains.
Other devices, like check valves, pressure reducing valves,
pressure sustaining valves, etc. will be dealt with in the next
chapter, within the framework provided by equation (8).
106
3 6 Extension of the gradient method to include PUMPS:
derivation of the recursive scheme for the extended method.
Using the link to node topological matrices for non-source and

1
source nodes (Al2 and A 1 0, respectively) , we can re-write the
head-loss (or head 'gain' in the pumps case) of each link

connecting two different nodes as:
A ll Q A l2 H = - A 10 HO (9)
where now
H : is the ((NN-NS)xl) column vector of unknown heads.
Ho : is the (NSx1) column vector of fixed (known) piezometric

heads (reservoirs).
and where, upon using the g eneral head loss (gain)/flow model
(8), All can be re-defined as:
El l in1-1+ 131/Q1
*11
* 21 Q21 n2-1+ 132/Q2
A 11 = (10)
5NP/NP
On coupling equation (9) with the flow (mass) nodal balance

condition, we still get equation (3), which is reproduced here:
A ll : Al2 -A10 HO
A21:0 lu a
107
Because the system represented by (3) is a non-linear one, a
direct solution is not possible, and an iterative procedure must
be followed to determine the unknown flows and heads.
On differentiating the system of equations (3) we get:
[ N A 11 * : Al2 [dQ1 [dE]

1 *
=
A21 0 d H d g
where:
dE : is the pipe head loss imbalance produced by a pipe flow
differential dQ and to a nodal piezometric head

differential dH.
dg : nodal flow imbalance due to a pipe flow differential dQ.
c'11Q1In1-1 13 1/ Q1
Q21n2-1
c<2I 132/Q2
A ll * = = A n - (12)
aNpAipInNP-1
13NP/QNP
N = (NPxNP) diagonal matrix of the exponents "n" of the

head loss-flow relationship.
At some iteration "i", when convergence has not yet been
achieved, the (NPx1) vector dE, representing the head loss

imbalance at each branch, and the ((NN-NS)xl) vector dg,
rep resenting the nodal imbalance, can be approximated as:
d E = A n" ) Q (i) Al2 H (i) +A 10 HO (13)

and
d g = A21 Q (i) - 2 (14)
where Q( i ) and HU) represent the pipe flow and the nodal heads
108
at the iteration "i". A11") stands for the matrix A ll evaluated
at Q=Q (i) , but from now on we drop the iteration superscript in
the matrices All and A11* for the sake of notational simplicity.
The algebraic solution of (11) is:
-1
[1:1Q1 N All* : Al2
(15)
d H A2 0 dg
where A1 1 * is evaluated at
The inverse of the partitioned matrix in (15) can be computed

as another block-partitioned matrix:
I -1
A 11 * : Al2
N [ B 11 : B 2 1
1
I = : (16)
A21 ' 0 B21 I B22
[
On using the following notation:

G = N A11* (17)
We can compute the blocks of the inverse in equation (16)

explicitly, according to Ayres (1974), as:
B 11 = G-1 -G-1 A l2 ( A21 G-1 A l2 )-1 A21 G-1

B22 = - (A21 G-1 Al2)-1
(18)
B 12 = G-1 A l2 (A21 G-1 Al2)-1
B21 = (A21 G-1 A l2 )-1 A 21 G-1
With the partitioning (16), the system (15) can be re-written

as:
d =B11 dE -I- B 12 dg
(19)
d H = B 21 dE + B22 dg
109
On introducing e quations (13) and (14), and the results of
(18) into equation (19), we get:
d Q = [I- G-1 A l2 (A21 G-1A l2) -1 A 21 ]G-1 (A llEi (1)+A l2H " )+A 10 HO ) +
+ ( G-1 Al2 (A21 G-1A l2 )-1) ( A21 ii(i)-g) (20)
Similarly:
d H = [ ( An G-1A l2) -1A21 G -1 1(A1 1 Q (i) + Al2 H (i) + A 10 HO ) -
- (A21 G-1A l2 )-1(A21 i (i) - a) (21)
The system of equations (20) and (21) can be solved for dQ and
dH via an iterative procedure, on defining:
d Q = Q( i ) - Q(i+1) (22)
and
d H = H(i) - H(i+1) (23)
And upon introducing (22) and (23) into (20) and (21), we
obtain:
from (23):
H(i+1)= -[A 21 G-1 A l2 ]-1 f A 21 G-1 (A ll Q(i) + A 10 HO ) -

- (A21 Q(i) - g) / (24)
from (22):
Q(i+1).= (I- G1 A 11 ) Q(i) - G-1 ( A l2 H(i+1) + (25)

A 10 HO)
Equation (24) represents a linear system of (NN-NS) equations

in the unknown piezometric heads, and it can be reordered into
the standard format of a linear system:
[ A21 G-1A l2] 11(i+1)= - f A21 G-1 (A 11 Q(i) + A 10 HO ) -

- (A21 (i) - g) 1 (26)
110
Starting with any flow distribution, not necessarily
balanced, and applying recursively the coupled system (26) and

(25), the problem can be solved. Note that once a balanced
solution has been obtained (second iteration and following), the
right-hand-side vector of equation (26) simplifies, since the

difference (A21 Q (i) - a) vanishes. Equation (25) guarantees the
nodal balance, as it can be proved via pre-multiplying (25) by
A21, which gives, after some algebra, the vector of nodal

consumptions g.
Figure 3.1 represents graphically the main steps of the
gradient method's numerical algorithm. The main distinctive
characteristic of this algorithm when compared with the rest of
the existing methods, is the fact that it is based on the
gradient operator being applied over both the heads and the
flows, and not only to one of them, as in all the methods
reviewed in the previous chapter.
The fundamental steps in the derivation of the gradient

algorithm can be summarised as follows:
a) The system of non-linear equations (3) are the necessary
conditions for the steady state flow in the network. The
sufficient conditions are dealt with in the following section

3.7.
b) The non-linear system (3) can be linearized by applying the
gradient operator to (3), in the space of the flows and heads,
leading to the linear system (11), in dQ and dff.
111
( start (i=0) )
4
Read data:
- General data: title, no. of links, nodes and
sources, convergence criterion, formula,
printout options.
- Link data: initial and final nodes, length,
diameter, roughness, type of link, a, 13.
- Nodal data: type of node, consumption, ground
level, fixed heads.
4
Assemble matrix of coefficients: A
= [A 21 G-1 Al2]
Assemble right-hand-side vector: h
f A 21 G-1(A 11 Q " )+A 10 11 0 ) - (A 2 Q(i)-a) 1
Set
Q(i)=1
or other
Solve equation (26): A 11( i + 1 ) = 12
Update flows: equation (25)
Q(].+1) = (1-G-1 A ll ) (i) - G-1(Al2 H (i+1)+A1 HO )
Print results
no Stop
= i+1
Figure 3.1. Flowchart of main steps in gradient method.
112
c) The system of e quations (11) can be solved explicitly by
inverting the partitioned matrix, producing the system (19)
which, after some algebra, leads to the coupled equations (26)
and (25).
d) The most computationally expensive step of the algorithm now
becomes the solution of the symmetric positive definite linear
system (26). Chapter Five deals with the efficient computer
solution of the system (26). Equation (25) is simply the product
of matrices and vectors and no linear system needs to be solved.
The inversion of matrix G is straightforward, since it is a
diagonal matrix.
e) Iterations: starting from any initial flow distribution,
equations (26) and (25), in that order, are applied recursively
until some convergence criterion is met; see Figure 3.1.
3.7. On the existence and uniqueness of the network analysis
solution.
It has been said earlier that the set of non-linear equations

(3), coupling the simultaneous fulfilment of nodal balance and
the head loss/flow relationship for each branch, are the
necessary conditions for the solution of the steady state flow in
water distribution networks. The sufficient conditions for that

problem remains to be defined.
Since the early stages of the study of network analysis, the

question of whether the solutions provided by the different
methods are unique, or whether there are multiple solutions, has
been considered by different authors.
113
From a mathematical point of view, since the solution to the
problem is in fact the solution of a set of non-linear
simultaneous equations, it seemed in the past quite natural to
think that the network analysis problem might have more than one
solution (more than one real solution, because the presence of
solutions in the complex plane can be discarded using a physical
argument).
Most of the work done in this area has been adapted from work
originally carried out for non-linear electrical networks, using
the well-known analogy between electrical and pipe networks.
D'Auriac (1947) proved that, in the case of a network of pipes
only, wherein the resistance function h = f(Q) is an increasing

function (-m < Q < m ), the solution for the steady state flow is
unique. D'Auriac suggested that this is not true when dealing

with a network containing pumps.
Warga (1954), on extending previous work by Duff in, provided a
set of sufficient conditions for the uniqueness of the solution
of the set of nodal balance equations (first Kirchoff law):
E. Qij = qi i=1,2,. ..,NN (27)

J
where:
= fii(Hi-Hi): flow between nodes "i" and "j", a
function of the head loss.
piezometric head at node
q i : nodal demand at node "i".
Warga's sufficient conditions can be summarised as:

a) f (x) = - f1(-x) : h is symmetric with respect to the
114
origin.
b) f(x) is continuous, for all "x"
c) f(x) is a non-decreasing function of "x".

d) it is possible to connect each node with the reference node
(which has a known piezometric head), via at least one
series of branches whose characteristic function fij is
strictly increasing and unrestricted (i.e. it can carry an
unlimited flow).
The presence of pumps doesnot satisfy condition c) of Warga,
since their characteristic curve is normally decreasing or, even
worse, in some cases pumps have characteristic curves with partly
increasing and partly decreasing limbs. Condition d) allows the
presence of some decreasing and restricted branches in the
network. Thus, devices such as check valves, allowing flow in
just one direction, can be present in the network and uniqueness
is guaranteed, provided that we can link a reference node with
any other node, via strictly increasing and unrestricted
branches.
Millar (1951), upon studying non-linear electrical systems,
defined the concepts of "content" and "co-content" which,
expressed in terms of their hydraulic analogue variables, can be
written as:
Content: G = f
0
h dQ (28)
hij
Co-content: J = Qij dh (29)

0
115
Both the content, and its dual, the co-content, have units of
power and they are actually proportional to the dissipated power.
From a geometrical point of view, the content and co-content
functions can be interpreted as the area underneath and above the
h-Q curve. Millar (1951) demonstrated that both quantities
(content and co-content) are constants when steady state flow
occurs, i.e.:
dG dJ
— = — = 0 (30)
dt dt
These content and co-content concepts have been used later on

by Collins et al. (1978, 1979), since they form the basis for
establishing the necessary conditions for the steady state flow
in a non-linear network.
Birkhoff and Diaz (1955-1956) proved the existence and

uniqueness of the solution of the steady state flow in a non-
linear network, under some particular conditions. They developed
their results within a general framework, including electrical,
thermodynamic and fluid flow problems. In the case of water
distribution networks, where the head-loss flow relationship is
expressed as:
Qjj = R ij sign(H i -H) 'Hi-Hi' 1/n, (31)
Birkhoff and Diaz (1955-1956) proved that, if Rij (the

conductance function) is continuous and strictly increasing, and
if some nodal heads are fixed (e.g.: reservoir levels), and when
some nodal demands are also fixed (or even when they are non-
increasing functions of h), then, the steady state flow problem
116
does have a unique solution.
In addition, Birkhoff and Diaz (1955-1956) determined that the

nodal balance (first Kirchoff law) holds, if R i j is continuous
and strictly increasing and, if and only if the function:
V(h) = E f Q(h) dh (32)

0
has an absolute minimum. The summation is carried out over all

the links of the network. This function has again, as in the
cases of the content and co-content functions, units of power and
it is similar to equation (29) of Millar.
Furthermore, Birkhoff and Diaz (1955-1956) proved that in the

case of exponential conductivity ,which is the case of water
distribution networks, and for (1/n) = constant for all the pipes
[see equation (31)], then the function V(h) is proportional to
the rate of energy dissipation (or power dissipation) D, where:
D =EQh (33)
Then, the previously mentioned results of Birkhoff and Diaz

(1955-1956) are also valid if V is replaced by D.
Even though there is no proof of the existence and uniqueness

of the solution when pumps are present, Millar (1951) and
Birkhoff and Diaz (1955-1956) have introduced a new factor in the
mathematical representation of the steady state flow in a
distribution network, that is the rate of energy dissipation
function or power dissipation.
117
In the 1960's and 1970's, the question of the existence and
uni queness of the solution remained without an explicit
mathematical proof. As pointed out by Shamir and Howard (1968)
and Shamir (1973), in the context of their nodal approach

(Chapter Two, Section 2.3.4.), this situation happens when:
i) Some nodal consumptions and branch resistances are part of
the set of unknowns of the problem.
ii) A pump with a non decreasing characteristic curve is

included in the network.
Lekane (1979) also recognised the lack of mathematical proof
for the uniqueness problem, except for the case of a network
including pipes only. Lekane expressed the view that, in a
network including pumps, check valves and pressure reducing
valves (P. R. V.'s), "the experience has shown" that the network
analysis problem does converge to a unique solution.
Collins et al. (1978,1979), using the results of Millar (1951),
formulated the problem of determining the steady state flow as
the minimization of the content function G(Q) , subject to the
nodal balance equations as equality constraints. An alternative
formulation based on the co-content function was also given. They
used an expanded directed graph of the network, introducing an
artificial one-way element to simulate two-way elements (like
pipes e.g.); also they introduced fictitious one-way elements to
link real reservoirs with the reference (ground) node. Then, the
objective content function to be minimised is:
Qij
øj. dx ] - E Hn Qgn + E Hn Qng (34)

Zi,j )E Ef E
G(Q) )EEL (g,n (n,g)EEL
118
subject to the constraints:
qn for all nodes "n" (35)

(n,j) d ni -(i ,IFI )E2 1n = n=1,2,..., NN
and
Qii 2 0 for all links and nodes (i,j) (36)
where:
E : is the set of all ordered pairs of nodes (i,j) of the
expanded network, which is partitioned into two sub-
sets: Ef and EL, the former corresponding to all real
nodes (i.e. those connected to real elements like
pipes, pumps and valves), while the latter corresponds
to those nodes connected to artificial (frictionless)
links, created by the expansion of the network and
linking real reservoirs and ground nodes.
Oii(Q): is the head-loss flow relationship, Hi-Hi = Oij(Qij)
Hi : piezometric head of node "i"

*
Hn : fixed piezometric head at reservoir node "n".
Qij : flow in element connecting nodes "i" and "j"
qn : demand on the node "n", negative if outflow.
g : subindex corresponding to the reference (ground) node.
Equation (35) is simply the nodal balance equation, and
constraint (36) is compatible with the expanded directed graph
where all the flows are assumed to be positive (i.e. in the pre-
specified direction, since, if a pipe is known to work in two
ways, a second one-way pipe has been introduced).
Following Collins et al. (1978), the condition for having a

unique solution in the non-linear optimisation problem
119
represented by equations (34), (35) and (36), is that G(Q), the
content function given by equation (34), must be a strictly
convex function.
Using a small example network with four pumps and one check
valve, where, due to directional and flow constraints, the
content function becomes a single variable function, Collins et
al. (1979) showed that the content function will be strictly
convex if each one of its summation terms is strictly convex and,
for this purpose, the pump characteristic curves of each pumping
station must be strictly monotone decreasing. Note that in the
previous statement, even though a particular pump within a
pumping station may not be strictly monotone decreasing, the
content function will be strictly convex, provided that the rest
of the pump curves of this station, make the characteristic curve
of the whole station a strictly monotone decreasing function.
In networks where pumps head "gains" are the dominant factor
(over pipe friction losses for example), Collins et al. (1979)
showed how the information provided by the content function can
be used to draw some conclusions with respect to the stability of
possible multiple operation points. The conclusions formulated
by Collins et al. (1979), are consistent with the well known
example of a single pump, having a characteristic curve with both
an ascendant and descendent limb, when two operating points can
be found where the characteristic curves of the pump and that of
the hydraulic system intersect.
More recently Carpentier et al. (1985), based on the content
and co-content formulation of Collins et al. (1978), and using a
120
variational approach, provided the proof of the uniqueness when
the functions Oii(x) in (34) are twice differentiable, strictly
convex and strictly monotonically increasing. Basically, these
conditions are imposing a restriction on the Hessian matrix, in
the sense that it has to be invertible in order to have a unique

solution. Additionally, Carpentier et al. (1985) showed that
convergence problems in some network analysis algorithms (the
nodal ones) are also related to singularity in the Hessian

matrix.
Thus, summarising, provided that we are dealing with water
distribution systems having dissipative elements, and pumps with
strictly decreasing characteristics curves, the uniqueness of the

solution -is guaranteed. The general model adopted here for
describing the head loss (gain)/flow relationship in the elements
of the network model (equation 8), provides the way for
automatically avoiding ill-posed problems dealing with pumps.
3.8. Other gradient formulations.
We concentrate now on some network analysis methods which have
not been reviewed so far, and which also fit into the class of
gradient-based methods. They are: a second mesh-nodal algorithm
[the first one was that developed by Hamam and Brameller,
(1971)], proposed by Stimson and Brameller (1981), and an
algorithm recently published by Dupuis et al. (1987). Here, we
intend to review these methods and to establish their differences
from the gradient method [Todini (1979), and Pilati and Todini
(1984)1 We have to remember that the improved version of the
121
linear theory method by Wood (1981a), can also be classified as
a gradient method; this method has already been p resented in
Chapter Two (Section 2.4.3).
The new "integrated mesh-nodal" algorithm of Stimson and
Brameller (1981) is an improvement of the previous mesh-nodal
algorithm of Hamam and Brameller (1971) and, though it is based
on some loop (mesh) concepts, the new algorithm has to solve a
linear system of equations in the unknown piezometric heads,
which has the same structure as that obtained in Todini's
gradient method. There are, nevertheless, other differences,
particularly in the way the flows are updated at each iteration,
which make the "integrated mesh-nodal" algorithm different from
the gradient method. We shall derive Stimson and Brameller's
algorithm, using the notation we have already used in this
chapter and in the previous one.
Stimson and Brameller (1981) recognised the fact that the loop
balance condition (second Kirchoff law), is simply establishing
the uniqueness of the piezometric heads at each node.
Nevertheless, they use the concept of tree and co-tree to

p artition the set of flows into subsets of dependent and
independent flows, exactly in the same way as in Hamam and
Brameller (1971), the independent flows corresponding to those in
the co-tree (chords) and the dependent flows corresponding to

those in the tree (branches).
The basic equations, which have to be fulfilled simultaneously
in the integrated mesh-nodal method, are:
Nodal (mass) balance: A21 = a (37)
122
Head loss per pipe: h = A l2 H A lo Ho (38)
and
Head loss/flow relationship: h = All (39)
Having a balanced flow solution at the iteration "i", say

means that:
A21 Q(i) = 2 (40)
If the rest of the basic equations are not fulfilled at
iteration "i", a pipe flow correction 8Q(i) is required, such

that:
Q(i+1) = Q(i) 8Q(i) (41)
and which still fulfils the nodal balance condition, i.e.:
A 21 Q(i+1) = a (42)
Upon introducing (40) and (41) into (42), we get:
A 21 8Q(i) = (43)
This is simply imposing the condition that the flow changes at
every node, with respect to the initially balanced condition,

must be null.
The p artitioning between dependent and independent flows is
introduced as:
(44) .
where:
Q2 : are the NN-NS dependent flows, i.e. tree flows.
Q3 : are the NP-(NN-NS) independent flows, i.e. co-tree

flows.
123
Similarl y , the flow correction vector SQ can be partitioned as:
8Q2 I
SQ = (45)
SQ3
The incidence matrix A21 can be partitioned accordingly:
A21 = [ A22 : A23 ] (46)

where:
A22 incidence (NN-NS)*(NN-NS) matrix corresponding to the
dependent flows.
A23 : incidence (NN-NS)*(NP-(NN-NS)) matrix corresponding to
the independent flows.

The node balance condition (43) can be written as:
6Q 2 I
[ A 22 : A 23 ]*[ --- =0 (47)
SQ3
and the dependent component of the flow correction can be

computed from (47), as a function of the independent component,
provided that the sub-matrix A22 is non-singular:
6Q2 = A22 -I A23 8Q3 (48)

1
The importance of equation (48) lies in the fact that the set
of unknown flow corrections (SQ.), can be reduced to the
independent flow corrections (8Q 3 ), the dependent flow
corrections (6'Q2) being computed from the independent flows,
using just the topological information of the system. In the

context of the integrated mesh-nodal approach, this is a key
124
issue, since (48) ensures that, starting with a balanced flow
distribution, the updated flows will remain balanced, despite
numerical approximations in the intermediate steps.
The only condition left is the fulfilment of the head
loss/flow relationship, which is a non-linear function. In order
to linearize it, the Taylor expansion is used.
If h(Q) = A ll Q then, at the i-th step of the iterative

scheme, the head loss/flow function can be expanded as:
h(Q(i) + 8Q(i)) = h(Q(i)) + j(i) sii(i) + E(i) (49)
where:
E (i) : is the summation of second and higher order terms in the
Taylor's expansion.
Ohl
J (i) is the tangent matrix J( i ) = [-- (50)
aQ
Q=C1( i)
or, in our notation (see equation 17):
J = N Ali* = G (51)
Because
h(i) = A ll Q(i) (52)
equation (49) can be introduced into the expression for the
head loss (38), giving:
A l2 H + A lo Ho = All Q (i) + J( i ) 6Q(j) + E( i ) (53)
where H is the "true" piezometric head vector, because we are
keeping the Taylor's expansion with all its terms, including the
higher order ones (in E(i)).
Equation (53) allows us to compute the flow correction vector
8Q (i) , provided that the tangent matrix is non-singular:
125
SQ (i) = [ J(i) ] -1 f A l2 H 4. A 10 HO - A 11 Q(i) - E(i) 1 (54)
which, when p re-multiplied by A21, allows us to re-introduce
the nodal balance condition (43), giving:
A 21 [J(i)]-1 f A 12 H + A 10 HO - A 11 Q(i) - E(i) 1 (55)

= °
Because, as a result of the Taylor's expansion, the residual

E (i) is not explicitly known, we can assume that its value is
nearly zero when the true heads H are a pp roximated by 11(i+1),

hence:
(3- 4-1) --> Q and ji( i+1 ) --> H if E (i) --> 0 (56)
and equation (55) becomes:
A21 [J(i)]-1 fA l2 H(i+1) + A 10 HO - A ll Q (i) 1 = 0 (57)
Re-ordering equation (27) gives:
A 21 [J( i)]-1A l2 H(i+1) = A 21 [J(i)]-1 (A 11 Q(i)-A 10 HO ) (58)
which is a linear system of (NN-NS) equations in H(i+1).
The flows can be updated from (54) which, when introducing the
assumptions (56), gives:
6Q(i) = [J(i)]-1 f A l2 H(i+1) + A 10 HO - A ll Q(i) / (59)

and, using (41):
(i+1) = Q(i) + [j(i)y4
f A l2 11(i+1) 4- A 10 HO - A ll Q(i) 1 (60)
or, reordering:
it(i+1) = {I-EJ(]- 1A 11
L 1Q(i) - [J(i)]-1{Al211(i+1)+A10H0} (61)
1
Instead of solving (61), Stimson and Brameller (1981) use the

corresponding equation for the independent flows:
126
ii (i+1) = II 3 -N3( i) ]-1 A 3 31 Q 3( i ) _[J 3 (i) ] - 1 {A 32 1:1( i +1) + AnHol (62)
where:
13 : identity matrix of order NP-(NN-NS)
J3 : partitioned {NP-(NN-NS)}*{NP-(NN-NS)} tangent matrix
corresponding to the independent flows.
A33: partitioned {NP-(NN-NS)}*{NP-(NN-NS)} matrix, obtained
from All, for the independent flows.
A30: partitioned {NP-(NN-NS)}*{NS} matrix, obtained from A10,
for the independent flows.
Upon computing Q3 (i4.1) with (62), the dependent flows are
updated using the equivalent of equation (48) for the flows:
Q 2 (i+1) = - A 22 -1 A 23 Q3(i+1) (63)
As it has been said before, Stimson and Brameller (1981) use
(62)+(63) to update the flows, thus avoiding the effects of round
off errors in the solution of the linear system (58), where they
found convergence difficulties in some networks.
In summary, Stimson and Brameller (1981) integrated mesh-nodal
algorithm is as follows:
a) Initialise the iteration counter (i=1) and find a balanced
flow solution: Q (1) . This is done via the construction of a
minimal spanning tree of the network, determining dependent and
independent branches, and assigning arbitrary values to the
independent flows Q3 (1 ). The dependent flows Q2 (1) are computed
with (63). All the matrix partitioning is done at this step,
having the tree and co-tree from the spanning tree.
127
b) Compute the tangent matrix J( i ) and the head loss vector:
A 11 Q(i), using the latest flows.

c) Solve the linear system (58) for H(i+1).
d) Update the independent flows using (62).
e) Update the dependent flows using (63).
f) If the latest flows satisfy the termination criterion stop,
otherwise update the iteration counter and go to step b).
As it can be seen from comparing equations (58) and (26), the
main structure of the linear system for the head update H (1+1) in
the Stimson and Brameller algorithm is fairly similar to that in
the Todini algorithm; indeed apart from the difference in the
sign, which is p roduced by the definition of dH in the gradient
method (equation 23), the only difference lies in the nodal
imbalance term (A2 1 Q( i )- a).

On the other hand, for the flow updating, equations (61) and
(25) are identical, although Stimson and Brameller (1981) do not
use (61) for the flow updating; they use (62) which comes out
directly from (61).
Thus the main difference between the integrated mesh-nodal
algorithm and the gradient algorithm, is basically in the
handling of the nodal balance. We have not found problems with
round off in the gradient method, which allows us to drop the
nodal imbalance in (26) once a balanced flow solution has been
obtained in the first iteration. On dropping the imbalance in
equation (26), both algorithms become equivalent in terms of the
head updating stage. The difference in the flow updating remains.
128
The integrated mesh-nodal approach requires the generation of a
minimal spanning tree, to determine the tree and co-tree of the
network. It is not clear for us how the algorithm performs in
cases when the topology of the network changes, for example if
check valves or pressure regulating valves close completely,
producing a change in the original topology of the network (26).
The gradient algorithm does not require the generation of the
minimal spanning tree (or loops); hence it is not affected by
such topological changes. This is possible because all the
topological information required by the gradient method is
contained in the incidence branch to node matrix Al2.
The integrated mesh-nodal method is currently being used by the
WATNET computer package; see Crabbe et al. (1982) and Stimson
(1982a, 1982b) for more details.
Dupuis et al. (1987) presented what they called a "new
method", which is based on the coupling of the nodal balance
equations and the loop balance conditions (summation of head
losses in a loop = 0). The resulting non-linear system is
linearized via the application of a Taylor's expansion, which
leads to a linear system having the flow increments as unknowns.

Thus, this is basically the same as the gradient version of the
linear theory method of Wood (1981a), already reviewed in Chapter
Two (Section 2.4.3) and which is not referenced by Dupuis et al.

(1987).
129
3 9 Comparison of the gradient method with some of the
the existing methods of network analysis .
Our objective in this section is to compare the performance of
the gradient method with those algorithms performing best in the
comparison carried out by Wood (1981a) [see also Wood and Rayes
(1981)], i.e. with the simultaneous loop adjustment methods
(looped Newton-Raphson) and the gradient version of the linear
theory method.
Since the integrated mesh-nodal model has been considered an
improvement of the original mesh-nodal method of Hamam and
Brameller (1971), we do not consider the latter in the
comparison. The integrated mesh-nodal has not been included,
because we expect similar behaviour to the gradient method, due
to the fact that their structures are similar; this is also due
to the fact that we do not have a source program based on the
integrated mesh-nodal model, which would allow us to determine
the number of iterations and the run-time for comparison
purposes.
We are mainly concerned with the convergence characteristics of
the algorithms, although the comparison is extended to other
parameters as well.
.9.2. The computer p rog rams used in the comparison.
A set of network examples was assembled and run with computer

p rograms based on the simultaneous path adjustment method, the
130
linear theory method and the gradient method . The computer
programs used for the comparison have the following
characteristics:
a) Simultaneous path adjustment method (Program LOOP)
This is a FORTRAN version of the BASIC program included in Wood
(1981a); the code requires both the initial flow distribution and
the definition of the energy paths (loops) to be provided by the
user, as part of the input data. The linear system of equations
is solved via a direct method, consisting of a Crout's LU
decomposition coupled with substitution routines, as presented by
Press et al. (1986). The program does not take advantage of the
sparsity of the matrices.
Although other computer codes based on the same method were
also used at earlier stages of this comparison, this program
(LOOP) was found to have the best p erformance, in the sense that
it was bug-free and reliable.
b) Linear Theory method (Program LT)
This code is a FORTRAN version of a BASIC program included in
Wood (1981a). The algorithm used is that described in Chapter Two
(Section 2.4.3). The program requires the identification of
energy paths by the user, but it normally generates its own
initial flows based on a velocity of 4 fps. The same direct
linear system solver of program LOOP was used in this program.
The program does not use sparse techniques for matrix

manipulation.
131
Because the main objective of this comparison has been to
evaluate the convergence characteristics of the methods, rather
than the computer codes themselves, we do not claim that a fully
fair comparison has been carried out. Indeed, the codes could
have been improved by simplifying the input data required (no
initial flow distributions and no identification of loops and
energy paths to be provided by the user) and by using better
methods for handling sparse matrices, but these improvements
would not have a positive impact on the convergence
characteristics of the methods on which the codes are based, noZy

on the effective storage and run-times.
c) Gradient method (Program GRAD)
This FORTRAN program is based on the algorithm described in
Section 3.6 of this chapter. The code does not need the
identification of loops and the provision of an initial flow
solution by the user; it uses sparse matrix manipulation and a

LDL T factorization for the direct solution of the linear system
of equations, as presented by Pissanetzky (1984). (See Appendix A
for details on the direct solution of linear systems). This
direct linear solver was used in order to have a relatively fair
comparison with the other programs, which also used a direct

solver.
3.9.3. The set of test examples.
The programs were tested with a subset of 6 examples of ill-
conditioned networks, all of them adapted from cases reported in

the literature.
132
i) Example A
This is a network of 66 pipes, 41 nodes and 2 reservoirs (nodes
1 and 2 in Fig. 3.2), taken from a case presented by Chin et
al.(1978) and reduced to its looped part (i.e. without the
branches connected at node 25). This network incorporates small
diameter pipes connected to larger diameter pipes, and its size
makes it similar to real networks found in practical
applications. The nodal demands have been reduced (producing
velocities of under 0.2 m/s) to simulate the behaviour of the
network during the night. This example network includes 2
gravitational sources and it does not include pumps.
ii) Example B
This is almost the same network as example A, except that 6
pumps with steep characteristic curves have been included; also,
the branched part connected at node 25 (see Fig. 3.2) has been
added which gives a total of 74 pipes, 48 nodes and 8 sources.
The steep characteristic curves have been taken from data of real
submersible multistage pumps with speeds in the range of

2900/3500 r.p.m..
iii) Example C
This is the same basic network as presented in example B, but
now the nodal demands have been multiplied by a factor of 10,
simulating extremely high daily consumptions.
133
Fi g . 3.2. Basic network for examples A, B, C and D, from Chin et
al. (1978).
134
iv) Example D
The previous example C has been modified to include 2 valves
(in links 67 and 70) which are completely closed in order to
simulate the eventuality that a particular link ceases to carry
flow (e.g. a check valve working under reverse flow or a pressure
reducing valve being closed by the system downstream pressure).
The valves are located in such a way that, when fully closed, the
original network becomes a disconnected system, with two
independent sub-networks, each one with its own sources.
v) Example E
This is a network of 16 pipes, 14 nodes and 2 sources, adapted
from a case presented by de Neufville and Hester (1969), as shown
in Figure 3.3. This network becomes ill-conditioned when one of
its pipe diameters is reduced (pipe 10), while the remaining
pipe diameters are unchanged.
vi) Example F
This is a network of 13 pipes, 10 nodes and 1 source, reported
by Barlow and Markland (1969), as shown in Figure 3.4. This
network becomes ill-conditioned for the particular diameter
distribution used.
12
•
5 (1/s)<=0
13
V
5 (us)
P
A16 15
.6
10 (Vs)
5 (us) 5(VO
14
V
5 (Vs)
1;t1)
Fi g . 3.3. Network for example E, from de Neufville and Hester

(1969).
100m
Fi g . 3.4. Network for example F, from Barlow and Markland (1969).
136
3.9.4. Comparison of the algorithms for the simultaneous path
adjustment, linear theory and_gradient methods.
In order to produce a comparison of the algorithms, independent
of the initial conditions, the same initial flow distribution was

supplied to all the programs. In standard applications, both the
linear theory and the gradient methods, do not need an initial
flow solution as part of the input data. For the sake of a fair
comparison, a direct solution was used for solving the linear

system of equations in all the cases.
The results of running the test networks with the 3 computer

programs are summarised in Table 3.1.
From the results obtained by Wood (1981a) the following are
the most usual causes of convergence problems in water
distribution networks:
- low resistance links connected to high resistance ones;
- the presence of pumps with steep characteristic curves.
Based on the results summarised in Table 3.1, the conclusions

of Wood (1981a) are partially corroborated.
From the pipe flows point of view, the simultaneous path
adjustment method (program LOOP) and the linear theory method
(program LT) converged for all the examples, even under cases
previously reported as ill-conditioned ( examples E and F). Both
programs failed to determine the piezometric heads when the
network included very small diameter pipes, or nearly closed
valves (examples D and E), where the high head loss (eventually
137
Table 3.1. Summary of the performance of the programs LOOP,
LT and GRAD with test examples.
PROGRAM
EXAMPLE
LOOP LT GRAD
,
converged converged converged
A iter=16 iter=17 iter=16
t=69.53 sec. [1] t=788.51 sec. t=29.77 sec.
converged slow convergence converged

B iter=12 iter=13 iter=10
t=91.67 sec. t=961.69 sec. t=34.00 sec.
converged slow convergence converged
C iter=13 iter=15 iter=12
t=100.13 sec. t=1109.82 sec. t=39.16 sec.
converged [2] slow convergence converged
D iter=21 iter=21 [2] iter=19
t=154.94 sec. t=1551.76 sec. t=57.28 sec.
some heads unavail. some heads unavail.
converged [3] converged -converged

E iter=12 iter=13 iter=12
t=4.67 sec. t=22.36 sec. t=5.38 sec.
some approx. heads
converged converged converged

F iter=4 iter=8 iter=4
t=2.69 sec. t=9.12 sec. t=2.47 sec.
Notes:
[1] All the times correspond to mean execution time on an

IBM-PC compatible computer (Amstrad PC1512, with
co-processor 8087).
[2] Program failed to compute piezometric heads in one

sector of the network.
[3] Computed heads are not exact.
[4] Only for fairness in the comparison, all the programs

are starting from the same initial flow solution. For
the same reason, a direct solution was used in all
the programs.
138
of "infinite" value) produced unreliable piezometric heads in the
section of the network which does not remain connected to the
reference node. This is in fact a problem inherent to all loop or
path-based methods, and is produced because the network is
broken-up into two (or more) disconnected sub-networks and
consequently the path to the reference node is lost. A solution
to this problem can be produced, by re-defining the reference
node, but it is not clear if this makes the program problem-
dependent, because, in more complex networks, the disconnection
of the system may happen anywhere (i.e. due to pressure
regulating valves, maintenance work, etc.).
On the other hand, the linear theory method, in its improved

form, showed slow convergence when applied to networks with steep
characteristic curve pumps (examples B, C and D).
The linear theory method produces a matrix of coefficients

for the linear system (see Chapter Two, equation 150), which is a
non-symmetric matrix. This is another shortcoming of the linear
theory method, since it makes the storage requirements critical
for real networks, even when efficient sparse matrix techniques
are used; in addition, pivoting is necessary in order to
guarantee the stability of the linear solver, which implies an
extra computational cost. The simultaneous path adjustment
method, and the gradient method, produce symmetric matrices of
coefficients in the linear systems of equations; as a result,
their storage demands are about 34% (simultaneous path) and 41%
(global gradient) of that needed for the linear theory method.
139
These are approximate values computed for a regular square
network.
The global gradient method did not fail to conver g e for any
of the test examples; moreover, since its development, we have
not been able to find a single example of failure or
unsatisfactory convergence rate. This method did not experience
problems when closed valves or high resistance links where
introduced into the networks. This is due to the fact that, at
each stage, the algorithm computes both heads and flows;
subsequently, it will converge properly provided that all the
sub-networks do have a source, which is a sensible condition
from the physical viewpoint. This is an important advantage of
the method, since the situation described will probably occur
when using this method for simulation purposes, in more complex
s ystems having check valves and pressure regulating valves.
Future implementations will include the detection of cases such
as those where a sub-network does not have a source, displaying a
clear warning to the user.
The proposed method requires an efficient and robust linear

s ystem solver, and we study this problem in Chapter Five.
We also ran other codes for the "simultaneous path
adjustment" algorithm, using different solvers for the linear
system of equations, and some of them failed to converge for some
of the test examples. This suggests that some reported ill-
conditioned cases are solvable, when a different linear system
solver is used. This seems to be a numerical problem and not a

p roblem of the water distribution analysis methods themselves.
140
3 10. Concluding remarks.
When compared with the simultaneous path adjustment method and

the linear theory method, the proposed gradient method has the
following advantages.
1. It can solve directly partly looped and partly branched

networks, while, in the simultaneous path method, the problem
needs to be transformed into an equivalent looped network, prior
to the application of the iterative algorithm. This has to be
done by the user (manually) or incorporated as an additional
subprogram.
2. The proposed gradient method does not need a loop or path

definition, as in the cases of the simultaneous path adjustment
and the linear theory methods. Even though this task can also be
done by the computer, it implies an additional computational cost
(in the examples tested in this paper, the paths were provided as
input data).
3. The proposed gradient method can solve in a straightforward

way networks that during certain periods of operation can become
disconnected (due to the action of check valves or pressure
reducing valves, for example). The simultaneous path adjustment
and the linear theory methods cannot cope with this situation,
although it might be possible to implement an additional
subroutine that could solve this shortcoming, although it is
doubtful if an absolutely reliable algorithm can be obtained,
especially when dealing with larger and more complex systems.
Either way, this means an additional computational cost.
141
In summary, it has been shown that the gradient method for
determining the steady state flow in water supply distribution
networks, offers some advantages when compared with other
gradient-based algorithms, such as the "simultaneous path
adjustment" method and the linear theory method. The reliability
of the method, in terms of its ability to converge for both the
flows and the piezometric heads, under extreme cases based on
ill-conditioned problems, makes it desirable for optimum design
and simulation applications, where manual intervention to avoid
convergence or disconnection problems is not possible.
142
CHAPTER FOUR
EXTENSION OF THE GRADIENT METHOD TO INCLUDE REGULATING VALVES:
A NEW PHYSICALLY-BASED APPROACH
4.1. Introduction.
So far, we have studied water distribution networks having
pipes, pumps and valves. The kind of valves we have been dealing
with are basically gate and butterfly valves; in other words,
valves with a fixed and known head loss-flow characteristic
curve. These valves may be operating fully open, partly open or
fully closed within the network, and the state of each valve is
assumed to be known prior to the analysis of the system; in fact,
the operator can alter the setting of any of these valves (by
giving instructions to a field crew or via a motor-operated
system), but basically the valves characteristics are known
and/or are determined by the operator.
From the hydraulic point of view, gate and butterfly valves
represent a local (minor) head loss whose magnitude depends on
the valve type, its design and its operational status (fully
open, partly open or fully closed). A short review of different
types of valves and their characteristics may be found in Ruus
(1981).
In the context of the gradient algorithm, we have introduced
a general model in Chapter Three (Section 3.5) for the head
loss/flow characteristic curve of all devices via a polynomial

expression:
143
= a ij Q ij n + (1)
where:
hi j : head loss (m) in the branch starting at node "i"

and ending at node "j".
: real exponent, typically 1.85-2.0, depending on the
formula used.
aij, Bij : characteristic parameters. Because (1) is our
general model, for pipes and valves we drop the
constant term (i.e. a ij =o), being the head loss/flow
characteristic dependent on aij only. This value is
either supplied by each valve manufacturer, or is
determined via laboratory or field head loss/flow
measurements, for different operating conditions.
The model (1) is then able to describe any type of device with
a fixed head loss/flow characteristic curve (valves, meters,
fittings, etc.).
Due to the variable behaviour of water consumption, the network

reacts by varying both pressures and flows, which the operator
must maintain within pre-specified limits to achieve an
acceptable service. Minimum and maximum pressures are required at
any point in the system. Also, constant flows may be required to
satisfy fixed demands. On other occasions, we may be interested
in avoiding flows in certain directions. To avoid overflows when
feeding a reservoir, it may be desirable to include in the
network an automatic level control valve (altitude valve). Some
of these requirements may even change during different periods of
a day (e.g. from day to night pressures can be reduced in order
144
to reduce leakages).
To cope with all these requirements (and others) a wide range

of pressure, flow and level control valves are available,
allowing the operator to fulfil the operational constraints in an
automatic way, without his permanent supervision. Since the
variety of control valves is quite wide, we shall refer to the
most widely used ones, namely check valves, pressure reducing
valves and pressure sustaining valves. A short explanation of
their physical characteristics is needed in order to understand
their behaviour, and to determine an adequate mathematical
representation.
4.2. Description of some control valves.
Pressure reducing valves (PRV's) and pressure sustaining valves

(PSV's) belong to the same type of pressure controlling devices,
but check valves (CV's) are completely different, because their
design and operational behaviour is different.
4.2.1. Check valves.
The main purpose of a check valve is to allow flow in one

direction only, preventing flow reversal. This situation may
arise when feeding a high pressure zone of the network from a
lower pressure zone, via pumps for example; in such cases, we
would be interested in avoiding backf low from the higher to the
lower pressure zone, when the pumping system is switched off.
The traditional hydro-mechanical design of a check valve has

been based on a flap, which is forced open by the flow itself (in
145
one direction), but that automatically closes when the flow
approaches zero, or when it is going to be in the opposite
direction. Due to the undesirable side effects of this type of
valve, especially from the pressure surge viewpoint, a number of
modifications of the basic valve have been implemented, in order
to reduce the waterhammer produced by a very fast closing of the
valve. Also, some hydraulically controlled designs are
available, in order to slow down the response of the valve, but
we shall not describe them here.
As a result, a check valve can operate under two mutually
exclusive modes:
a) Fully open: when the inlet pressure is greater than the
outlet pressure.
b) Fully closed: when the outlet pressure becomes greater than
the inlet pressure.
Hence, the valve has an associated direction, and its
installation must be made accordingly. The valve operating mode
is controlled by its outlet and inlet pressures, these being
independent of the operator's commands.
Mathematically, the valve has to be described with two
different head loss-flow relationships
i) Fully open (inactive mode):
= aij Qij n when Hi > Hj (Qij > 0) (2)
where:
= constant, dependent on the valve design.
Qii = flow from node "i" to "j" (1/s)
Hi = piezometric head at node "i" (m)
146
ii) Fully closed (active mode):
hi j = Hj - Hi (Qij = 0) when Hj > Hi (3)
4.2.2. Pressure controlling devices: main feature .
It is possible to identify three main components in a pressure
controlling valve: main valve, relay system and connecting piping
plus fittings. The main valve and the relay system are basically
the same for most of the pressure control valves; they differ
only in the connecting piping and fittings, which are added to
the main valve to achieve its particular objectives. A more
detailed description of the construction and operational
characteristics of pressure reducing valves can be found in
Ratcliffe (1986).
For convenience, we shall refer to Glenfield & Kennedy control
valves, a leading British valve manufacturer, but most of their
features are applicable to other manufacturers as well.
a) Main valve.
Although there are different design concepts for the valve body
we shall refer to control valves based on globe valves.
The main valve body is a cast iron globe valve design, modified
to include an upper cylindrical section, which contains a mobile
element that fits into the main valve seat. Either Figures
4.1.a. or 4.2.a. can be used to follow the description of the
different valve components. From the hydromechanical viewpoint,
the central mobile "valve element" (component 1 in Fig. 4.1.a.)
introduces a variable head loss throughout the valve, depending
147
1 Valve Element
2 Upper Cylinder
3 Isolating Cocks
4 Strainer
5 Orifice
6 Needle Cock
7 Indicator
8 Relay Valve
9 Diaphragm
10 Spring
11 Adjusting Screw
12 Pressure Gauge
13 Alternative Constant
Pressure Connection
14 Electric Motor
a) Valve Nos. 1301 and 1302

Standard and Motor Operated Valves
lithe outlet pressure P2 becomes too high to

balance the spring load, the relay valve
tends to close, allowing pressure P3 to
increase, causing the main valve to tend to
close, thus reducing P2 to the set value.
Similarly, if P2 becomes too low to balance

the spring load, the relay valve opens, P3
decreases, the main valve opens further and
P2 rises again to the set value..Thus the valve
continuously maintains the desired constant
downstream pressure P2, which is determined
only by the spring load and is entirely
b) Operation independent of the upstream pressure.
STATIC PRESSURE UPSTREAM STATIC PRESSURE WITH VALVE SHUT
VALVE SHUT IF DOWNSTREAM PRESSURE

EXCEEDS SET OUTLET PRESSURE
RANGE OF GRADIENTS WITH INLET PRESSURE RANGE

CORRECT OUT LET PRESSURE
CONSTANT OUTLET PRESSURE
VALVE FULL OPEN AND DOWNSTREAM STATIC PRESSURE WITH VALVE SHUT
IN ACTIVE WHEN INLET w
R
SET OUTLET
PIRLEVIEJ IR?
RANGE OF GRADIENTS WITH CORRECT
UT LET PRESSURE
SET OUTLET
PRESSURE RANGE OF GRADIEN'T---
RANGE OF GRADIENTS WITH
VALVE FULL OPEN UNDER EXCESS DRAW OFF
rossistE WITH E aCE SS DRAW-OFF -

UPSTREAM OR DOWNSTREAM OF VALVE
c) Hydraulic Diagram for Pressure Reducing Valves
Fig . 4.1. Pressure Reducing Valve. Design and operational

features, from Gle
215/R3. nfield & Kennedy Ltd., Pub.
148
1 Valve Element
2 Up p er Cylinder
3 Isolating Cocks
4 Strainer
5 Orifice
6 Needle Cock
7 Indicator
8 Relay Valve
9 Diaphragm
10 Spring
11 Adjusting Screw
12 Pressur• Gauge
a) Valve No. 1310 Standard Valve with manually operated relay valves
When the inlet.pressure P1 becomes too high

to balance the spring load, the relay valve
tends to open, allowing pressure P3 to
dirninish so that the main valve tends to open,
thus reducing P1 to the pre-set value.
When used in the Sustaining application, if

Pi becomes too low to balance the spring load,
the relay valve closes, P3 increases, the main
valve closes and P1 rises again to the pre-set
vahisi
b) Operation
STATIC PRESSURE-VALVE FULL OPEN
WHEN VALVE IS FULL OPEN

OUT LET PRESSURE MAY RISE
ABOVE SET INLET PRESSURE
RANGE OF GRADIENTS
WITH VALVE FULL OPEN
•
LOSS
THAT.
POSSIBLE GRADIENTS VALVE
WITH coRRECT
INLET PRESSURE TERMINAL UPSTREAM PRESSURE
C)
VALVE SHUT WHEN TERMINAL
UPSTREAM PRESSURE IS LESS
THAN SET INLET PRESSURE
1
SET INLET
PRESSURE
INLET
RANGE OF GRADIENTS
Valve No, 1310 Pressure Sustaining Valve Hydraulic Diagram

WITH CORRECT INLET PRESSURE
EXCESSIVE DRAW OFF
Fig. 4.2, Pressure Sustaining Valve. Design and operational

features, from Glenfield & Kenned y Ltd., Pub.
215/R3.
149
on the relative values of the pressures at the inlet, outlet and
upper cylinder (in Fig. 4.1.a.: P1, P2 and P3, respectively).
This component (the valve element) allows the valve to move from
a fully open position, with a minimum head loss (component 1 up,
see Fig. 4.1.a.), to a fully closed position (component 1 down)

with a maximum head loss.
b) Relay system.
The relay system allows the operator to set a pre-specified
pressure to be maintained, either at the inlet or outlet of the
valve (depending if we are dealing with a PSV or a PRV). This
relay system consists of a small relay valve (component 8 in
Fig. 4.1.a.), which is operated with a spring (component 10 of
Fig. 4.1.a.), whose tension is set via an adjusting screw
(component 11 in Fig. 4.1.a. ). The screw can be operated by
hand ( local control mode) or via an electric motor (remote
control mode). The force produced by the spring tension acts
against the pressure being controlled (either downstream or
upstream of the valve), opening or closing the relay valve when
the controlled pressure differs from the pressure produced by the
spring. A reinforced synthetic rubber diaphragm (component 9 in
Fig. 4.1. a.) allows the controlled pressure to be applied

directly against the compressed spring.
c) Connecting piping and fittings.
The inlet, outlet and upper cylinder of the valve, and the
relay system, are connected by a series of small pi pes and

•
150
valves. A change in their connectivity implies a change in the
objective of the control valve, transforming the basic globe
valve body into a PRV, a PSV, altitude valve, constant flow valve
or a combination of some of them. This flexibility provided by
the connecting piping system, added to the fact that the pressure
setting s can be adjusted on-site or by remote control, makes this
type of device a very helpful piece of hardware in any

efficiently operated water distribution network.
a) Pressure Reducing Valves (PRV's).
A PRY is a pressure controlling device designed to perform the

following duties:
- Maintain a constant downstream pressure when the upstream
pressure exceeds a pre-established pressure setting.

- Avoid reverse flow when the downstream pressure becomes
higher than the pressure setting and/or when the inlet
pressure becomes less than the outlet pressure .
The first objective is met via the variable head loss
introduced by the "valve element", as explained before. Figure
4.1.b. explains how the PRV reacts to outlet pressure variations,
different from the valve setting, adjusting the position of the
valve element (component 1 in Fig. 4.1.a.), until the head losses
within the valve produce the desired outlet pressure. In this

operational mode, the PRY is said to be in the "active mode".
The second objective is met by the PRV being fully closed,
acting in fact as a check valve and preventing backf low or
backpressure. This is the "check valve mode" of operation of the
151
PRV. This operational mode is achieved in two situations:
i) The outlet pressure (P2, see Fig. 4.1.a.) is greater than
the pressure setting; then, the difference between spring tension
and outlet pressure P2 is acting upwards, closing the relay valve
(component 8 in Fig. 4.1.a) and increasing the pressure in the
upper cylinder (P3 in Fig. 4.1.a.). As a result of a greater P3,

the valve element moves downwards, closing the valve.
ii) The outlet pressure (P2 in Fig. 4.1.a.) is lower than the
setting pressure, but greater than the inlet pressure (P1 in Fig.
4.1.a.). The relay valve keeps open and P1 < P3 < P2, then the
pressure in the upper cylinder is greater than the inlet one (P1)
and the PRV closes.
A third situation arises when the inlet pressure is less than
the pressure setting, but greater than the outlet pressure. In
this case the PRV operates fully open in its "inactive mode",
introducing a minimum head loss into the system. In this
particular situation, because P2 is lower than the setting
pressure, the relay valve remains open and, then, P1 > P3 > P2
which means that the valve element tends to stay fully open.
Figure 4.1.c. explains graphically all the possible operating
modes of a PEST and shows the corresponding pressure profiles.
As a result, the head losses through the PRV are essentially
variable. When the PRV is fully open, this introduces a minimum
local head loss, while, when in the active mode and check valve
mode, the PEST can be understood as a variable head loss device.

This feature is shown in Figure 4.3.
152
Another way of understanding the PRV's behaviour, is by
plotting both the inlet (Hi) and outlet (Hi) pressures
simultaneously, as shown in Figure 4.4, for a constant positive
flow through the valve.
In the operating point "1" of Fig. 4.4 (Hi < H setting) the
valve is fully open, and the difference between the inlet and
outlet pressures depends only on the flow (or velocity) and the
minimum head loss characteristic coefficient (a 0 ) of the valve.
This coefficient is generally of the order of 5 to 10 velocity
heads.
In the operating point "2" of Fig. 4.4 (Hi = Hsetting 4 h1),
ill is equal to the local head loss produced through the valve in
the inactive mode (wide open). At this point, as Hi increases,
the valve starts closing, increasing its internal head loss in
order to maintain a pre-specified setting.
In the operating point "3" (H i > Hsetting + h 1 ), Hi is

maintained at H setting via extra head losses produced by closing
the main valve.
Figure 4.4 is valid for positive constant flows only, that is
to say, flows from the specified initial to final node.
A simple graphical representation of a PRV, involving the main

variables H i , Hi, Qii, H setting and aii has not been found in the
literature. Figures 4.3 and 4.4 involve only a partial
representation of the PRV's behaviour. An attempt to represent
most of the variables involved is provided by Figure 4.5, in a

tri-dimensional plot.
153
P. R. V. in the ACTIVE or
h ii = H• —H
I i CHECK VALVE operating mode
P. R. V. INACTIVE (fully open)

h ii =c(cprili
Qij
Fi g . 4.3. Feasible head loss/flow region for a pressure

reducing valve. (For a variable flow).
Hj
outlet head
A
H setting
Hi
inlet head
H setting H setting + hl
..
1( >
Valve fully open Valve partly closed
Fi g . 4.4. Relationship between inlet and outlet heads in a
pressure reducing valve. (For a constant flow).
154
Figure 4.5 is valid for positive constant flows only. In the
horizontal plane, Hi versus Hj, we have the previous Figure 4.4.
When Hi < Hsetting + hl, the PRV is in the active mode (fully
open), and its head loss-flow relationship is described by:
= Hi - Hj = ao Qijn
where
ao a constant corresponding to the minimum head loss

coefficient of the valve.
the flow which, for Fig. 4.5 has been held

constant, then hij = constant for Hi < Hsetting + hl.
For Hi > Hsetting + hl, the valve increases its aij , i.e. we
are in the active mode, and the relationship between hij and aij
is linear, with a proportionality factor (slope) equal'to Qij n .
Figure 4.5 has been obtained assuming that a variable head
reservoir (head variation is independent of the flow), is

connected directly to the upstream node of the PRV. The flow is
assumed to be constant and the setting of the PRV is also
constant. This is an ideal situation, since when a PRV is
inserted in a more complex real network, the upstream head Hi and
the flow Qij are not independent, and the flow is not constant,
because a change in the internal head loss of the PRV causes a
redistribution of flows in the network.
Figure 4.6 attempts to represent the more general situation,
for a variable flow, showing the relationship between Hi and Hj.
The coefficient ocij can be incorporated as a third variable, as
in the case of Figure 4.5, for showing the non linear effects,
155
but its practical usefulness is negligible. Figure 4.6 has been
obtained from a system where the reservoir is not connected
directly to the valve inlet, but through a network; then, the
variable inlet pressure is due to the head losses throughout the
network. Starting from zero flow, we move from the point "3" (in
Fig. 4.6) to the left.
The main difference between Figures 4.6 and 4.4 is that the
ascending limb of the curve is no longer a straight line, since
now it has to include the effect of the variation of flow.
The practical application of the previous figures is rather

limited, but they help to understand the behaviour of the PRV's
and also to point out their complexity.
For practical purposes, a lumped representation of the response

of the network to changes in the internal resistance of the PRV's
can be obtained by plotting the downstream head of the PRV versus
the characteristic coefficient aij ; the result is shown in Fig.
4.7. Two possible situations are clearly differentiated, namely
curve "a", where the downstream head is dependent upon the
upstream conditions (i,e. dependent on the internal head loss in
the PRV), whereas in curve "b", the system characteristic does
not allow the downstream head to be de pendent on the upstream
conditions. The latter is typically the case when the downstream
conditions produce heads greater than the PRV setting, and the
valve tends to close; in other words, the PRV is controlled by
the downstream conditions rather than the upstream ones; in this
case the PRV cannot produce the desired Hsetting at its outlet
and simply closes.
156
•
Hsettin9.,
Hi
Fi g . 4.5. Relationship between inlet and outlet head, and aij

in a pressure reducing valve. (For constant flow).
Hi
outlet head
•
hii= — H setting
el 3
H setting as
2
.1= °CoOrilj
Q>0
Hi
inlet head
•
H setting H setting + hi Hi max.
Valve fully open Valve partly closed Valve

fully closed
Fi g . 4.6. Relationship between inlet and outlet heads in

a pressure reducing valve. (For a variable flow).
157
In the case when the outlet pressure is dependent on the head
loss through the valve, Fig. 4.7 clearly suggests how can we find
the value of aii* which produces an outlet pressure equal to
Hsetting.
Essentially, to include PRV's in our model, we shall be using
the feature represented in Fig. 4.7 to balance the network
iteratively. This procedure shall be explained in more detail in
subsequent sections.
b) Pressure Sustaining Valves (PSV's).
A PSV is a pressure controlling device designed to perform the
following duties:
- Maintain a constant minimum upstream pressure regardless of
the downstream pressure.
- Avoid reverse flows when the upstream pressure becomes
lower than the pressure setting and/or when the inlet
pressure becomes less than the outlet pressure .
158
H-
J
downstream head
4
1
/ outlet pressure
independent of
valve resistance
H--+
H setting n ..
a . outlet pressure
is dependent on
valve resistance
Fig. 4.7. Response of the network at the downstream end of a

PRY, to variations in the resistance characteristic
coefficient au.
The first objective is met via the variable head loss
introduced by the "valve element", as explained before in the
case of PRV's. Figure 4.2.b. explains how the PSV reacts to
outlet pressure variations, different from the valve setting,
adjusting the position of the valve element (component 1 in Fig.
4.2.a.), until the head losses within the valve produce the
desired inlet pressure. In this operational mode, the PSV is

said to he in the "active mode".
The second objective is met by the PSV being fully closed,
acting in fact as a check valve and preventing backf low .This is
the "check valve mode" of operation of the PSV. This operational

mode is achieved in two situations:
i) The inlet pressure (P1, see Fig. 4.2.a.) is lower than the
pressure setting; then, the difference between spring tension and
inlet pressure P1 is acting downwards, closing the relay valve
(component 8 in Fig. 4.2.a) and increasing the pressure in the
upper cylinder (P3 in Fig. 4.2.a.) and, as a result, closing the

valve element downwards.
ii) The outlet pressure (P2 in Fig. 4.2.a.) is greater than the
inlet pressure (P1 in Fig. 4.2.a.) and lower than the inlet
pressure setting. The relay valve keeps open and P1 < P3 < P2,
then the pressure in the upper cylinder is greater than the inlet
one (P1) and the PSV closes.
A third situation arises when the inlet pressure is greater
than the pressure setting, and greater than the outlet pressure.
In this case the PSV operates fully open in its "inactive mode",
introducing a minimum head loss into the system. In this
particular situation, because P1 is greater than the setting
pressure the relay valve remains open and, then, P1 > P3 > P2
which means that the valve element tends to stay fully open.
Figure 4.2.c. explains graphically all the possible operating
modes of a PSV and shows the corresponding pressure profiles.
As a result, similar to the case of the PRV's, the head losses
through the PSV are essentially variable and equally complex to
160
represent graphically. Figure 4.8 is the equivalent of Figure 4.4
for the case of a PSV. We shall O m it the rest of the figures,
corresponding with those displayed in the case of a PRV, and we
shall show only the corresponding lumped schematic representing
the response of the system at the inlet of the PSV versus the
head loss characteristic parameter aij, which is shown in Fig.
4.9.
4.3. Existing models for Pressure controlling devices.
Although it is known that many of the network analysis codes
used in consultancy work are able to deal with pressure
controlling devices, the open literature provides only a few
references on the modelling of such devices.
In the early 1970's, Zarghamee (1971), following a nodal
approach, explicitly considered the presence of PRV's. He
actually modelled the set PRY plus the pipe in which the device
was inserted and, for the case when the PRV is next to the
upstream node "i", the model is:
= R ij (H setting H j ) 1/n when Hi k H settin g 1 Hi (4)
where:
is the conductance of the pipe branch between the outlet
of the PRV and the downstream node "j".

H.
outlet head
H setting
H
inlet head
Fig. 4.8. Relationship between inlet and outlet heads in a

p ressure sustaining valve. (For constant flow).
upstream head
inlet pressure
b
independent of
valve resistance
inlet pressure
is dependent on
valve resistance
H sew"
(Kips
Fig. 4.9. Response of the network at the u p stream end of a

PSV, to variations in the resistance characteristic
coefficient aij.
162
When Hj < Hi < H setting , then H setting must be re p laced by Hi
in equation (1), and the same model can be used to represent the
PRV's behaviour. It is implicit that the valve closes when Hi<Hj
or Hj > Hsetting (i.e. Qij = 0.0).
Zarghamee (1971) reported a successful application of his

method to a water distribution system in Teheran, containing 9
PRV's and located in a steep area of the city. Although some
details are not completely clear from the paper, Zarghamee did
get convergence in two study cases (convergence is reported to be
slower when pumps are operating). No problems were reported
due to the presence of the PRV's.
Donachie (1974), within a nodal formulation, proposed a two

stage process for modelling the PRV's. In the first stage of the
iterative procedure, when the difference between successive heads
is greater than 1 m., the outflow is computed with an upstream
pressure pi*, such that:
Pi * = m i n (Pi, Ps) (5)
where:
p i : is the actual upstream pressure.
ps : is the downstream pressure setting of the PRV.
In the second stage (closer to the final solution), the PRV is

considered as a fixed outlet pressure device, i.e.: with a
163
constant outlet pressure equal to ps, p rovided that pi > ps.
Jeppson (1976) and Jeppson and Davis (1976), reviewed all the
modifications needing to be introduced in the different
formulations of the network analysis equations in order to
include PRV's. Basically, the PRV's are modelled by splitting
the branch where the PRV is, inserting a pseudo-reservoir whose
head is equal to the PRV setting and imposing flow continuity
over the two halves of the resulting pieces of pipe, as shown in
Fig. 4.10.
As a result, the introduction of the pseudo-reservoir in the
network analysis equations affects the different formulations, in
a way which is summarised in Table 4.1, using the notation of
Fig. 4.10 and considering the usual head loss/flow model:
hij = aij Qij n + Bij (with Bij = 0 for pipes)
Hsetting
a) Real Pressure Reducing Valve. b) Pressure Reducing Valve modelled as pseudo—reservoir.
Fig. 4.10. Modelling PRV's as pseudo-reservoirs.
164
Table 4.1. Summary of effects of PRV's in different
formulations of network analysis equations
[following prop osition of Jeppson (1976) and Jeppson
and Davis (1976)).
Original set of Modified set of

equations (no PRV's) equations (with PRV's)
Q-e quations:(NN-NS+NL eq.) a) Nodal flow continuity:
a) Nodal flow continuity: same as in the original set
E Qik = qi i=1,..,NN-NS b) Loop head continuity:
k E aij Qij n = 0 1=1,.,(NL-NPRV)
b) Loop head continuity: 1 (same as before)
E cqj Qij n = 0 1=1,.,NL E a ij Q ij n = Hsetting-Hreservoir
1 1" 1=1,...,NPRV
(changed, see Note 1)
H-equations:(NN-NS eq.) Nodal flow continuity:
Nodal flow continuity: (HOHs etting>HP as before,except:
a) Node "i":
va ik = E ((H i _H k )/a ik )1/n 4.
kEQik=E((Hi-Hk)/ai'))1/n=cli
k 1=1,...,N-NS . k k 4. ((H i _H x )/a ix )1/n = qi
i=1,...,NN-NS (see Note 2+Fig. 4.10)
b) Node "j": .
EQ ik = E (([1.1-Hk)/aik) 1/n 4.
k
k=i 4-(THsetting-H.Way.01/n=qi
j=1,...,NN-NS (see Note 3+Fig. 4.10)
Additional flow continuity equation:
Qix = Qyj = Qij i.e.:(see equ. 10)
(H i -H x )/a ix = (Hsetting-HP/ayj
SQ-equations:(NL eq.) Loop head continuity:
Loop head continuity: E u ij (Q ij +8Q 1 ±8Q N) n = 0 (same as
E ai4(Qij+8Q1±eQN) n =0 1 1=1,...,(NL-NPRV) before)
1 ' 1=1,...,NL
(see Note 4)
1„ a ij (Q ij +8Q 1 ±8QN )n=H setting -H reserv
E
1"=1,...,NPRV (changed)
Notes:(0) NN = total number of nodes.
NS = number of source-nodes.
NL = number of loops (or energy paths).
NPRV= number of PRV's.
(1) 1" considers a pseudo-loop replacing the original loop
broken-down by the introduction of the PRV and linking
the pseudo-reservoir with a real reservoir .
(2) aik: a corresponding to the pipe from node "i" to "k".
aix: a corresponding to the pipe from node "i" to "x".
Hx : head at the inlet of the PRV.
(3) axj. : a corresponding to the pipe from node "y" to "j".
(4) 610/N: flow correction in neighbouring loops.
165
As it is clear from Table 4.1 that every PRV implies that the
original pipe (from nodes "i"- "j") is split up into two parts,
we need to restore one loop equation for every PRV introduced.
For this purpose, it is necessary to link every PRV with some
existing (real) reservoir, create a path between them, and write
down the equation establishing that the head loss in this path
must be equal to the difference between the pseudo-reservoir head
(Hsetting corresponding to the PRV) and the head of the real
reservoir.
As far as the H-equations are concerned, the model proposed by

Jeppson and Davis (1976) introduces an additional unknown (the
head at the inlet of the PRV: Hx) for each PRV, but also
introduces an additional flow continuity equation
Q ix = Qyj = Q ij (6)
because
Q ix = ((Hi-Hx)/aix)1/n (7)
Qyi = ((Hsetting41p/ayj)1/n (8)
then, from equation (6):
[(Hi-Hx)/a ix] lin = rtu
L, -setting-H Wayil l /n(9)
or
(Hi - Hx)/ a ix = (H setting - (10)
H)/ c(Yj
Jeppson (1976), and Jeppson and Davis (1976) gave a number of

examples on how to introduce PRV's into the networks using these
procedures. We have tried to reproduce them, but we have not
been able to attain the same results. This disagreement will be
discussed in a subsequent section.
166
Lekane (1979) followed a similar approach to that proposed by
Zarghamee (1971) in modelling PRV's within a nodal formulation.
He also reported the results of the application of the method to
a small network ( 18 nodes and 24 branches with 2 PRV's).
Another application to a bigger network (314 nodes and 424
branches) is also mentioned.
Lekane (1979) raised the question of the uniqueness of the

solution of the steady state, especially when dealing with
devices like PRV's, CV's and pumps. He pointed out that,
according to experience, a unique solution can be expected.
Lekane also mentioned the fact that the presence of PRV's turns
the symmetric Jacobian matrix into a non symmetrical one. This
is relevant from the computational viewpoint.
Collins (1980), in discussing several common possible causes of

difficulties in network analysis, drew attention to the
uniqueness problem, especially when dealing with a network with
control devices. He concluded, based on a simple example of a
network with 7 nodes (3 of them reservoirs) and 6 branches (3
pumps, 2 PRV's and one gate valve), that no definitive answer has
been obtained for this problem. We shall discuss this
interesting case in more detail later on. Collins also questioned
the validity of the usual procedure for dealing with multiple
pressure controlling devices within an iterative solution, where
an initial status is assumed for these devices, and a checking
procedure is implemented at the end of the iterations, in order
to prove that the control devices are indeed working in the
assumed conditions. Collins queries if such a check should be
167
carried out at the beginning of the iterations or after having
reached a certain degree of convergence ( if so, what is the
"ri ght" degree of convergence ?); clearly, because in the first
iterations, the network can be working in conditions that are
very far from the final ones, the checking procedure can lead to
a completely wrong solution, if it leads to any one at all

(dep ending on the termination criteria).
Chandrashekar (1980), based on graph theory concepts, developed
the models for introducing NW's, CV's and booster pumps into the
network analysis equations (nodal approach). Basically, the
model for PRV's used by Chandrashekar is the same as that used by
Jeppson (1976), adding a pseudo-reservoir whose head is equal to
the PRV setting. This author also gave some warnings about the
possibility of not finding a solution via the application of this
method (and others reported in the literature); he suggested
that an oscillatory behaviour and slow convergence can be
expected due to numerical problems, although he did not encounter
such problems . Chandrashekar also gave some examples of the
application of this procedure; we agree with his results.
Gessler (1981), in a review paper, recognised the complexity
added by PRV's in the network analysis, especially when dealing
with a number of devices simultaneously. According to his review
"at the present time there is no generally accepted procedure for
handling PRV". Gessler insisted that in introducing PRV's within
a nodal formulation, the symmetry of the Jacobian is lost and
that, within a loop formulation, the procedure "is extremely
cumbersome" due to the fact that the state of these devices
168
depends on the heads at the nodes, something which is determined
at the end of the process in the loop approach. Consequently,
Gessler recommended a nodal formulation. He also gave a couple of
detailed examples on how to include PRV's; we agree with his
results, as discussed in a subsequent section. Finally he
addressed the problem of uniqueness, recommending a reassessment
of the physical problem whenever mathematical and numerical
problems are found.
In summary, from the literature reviewed, it becomes apparent

that there are some common characteristics in many of the
existing models for handling pressure controlling devices. All
the authors reviewed, with the only exception of Gessler (1981),
describe the behaviour of the PRV in conjunction with the pipe
where it is inserted. Actually, in so doing, they do not describe
the PRV itself but its effect on the pipe.
All the referenced authors deal only with PRV's, without even
mentioning PSV's or other pressure controlling devices, like
altitude valves, constant flow valves, etc. As it can be seen
later on, in the section corresponding to the examples found in
the literature, not a single example dealing with PSV's was
reported and, more importantly, there is no example dealing with
PRV's and PSV's simultaneously. The reason for this may lie in
the fact that, roughly speaking, PSV's are dealt within a similar way
to PRV's, but instead of the maximum downstream pressure we are
concerned with the minimum uPstream pressure. No reason
whatsoever can explain the lack of evidence on the behaviour of
real networks having more sophisticated controlling devices.
169
All the authors reviewed, with the exceptions of Jeppson (1976)
and Jeppson and Davis (1976), referred to a nodal formulation of
the set of equations. The attempt of Jeppson and Davis is
commendable, in the sense that they focused on all the possible
formulations.
Gessler (1981) recommended the nodal a pproach as the most

suitable for dealing with networks including PRV's. This seems
to be implicitly accepted by most of the authors reviewed, except
Jeppson and Davis (1976). This is in clear contradiction with
the conclusion drawn by Wood (1981a), who recommended either the
simultaneous path method (loop approach) or the linear theory
method.
The question of lack of symmetry in the Jacobian matrix, when

following a nodal approach, when considering the PRV in
conjunction with the pipe, is relevant from the storage point of
view. As far as the solution of the linear system of equations is
concerned, the symmetry issue is relevant from the stability
point of view, and also because it is known that an efficient
method for solving symmetric linear systems of equations does not
need to be as efficient when applied to a non symmetric system
(if applicable at all).
Although it has not been emphasised in the literature, it is

evident that all the methods proposed for dealing with NW's need
to re-assemble the structure of the matrix of the linear system,
every time a device is found to be working in the wrong way.
This is, for example, the case of a CV, which must be closed
170
because its downstream head is found to be greater than the
upstream one; this implies eliminating the branch containing the
CV, re-formulating the topology and re-building the corresponding

Jacobian matrix. The same re-shaping of the equations is
necessary in the case of the PRV's, when sometimes (in the nodal
approach) an additional flow continuity equation must be added to

the original set.
Also relevant are Collins (1980) worries about the risks of
changing the status of a PRV at early stages within the iterative
procedure, when convergence has not been fully achieved. This
seems to be a common problem in existing codes, although not

openly discussed in the literature.
Finally, the question of the uniqueness of the solution was
raised by some authors [Lekane (1979), Collins (1980),
Chandrashekar (1980) and Gessler (1981)], leaving a shadow of
uncertainty over the solutions when dealing with general systems
with multiple PRV's and other devices simultaneously.
All these ideas led us to a feeling of dissatisfaction about
the way pressure controlling devices are handled at the present
moment in the open literature, and thus motivated a search for an
alternative, completely different, approach, which hopefully can
cope with the disadvantages and uncertainties of the reported
methods. Also, we are looking for a method able to take full
advantage of the positive attributes of the gradient method.

4.4. Proposed model for pressure controlling devices.
4.4.1. Needibr an alternative model.
Due to all the shortcomings affecting the modelling of pressure
controlling devices in the existing methods, a new approach is
needed, especially in the context of the gradient method.
Putting the existing methods in perspective, and considering
the physical behaviour of the pressure controlling devices, it is
clear that the fact that all these valves are variable resistance
devices is not explicitly considered in the existing models. All
of them are based on a post-analysis check of whether the device
is working in the assumed mode or not and, as a result, the
existing approaches are more combinatoric than physical, in the
sense that all the possible combinations of status may have to be
checked.
Also, the fact that the state of the pressure controlling
devices is influenced, and indeed determined by the hydraulic
system (i.e. the network as a whole), is not explicitly taken
into account in the dynamics of modelling pressure controlling
devices. We believe that this can lead to some misunderstandings
of the way pressure controlling devices work. Clearly, the fact
that we can fix the PRV setting does not necessarily mean that
this is the pressure we are going to have, since in the final
analysis this will depend on the system characteristics.
We shall follow what we believe is a more physically meaningful

approach to modelling pressure controlling devices, aiming at a
robust algorithm on which we can rely and with which a unique
172
solution can be obtained. We believe that because pressure
controlling valves are in essence variable resistance devices,
consequently dissipative devices, their steady state solution
must be unique. What we shall try to do in our model is to
follow, as closely as possible, what we think should be the
response of the physical water distribution network.
4.4.2. Development of the computer Program .
At the present moment, PRV's, PSV's and CV's have been

introduced into our computer program for the gradient method. In
the future new devices can be introduced using the same basic
model explained in the following paragraphs. For PRV's, PSV's
and CV's we define the following possible operational modes (or
"STATUS" in the notation used in the program):
- Inactive mode: valve fully open, this corresponds to the valve
having a minimum resistance characteristic parameter ao. This
corresponds to STATUS = 0 in the computer program.
- Active mode: this corresponds to the closed mode for a CV
(this is, for CV's, STATUS = 1 in our program). For PRV's and
PSV's this mode has two sub-modes:
* Partly closed mode: the valve has a resistance
characteristic parameter greater than ao and is not
completely closed, so that a flow throu gh the valve
is possible. This corresponds to STATUS = 1 in the
computer program.
* Fully closed mode: no flow can pass through the
valve and its characteristic resistance parameter is
set to an "infinite" value. This sub-mode corresponds
173
to STATUS = 2 in the program.
In addition to its use for pressure controlling valves, this
STATUS index of the branches is used to introduce closed pipes,
on/off pump switches, with minor changes in the data, and without
having to re-define the topology of the whole network.
We also define the imbalance in each pressure controlling
device, which is proportional to the force that will tend to move

the valve from one operational status to another:
- For CV's this imbalance is simply defined as
X = IH j - Hil
- For NW's the imbalance may be:
X = I H setting - H it when the PRV is partly open or

X = lif j - H i l when the PRV is fully closed.
- For PSV's the imbalance may be:
X = I H settin g - Hil when the PSV is partly open or

X = IH j - H i l when the PSV is fully closed.
Our model operates in two main stages:
i) Detection stage : This aims at determining which devices are
working in a status different to that originally assumed; for

example, a CV with current STATUS = 0 and whose outlet head is
higher than the inlet head. Also, we determine the two maximum
imbalances, say X and X*, and the indices of the devices
corresponding to these imbalances, say I and I*. On finishing the
detection throughout the whole network, the index corresponding

to the most imbalanced valve and its imbalance X, and also the
next most imbalanced one (if any) are known. When there is only
174
one valve out of balance, X* is set equal to the accuracy
specified for the problem, which is the maximum head difference
between the outlet head and the setting for a PRV (or inlet head
in the case of a PSV). In addition, this stage identifies what

is the kind of action that has to be taken in the correction
stage.
ii) Correction stage : In modifying the internal resistance of
the device, this part of the algorithm tries to match the
controlled head (outlet head in PRV's, inlet head in PSV's, etc.)
with the setting. For CV's the correction stage simply changes
the characteristic resistance parameter to ao if the valve needs
to be inactive, or to "infinite" if the CV ought to be closed;
the index describing the status of the CV is also changed
accordingly. For PRV's and PSV's the matching procedure is based
on the lumped curve presented in section 4.2 (Figure 4.7 for the
PRV's and Figure 4.9 for the PSV's). This explanation will
relate to the case of the PRV's only, because the case of PSV's
is completely similar.
The correction stage for PRV's and PSV's operates in the

following way. Let us assume that a l is the current resistance
characteristic parameter of the PRV we are trying to correct at
this stage (i.e. this is the maximum imbalanced device). Let us
also represent this situation in Figure 4.11. The lumped curve
suggests that, for finding a*, the a which matches outlet head
and setting, a Newton procedure can be implemented, starting at
the point a l . At this point let X1 = IH1-Hsettingl be the current
imbalance. Since the analytical expression of the function
175
represented by the lumped curve is not known, the derivative of
that function must be computed numerically. To do so, we perturb

a l by a small quantity, say 8a, leading to 0(2 =0(1 + ea. At this
point the evaluation of the outlet head for the perturbed *2 is
needed; this is done by running the program and determining the
corresponding head. Let us call the new outlet head H2, of course
the imbalance at this point becomes X 2 = IH2-Hsettingl, then the
derivative we are looking for can be approximated by:
f' P.: (H1 - H 2)/( 13C 1 - *2)
Hence, the new approximation for a* can be computed using

Newton's method as:
a3 = a l - i(H 1 -Hsetting ) / f ' ]. (12)
With this new a3, a new evaluation of the outlet head is
needed, which means that we have to run the main program again to
produce a new head, say H3, which allows us to determine the

imbalance at this point X-=
4 I H 3 - Hsettingl which must be compared
with the maximum admissible tolerance. At this point of the
procedure, we are faced with two main alternatives, namely carry
on the Newton's iterations until H3 is sufficiently close to
Hsetting, or stop the iteration when X3 becomes smaller than X*
(the second maximum imbalance determined at the detection stage);
since because there is another control valve which is more
imbalanced than the current one, there is no point in carrying on
the Newton iterative procedure on up to the limit. We chose the
second alternative, which was found to be the most adequate, in
order to avoid wasting computational resources. Actually, we
should compute the current value of X* after each change in a,
176
but this was not found worthwhile during the development of the
program. If we find that that we have to stop the Newton
iterations, we go back to the detection stage, in order to check
if any other device has gone out of balance. Checking the valve
status is not as computationally expensive as re-running the
program.
H outlet
H1
H2
X -1 X2
H3
H4
Hsetting
Fig. 4.11. Correction of the resistance characteristic

parameter (a), for a PRV, based on the lumped
curve.
177
In summary, the method for determining the resistance parameter
a* uses the network analysis routine intensivel y . As far as the

network analysis program is concerned, the PRV's and PSV's are
modelled as fixed head loss/flow characteristic devices, since
the changes needed in the resistance (i.e. the sequence
**) are handled outside the analysis routine, as shown
in Figure 4.12.
Figure 4.12 clearly suggests that the proposed approach for
modelling regulating devices can be implemented with any network
analysis algorithm, since we consider the correction of the
resistance parameter a as an independent process from the network
analysis. This means that the proposed algorithm for modelling
pressure regulating devices can be implemented with any other
network analysis algorithm. From the theoretical point of view,
this also means that the pressure regulating devices can be
considered as any other resistive (dissipative) device, hinting
that no problem with the uniqueness and existence of the solution
should be expected when following this approach.
I
I.( fixed ai Program modelling
Standard network pressure regulating
analysis algorithm devices: performs
(gradient method detection and correction
or ) stages, determining ai
Heads: by Newton's method.
Fig. 4.12. Interaction between pressure regulating device model

and the standard network analysis algorithm.
178
To establish the robustness of our proposed model, we have rtAn
it with a number of examples reported in the literature, and the
results are detailed and analysed in the next section.
4.5. Comparison between the results given by the ProPosed model

and some examples found in the literature.
In order to compare some of the results produced by some of the
authors reviewed in previous sections, we have taken the
published data and computed our own results, using a network
analysis program based on the gradient method. Table 4.2 is a
summary of the comparison between the published results and our
results. At the end of Table 4.2. some of our own test examples
are also included.
Surprisingly, we have found disagreement between our results
and those reported by some authors. In this section we shall
analyse in some detail these differences, in order to find out
the reasons (if possible) for such discrepancies.
In the rest of this section we shall concentrate on those
examples where no agreement was detected.
4.5.1. Example JEPPO. from JePPson and Davis (1976).
The network with 15 branches and 14 nodes shown in Fig. 4.13,
including one PRV, was analysed by Jeppson and Davis (1976) under
two possible settings for the PRV:
a) H setting = 149 m
b) H setting = 140 m
179
Table 4.2. Comparison of results reported in the literature and
our results for networks containing pressure control
devices.
Example Source Network Size:Number of Devices Comments

Name (*) Branches Nodes Sources CV PRV PSV
Jepp0 (1) 15 14 2 - 1 - no agreement
Jeppl (2) 10 a 2 - 1 - no agreement
Jepp2 (2) 17 15 2 - 1 - agreement
Jepp12 (2) 15 13 2 - 2 - no agreement
Jeppl3 (2) 15 13 2 - 2 - no agreement
Lekane (3) 24 18 3 1 2 - agreement
Collins (4) 6 7 4 - 2 - no agreement
Chanl (5) 12 10 1 1 1 - agreement
Chan2 (5) 12 10 1 1 1 - agreement
Chan3 (5) 12 10 1 1 1 - agreement
Chan5 (5) 37 28 2 — 1 — agreement
Gessl (6) 14 12 1 - 3 - agreement
Gess2 (6) 14 12 1 - 3 - agreement
Gess3 (6) 21 17 3 1 2 - agreement
,
Check (7) 6 6 2 1 — -
Pry (7) 6 6 2 — 1 —
Psv (7) 6 6 2 — — 1
Prv5 (7) 74 48 4 1 2 2
(*) Sources:
(1) :
Jeppson and Davis (1976)
(2) :
Jeppson (1976), PP. 86, 88, 110 and 112
(3) :
Lekane (1979)
(4) :
Collins (1980)
(5) :
Chandrashekar (1980)
(6) :
Gessler (1981),pp. 88 for gessl and gess2, 91
for gess3.
(7) : our test examples.
Note : In some cases, the number of branches or nodes does not

coincide with the original source, because we use an
initial and final node for every device.
180
Fig. 4.13. Example JEPPO, from Jeppson and Davis (1976).
The comparison between the results presented by Jeppson and

Davis and those computed with our method, corresponding to both
cases a) and b), are shown in Table 4.3. Basically we differ
from Jeppson and Davis in the operational mode determined for the
PRV; thus, when they say that the PRV is active, we find that it
is actually closed (in case a) .
According to our findings, the PRY closes for any setting under
151.34 m. This is the limit where the downstream head of the PRV
is determined by the system, rather than by the PRV
characteristics.
Unless there is a misprint in the data published by Jeppson and

Davis (1976), we could not find a reason for the discrepancy.
Table 4.3. Comparison of results presented by Jeppson and Davis
(1976) and our results, for example JEPPO.
PIEZOMETRIC HEADS (M)

NODE
OUR RESULTS (1) Jeppson and Davis a) Jeppson and Davis b)
1 176.259515 168.95 169.02
2 162.834336 153.44 153.89
3 162.200871 153.42 152.96
4 182.390276 173.81 173.76
5 151.239548 147.88 141.88
6 163.404912 (2)
7 159.146872
8 176.900400
9 189.626238
10 178.033607
11 170.000000 170.00 170.00
12 200.000000 200.00 200.00
13 162.834336 152.88 153.89
14 151.239548 (3) 149.00 141.88
FLOWS PER BRANCH (L/S)

LINK
OUR RESULTS (1) Jeppson and Davis a) Jeppson and Davis b)
1 102.0292 108.0 107.0
2 22.0291 2.6 26.9
3 107.9708 102.0 103.0
4 76.9851 67.3 66.0
5 0.0001 25.4 0.0
6 79.9999 54.6 80.0
7 55.0441 70.7 71.0
8 184.9559 169.0 170.0
9 55.0441 70.7 71.0
10 184.9559 169.0 170.0
11 102.0292 108.0 107.0
12 107.9708 102.0 103.0
13 0.0001 25.4 0.0
14 55.0441 70.7 71.0
15 0.0001 25.4 0.0
Notes:
*****
(1) Our results are exactly the same for both cases a) and b).
(2) These values are not given by Jeppson and Davis.
(3) Node 14 corresponds to the outlet of the PRV and the settings
are 149 m and 140 m, for cases a) and b), respectively. Link
15 corresponds to the PRV.
182
4.5.2. Example JEPPl. from JePPson (1976).
A small network of 10 branches and 8 nodes, including one PRV,
was studied to illustrate the different modes of operation of
these devices. The schematic of this network is presented in Fig.
4.14 and the comparison of the results obtained here with those
of Jeppson is shown in Table 4.4.
Although the numerical results seem to be in agreement, for
this particular example, we have the following comments.
30.48 m
Fi g . 4.14. Example JEPP1, from Jeppson (1976), page 86.
183
Table 4.4. Comparison of the results given by Jeppson (1976)
and our results for example JEPPl.
PIEZOMETRIC HEAD (M) FLOW (L/S)

NODE LINK
OUR RESULTS JEPPSON OUR RESULTS JEPPSON
1 35.740321 35.726 1 29.0793 29.167
2 29.199594 29.197 2 27.8774 27.751
3 29.201681 29.197 3 0.4396 0.481
4 16.756944 17.672 4 27.5547 27.467
5 35.733260 5 27.1839 27.184
6 16.764005 (*) 16.764 6 1.1331 1.161
7 42.851544 7 0.0688 0.085
8 27.431999 27.432 8 1.1331 (**) 1.161
9 30.480000 30.480 9 1.1331 1.161
10 29.0793 29.167
Notes
(*) Node 6 corresponds to the outlet of the PEW, the PRY setting
was 16.764 m.
(**) Link 8 corresponds to the PRV itself.
The pump characteristic curve used by Jeppson has an ascending

limb, which means that we are in one of those cases when two
solutions to the problem are possible, depending on the system
characteristics.
According to Jeppson, the PRY in this example is working in the

active mode, i.e. maintaining the head in node 6 at the preset
value H6= 16.764 m. This is corroborated, in Jeppson's words, by
the fact that
the pressure upstream from the PRV equals 117.19 ft.
and downstream equals 55 ft. Consequently the
assumption used in writing the final loop equation
(Jeppson is referring to the pseudo-loop equation
linking the outlet of the PRY with the reservoir in
184
node 9) above is correct. Had the solution given a
negative flow rate in pipe 6, this assumption would
be incorrect, since the PRV would then have acted as
a check valve and allowed the elevation of the HGL
downstream from the PRV to rise above 55 ft. Should
this have occurred, the the flow rate in pipe 6 would
no longer be unknown, but equal to zero"
In our opinion, Jeppson's statement is wrong, simply because
according to his own results the head in node 4 is greater than
the head in node 6, which automatically means that there should
be some flow from node 4 to node 6, contradictin g his assumption

that the flow is in the opposite direction. Perhaps Jeppson's
algorithm did not detect the situation where apparently the
pseudo-reservoir is overflowing (which happens when Hi > Hy in
Fig. 4.10. b).
4.5.3, Example JEPP12, from JePPson (1976).
Another example provided by Jeppson is a network consisting of
15 branches and 13 nodes, this time with 2 PRV's. The schematic
of this network is shown in Fig. 4.15, and Table 4.5 presents the
comparison between the results published by Jeppson and ours.
Again we differ in the operational mode for the PRV's, because
where Jeppson reports a PEW closed ( branch 12) we found that it
is in the active mode with a flow of 117.3475 l/s.

and our results, for example JEPP12.
PIEZOMETRIC HEAD (M) FLOW (L/S)

NODE LINK
, .
1 292.385632 294.00 1 736.1791 640.2
2 276.698475 277.79 2 283.7416 285.7
3 264.870323 265.48 3 183.7415 185.7
4 263.190571 263.68 4 -83.7415 -85.7
5 250.442362 251.91 5 249.7851 234.4
6 146.965080 237.54 6 -253.5266 -240.2
7 248.274502 251.91 7 117.3475 0.0
8 150.000085 (*) 237.54 8 0.0000 0.0
9 264.870323 265.48 9 202.6525 120.0
10 146.965080 (**) 237.54 10 -86.1792 -190.2
11 284.000067 11 117.3475 0.0
12 300.000000 300.00 12 117.3475 (*) 0.0
13 250.000000 250.00 13 0.0000 (**) 0.0
14 0.0000 0.0
15 249.7851 234.4
Notes:
(*) First PRV is in link 12, outlet node 8 and its setting is
H = 150 m.
(**) Second PRV is in link 13, outlet node 10 and its setting is
H = 145 m.
4.5.4. Example JEPP13. from JePPson (1976).
In this example Jeppson presents a variation of the previous

case, by changing the connectivity of branch 6, from node 4 to
node 2, while maintaining the remaining characteristics of the
network.
Figure 4.16 shows the schematic of the network and Table 4.6
presents the comparison between Jeppson's results and ours.
300 m
80 (I/s) 100 (I/s) 320 (I/s)
Fig. 4.15. Examp le JEPP12, from Jeppson (1976), page 110.
300 m
Fig. 4.16. Example JEPP13, from Jeppson (1976), page 112.
187
and our results, for example JEPP13.
PIEZOMETRIC HEAD (M) FLOW ( L/S)

NODE LINK
1 290.081524 291.745 1 849.3089 761.1
2 258.310816 266.422 2 415.6674 364.0
3 259.522824 266.583 3 -52.6691 -16.9
4 264.547302 269.737 4 152.6691 116.9
5 252.034725 250.718 5 232.6691 196.9
6 146.885112 146.781 6 -368.3366 -280.9
7 249.809812 248.419 7 119.0277 119.7
8 149.999990 (*) 266.583 8 0.0000 0.0
9 259.522824 266.583 9 200.9723 200.3
10 146.885112 (**) 10 -199.3087 -111.1
11 282.805791 11 119.0277 119.7
12 300.000000 300.000 12 119.0277 (*) 119.7
13 250.000000 250.000 13 0.0000 (**) 0.0
14 0.0000 0.0
15 232.6691 196.9
Notes:
(*) First PRV is in link 12, outlet node 8 and its setting is
H = 150 m.
(**) Second PRV is in link 13, outlet node 10 and its settin g is
H = 145 m.
As before, there is disagreement, although this time the status of

the PRV's is the same in both cases; however, the heads and flows
do not correspond.
As a general comment on Jeppson's examples, we can say that he

has been dealing with networks which are normally affected by
backpressure, which is provided by some pipe connecting high
pressure zones of the network with lower pressure zones (i.e.
with the outlet of the PRV). This is a situation which is
inconsistent with the objective of a PRV and is bound to produce
problems both in the real network and in its model.
188
4 5 5. Example COLLINS. from Collins (1980)
On presenting a case for his statement that non-unique
solutions can be found in some networks, especially when
pressure controlling devices are included, Collins (1980)
introduced the network of Figure 4.17, consisting of 6 branches
(3 Pumps, 2 PRV's and 1 valve) and 7 nodes (4 of them
reservoirs). As an additional condition Collins, established
that each pump has a non-return device included, so that flow is
only possible in the positive direction; also, the pumps are
equipped with a device which closes when the downstream head of a
pump is lower than the upstream head. According to our model,
the PRV's are always acting as. non-return (CV) valves; thus, the
problem can be modelled without any additional feature. The
results presented by Collins (1980) are summarised in Table 4.7.
Collins analyses the problem following a combinatoric approach,
assuming that the operational mode of the PRV's is an independent
variable of the problem; the same assumption is made for the
pumps. This led Collins to assume that 16 potential operating
modes are possible in the system, corresponding to all the
combinations of PRV and pump states. We do not agree with this
assumption, since clearly the final status of the PRV's will be
determined bY the system ,once their settings are established by
the operator. In this problem, since no automatic control is
acting on the pumps, in order to start or stop them (for example
according to the reservoir levels or some nodal p ressures), their
operating mode is controlled by the operator,i.e. it is an

external (independent) variable. In this context, the 16
189
Table 4.7. Network solutions for example COLLINS, from Collins
(1980).
-
Operating modes Network
M of control elements solution (a)
o
D PRV's Pumps Head (m) Flowrates (l/s)
E status P 1P 2 P 3 H6 Q26 Q46 Q56 Q67
1 Passive on on off 238.0476 3.5813 0.8927 0.0000 4.4739
2 Passive on on on No solution
3 Passive on off off 190.0000 2.0000 0.0000 0.0000 2.0000
4 Passive off on off No solution
5 Passive off off off No solution
6 Passive off on on No solution
7 Passive off off on No solution
8 Passive on off on No solution
9 Active on on off No solution
10 Active on off off No solution
11 Active off on off No solution
12 Active off off off No solution
13 Active on off on No solution
14 Active off on on No solution
15 Active off off on No solution
16 Active on on on 245.0000 --(b) --(b) 1.0000 4.7258
_
Notes: (;,6"- 1 A
#.7,12 = Q26 ; Q34 = Q46
t ° 1 '426 + Q46 = 3.7526
215 rn
Fig. 4.17. Example COLLINS, from Collins (1980).
190
possible operating modes suggested by Collins are reduced to
eight cases (see Table 4.8), this number being the possible
combination of operational modes of the 3 pumps (with two

exclusive modes: on/off).
Of course the eighth operational mode is trivial, because no
flow can be possible, unless we accept that the pumps can operate
as turbines, in which case some flow can be established
throughout the pump in branch 5. This case is clearly outside
the scope of this example.
Table 4.9 presents the results of running our program in the
first seven modes defined above, assuming that the PRV's are set
to a head of 245 m in their outlet (node 6).
According to Collins (1980), of the 16 possible operational
modes he identified, only 3 of them are feasible and compatible
with the conditions imposed to the problem. Only in his mode 16,
which corresponds to our case 1, we get some similar results,
although he finds that the flows between nodes 2,6 and between
4,6 admit more than one solution. In our case we found that this
Table 4.8. Operational cases for example COLLINS, based on all

the possible combinations of pump operating modes.
Case Pump 1 Pump 2 Pump 3

(link 1) (link 3) (link 5)
1 on on on
2 on on off
3 on off on
4 on off off
5 off on on
6 off on off
7 off off on
8 off off off
,
191
solution in also unique, the flows being controlled by the
difference in the levels (available potential energy of the
reservoirs connected to nodes 3 and 1, see Fig.4.17). In his mode
1, which is equivalent to our case 2, the head in node 6 was
238.0476 m, while we found that this head is the PRV's setting.
In his mode 3, equivalent to our case 4, Collins found that the
head in node 6 is 190.0000 while we obtained 205.000. We
reproduced exactly Collins results in modes 3 and 6, provided
that we change the settings of the PRV's to the heads found by
Collins, but this is clearly a change in the original settings
adopted for the PRV and, therefore, is another problem.
Collins (1980) did not consider as a possible operating mode

the case where one PRV is active while the other is passive,
which really happens in our case 5, thus confirming the weakness
of his approach.
Table 4.9. Results of running our program for the operational

modes possible in example COLLINS.
#
C M PRV status Flow (1/s) Head (m)
A0
S D PRV2P1W4 Q26 Q46 Q56 Q67 H2 H4 H6
E E
1 16 act. act. 1.0000 2.7258 1.0000 4.7258 245.00 277.85 245.00
2 1 act. act. 0.9982 3.7276 0.0000 4.7258 245.02 245.52 245.00
3 mac mac 1.5709 0.0010 2.8876 4.4595 237.66 237.66 237.66
4 3 mac mac 3.0000 0.0000 0.0000 3.0000 205.00 215.00 205.00
5 mac act. 0.0001 3.7257 1.0000 4.7258 245.00 245.59 245.00
6 mac mac 0.0005 4.1379 0.0005 4.1388 229.39 229.39 229.39
7 mac mac 0.0005 0.0005 4.1224 4.1224 229.01 229.01 229.01
8 pumps off, not considered.
Case : refers to our 8 operational cases.
Mode : refers to Collins' 16 operational modes.
Note that = Q26 and Q34 = Q46
192
In our opinion, what is happening in this example is a
conceptual problem at the formulation stage, in the sense that
the status of the PRV's cannot be determined a priori , but are
determined bY the hydraulic system under study, accordin g to the

PRV's settings.
4.6. Concluding remarks.
We have proposed a physically-based model for representing
pressure regulating devices, which apparently does not have the
problems detected and documented in the previous section for some

traditional formulations.
Although the convergence of the model is rather slow when
analysing a network from scratch, i.e. without any knowledge of
the previous state of the network and its regulating devices, we
believe that this apparent shortcoming is not relevant when
applying this program for continuous simulation and operational
purposes, when the final solution of the previous stage can be
used as an initial solution for the next stage. Also, because
the changes in heads and flows with respect to time are
relatively slow, the changes in the operational mode of pressure
regulating devices should be slow as well, which implies that, in
correctly operated networks, the pressure regulating valves are
not changing their status too often.
We believe that, in this section, we have proven that, when
dealing with regulating devices, reliability and robustness
rather than speed should be the major concern, because obviously
193
there is no point in having a fast program leading to the wrong
solution.
Nevertheless, in order to reduce the execution time of the
proposed algorithm we concentrate in the next chapter on the

search for efficient fast direct linear solvers.
The main advantage of our method seems to be its robustness, as
shown in Table 4.2, in detecting problems in the results given by
other programs. The algorithm is able to deal with multiple
pressure regulating devices simultaneously, which is something
that is not clearly found in the open literature, as can be seen
from Table 4.2.
We also have to emphasise the fact that the algorithm does not
need a re-definition of the network topology when a regulating
valve changes its status; as a result, the linear system of
equations remains symmetric all the time, which is efficient from
the storage and computational point of view.
We recognise the need to include other regulating devices, like
flow-modulated PRV's, double relay valves (PRV's with two
different settings, one for the day and the other for the night),
altitude valves, constant flow valves, etc., some of them shown
in Figure 4.18. This should also be done in the future to expand
the capabilities of our program.
194
1 Valve Element
2 Upper Cylinder
3 Isolating Cocks
4 Strainer
5 Orifice
6 Needle Cock
7 Indicator
8 Relay Valve AS FOR 1301
9 Oiaphragm
10 Spring
11 Adjusting Screw
12 Pressure Gauge
13 Alternative Constant
Pressure Connection
14 Electric Motor
15 Solenoid operated valve
A. Low Pressure Relay Valve

a) Valve No. 1303 Valves with two B. High Pressure Relay Valve
relay systems
1 Valve Element
2 Upper Cylinder
3 Isolating Cocks
4 Strainer
5 Orifice
6 Needle Cocks
7 Indicator
8 Riley Valve
9 Diaphragm
10 Spring
11 Adjusting Screw
12 Bill Cock
13 Non-Return Valve
b)
Valve No. 1330 Standard Altitude Valve — Flow into tank only
Valve No. 1331 Altitude Valve allowing flow in and out of tank
(Illustrated by dotted lines)
1 Valve Element
2 Upper Cylinder
3 Isolating Cocks
4 Strainer
5 Pressure Orifice
6 Needle Cock
7 Indicator
8 Relay Valve
flow 9 Diaphragm
=-4.-
10 Spring
11 Adjusting Screw
12 Flow Orifice
c)
Valve No. 1340 Standard Constant Flow Valve
Fig. 4.18. Other pressure regulating valves, from Glenfield &

Kennedy Ltd., Pub. 215/R3.
195
CHAPTER FIVE
EFFICIENT COMPUTER PROGRAM IMPLEMENTATION OF THE GRADIENT METHOD
FOR WATER SUPPLY DISTRIBUTION NETWORKS
In the computer implementation of the gradient method, some
issues become relevant from the computational point of view,
especially if an efficient microcomputer-oriented implementation

is sought. In this Chapter we are mainly concerned, for example,
with the generation of an initial flow distribution and the
possibility of using previous flow solutions as a starting point
for a new network analysis. Another relevant issue, which perhaps
is instrumental in the performance of the gradient method, is the
numerical scheme for the solution of the linear system of
equations generated by the gradient method. Due to the successive
linearizations carried out within the gradient method, each
network analysis implies the solution of a number of linear
systems, typically 6-10.
In this chapter we intend to study these issues, particularly
the solution of the linear system of equations, which is the most
time-consuming step in the gradient method.
5,1. Automatic generation of initial flow distribution.
Unlike other network analysis algorithms reviewed in Chapter
Two, the gradient method does not require an initially balanced
flow distribution. In fact the algorithm itself is able to
generate its own initial flows, without requiring to input the
flows through the data set. This is possible because the coupled
196
equations which are the basis of the gradient method
automatically re-balance the network at each iteration. These
equations are:
Heads update:
H(1-' 1)= -[ A 21 G-1 A l2] -1 f A2 1 G-1 ( A ll ii(i) + A lo H o) -

- ( A 21 Q(i) - a ) 1 (1)
Flows update:
Ei(j+1)= (I- G-1 A ll ) Q(i) - G-1 ( A l2 H(i+1) + A 10 HO) (2)
Equation (1) represents a linear system of NN-NS equations in
the unknown piezometric heads, and it can be reordered to the
standard format of a linear system, by premultiplying equation
(1) by the matrix [A21 G- 1-12,,

A 1 which becomes the matrix of
coefficients of the linear system. For more details on the
derivation of the gradient method see Chapters Two (original
Todini's formulation) and Three (extended version).
Indeed, the last term of equation (1), A 21 Q (i )-g, allows us to
start the first iteration with any flow Q( i) , since it accounts
for the flow imbalance produced by Q (i) . Equation (2) guarantees
that, with the heads computed via (1), the next flows Q( i+1) will
fulfil the nodal balance condition. This can be easily proved by
premultiplying equation (2) by A21 1 and verifying that the result
is always the nodal demand, i.e.:
A21 ii(i+1) = a (3)
197
Thus, after the first iteration, equation (2) produces a
balanced flow distribution [Q( 1+1 )] which, when used during the
second iteration [Q (i) in equation (1)], allows the nodal
imbalance to vanish.
In the present computer implementation of the gradient method
we start with Q = 1 (1/s) in the first iteration although, for
large problems, a great deal of computer resources can be spared
if we have stored a previous flow solution, not necessarily

balanced for the new analysis conditions.
This feature of the gradient method represents the main
difference with respect to the "integrated mesh-nodal" method of
Stimson and Brameller (1981), which requires the generation of
the tree and co-tree of the network, and the corresponding
partitioning of the flow vector into dependent and independent
flows. The dependent flows, computed from the independent ones,
are then forced to be balanced before starting a new iteration.
These interesting properties of the gradient method allow us to
reduce considerably the execution time in applications like the
extended period simulation, where the flow solution of the last
time-step is used as the initial (non-balanced) flow solution for
the next time-step. Also, this helps us to reduce considerably
the execution time in the modelling of pressure regulating
devices (see Chapter Four for details), where the previous state
of the system (flows and regulating valve's resistance) is used
to obtain a new approximation to the equilibrium of the

reg ulating valves.
198
As a result, the problems of generating an initial flow
solution, and that of using an already existing flow solution to
start a new analysis, can be efficiently handled within the
gradient algorithm itself, without requiring additional

programming.
The remaining problem from the computational standpoint is,
then, the solution of the linear systems of equations generated

by the gradient algorithm.
5 2. Solution of the linear system of equations generated by the

gradient method.
In Appendix A we have reviewed the main approaches to the
solution of linear systems of equations applicable to sparse
symmetric positive definite systems. The two basic alternatives

are direct and iterative methods.
5.2.1. General overview of direct sparse linear solvers.
All the direct methods for the solution of linear systems of
equations are based on Gaussian elimination which, as shown in
Appendix A, consists in a factorization stage followed by a
substitution stage. The Cholesky factorization is the most common
factorization scheme for symmetric positive definite matrices.
It is in the factorization stage where the problem of fill-in
comes about, creating new non-zeros in the Cholesky factors with
respect to the amount of non-zeros of the original matrix. The
p roblem of the amount of fill-in can be handled via appropriate
re-ordering of the original system; different orderings are
199
possible, leadin g to more or less efficient factorizations,
ranging from a completely filled factor up to one with no fill-in

at all, as shown via a small extreme example in Appendix A. Thus,
the efficient handling of the fill-in problem is basically
concerned with finding the most adequate re-ordering of the
original system of equations. This task is eased by the
introduction of graph theory concepts.
Because, in symmetric positive definite systems, the Cholesky
factorization is inherently stable, we can concentrate our
attention on the problem of reducing the fill-in, leaving the
stability problem aside. This allows us to split up the direct
solution of a linear system into a three stage process:
a) ANALYSIS stage: which searches for the re-ordering of the
original system (or, equivalently, the re-labelling of the system
graph), which produces the minimum fill-in. Also, at this stage,
the data structures required are set up; this refers to all the
auxiliary vectors and pointers required to handle the numerical
parts of the algorithm.
b) FACTORIZATION stage: which carries out the numerical part of

the factorization, using the ordering and data structures
generated in the previous stages.
c) SOLUTION stage: which actually solves the linear system by
substitution, using the Cholesky factorization produced in the
previous stage, for any right-hand-side vector specified.
In Appendix A we have followed the sequence corresponding to
the historical development of sparse direct linear solvers,
comprising:
200
- Band methods.
- The envelope ("skyline") method.
- Minimum degree algorithms.
- Quotient tree algorithms.
- One-way dissection method.

- Nested dissection method.
- Frontal and multifrontal methods.
All these methods follow an increasing degree of sophistication
and complexity, in parallel with an increasing degree of
computational efficiency, both in terms of storage and speed.
Band and envelope methods are closely related, since both of
them are based on the observation that fill-in occurs within the
border lines separating the non-zeros and zeros of the matrix
(see Figure 5.1). Thus, roughly speaking, the problem becomes
that of interchanging the rows and columns of the original
matrix, in such a way that the "area" of the non-zero region is

minimized.
the objective is to
reduce this "area".
A=
4m.
Fig. 5.1. Target "area" for band and envelope methods.
201
Although, for general matrices, the envelope methods are better
than the band methods, the main shortcoming is concerned with the
fact that the amount of fill-in is unpredictable, before actually
running the numerical part of the algorithm. This means that we
are constantly threatened with running out of storage, unless a
large overhead is initially specified, leading to a sub-optimal

storage utilisation.
As a contribution to solving the storage "predictability"
problem, and to exploit even further some zeros still unexploited
by the envelope methods, the minimum degree algorithm has been
proposed. The minimum degree-based methods choose the next pivot,
in the Gauss elimination process, from the row (or column) with
the minimum number of non-zeros in it. A so-called "hybrid"
implementation, due to George and Liu (1981) allows us, during
the ANALYSIS stage, to determine the amount of memory required in
the factorization stage, which happens to be not greater than
that required to store the original matrix. Our experience with
this algorithm, within the context of the gradient method, has
been quite bad, particularly for networks larger than 900 nodes,
where it actually failed.
Quotient tree algorithms take advantage of the fact that, on
joining the neighbouring already eliminated nodes of the matrix'
graph into "supernodes", a much simpler tree-like structure of
"supernodes" is obtained. This "supertree" induces a partition of
the original matrix, which can be efficiently handled by blocks,
together with an implicit scheme, where only the diagonal blocks
of the Cholesky factorization and the off-diagonal blocks of the
202
original matrix are explicitly needed, the rest being handled
implicitly.
Dissection algorithms exploit the idea of partitioning the
matrix even further. In one-way dissection the whole matrix graph
is re-numbered, slicing it up via "dissectors", which are subsets
of nodes running either horizontally or vertically, cutting the
graph through the shorter path (i.e. in only one direction). This
induces a reordered matrix structure which has low fill-in. The
nested dissection algorithm exploits the dissection idea in two
directions, creating a matrix which leads to very fast solutions,

though requiring more storage than the one-way dissection.
The frontal method has originally been produced for finite
element applications, and it takes advantage of the way the
global matrix is assembled; thus, a row (node) can be eliminated
as soon as all the elements connected to it have been assembled.
An element numbering system, related to the order of assembling
the global matrix, keeps the front-width as short as possible
and, in multifrontal methods, eventually allows more than one
front to be handled simultaneously, increasing the efficiency of
the algorithm.
5.2.2. General overview of iterative methods for solving linear
systems of equations.
Iterative methods for the solution of linear systems of
equations have the common aim of constructing a sequence of
vectors: x(o) , x(1) , x(2),....,x(k), converging to the solution
of the linear problem A x = h.

203
Traditionally, iterative methods like Jacobi, Gauss-Seidel and
successive over (or under) relaxation have been used to solve
sparse linear systems. This is due to the fact that, in all
iterative methods, sparsity is handled quite naturally and no
fill-in occurs. Unless certain specific conditions are met by the
matrix of coefficients (see Appendix A for details), some of
these iterative methods may fail to converge, or converge slowly.
The conjugate gradient method, developed in the early 1950's by
Hestenes and Stiefel (1952), and re-vitalised in the seventies by
Reid (1971), has gained a reputation as an efficient linear
solver. This is basically an unconstrained minimization algorithm
which, when applied to the quadratic function:

z = k xT A x 4. xT 12 (4)
produces the solution of the original linear system A x = h.
Starting from any initial solution (x( o) ), the method ideally
converges in a fixed number of steps, though roundoff errors
spoil the convergence, requiring in practice some additional
iterations. Convergence is guaranteed for symmetric positive
definite systems.
Acceleration of the convergence of the conjugate gradient

al g orithm can be achieved via a transformation, consisting of
p re-multiplying the original matrix by a "pre-conditioning
matrix" ( L -1 ), such that:
( L -1 A LT ) ( L T x ) = L -1 12 (6)
where the new transformed matrix ( L -1 A L -T ) is still
symmetric, and is better conditioned than the original one. The
204
same original conjugate gradient algorithm, when applied to the
transformed problem (5), is known as the preconditioned (or

modified) conjugate gradient method.
In selecting a preconditioning matrix ( L -1 ), the problem of
fill-in strikes back, as in the direct linear solvers, but an
incomplete Cholesky factorization scheme proposed by Kershaw
(1978), allows us to decompose the original matrix
A z L LT (6)
where L has exactly the same sparsity pattern as A. The non-zeros
outside the original pattern are simply thrown away, thus solving
the fill-in problem. Note that with (6), the transformed matrix
( L -1 A L -T ) in (5) becomes nearly an identity matrix, thus
being easier to invert than the original matrix A.
5.2.3. Comparison of the Performance of the gradient method
with different linear solvers.
In sections 5.2.1. and 5.2.2. the two main alternatives for the
solution of sparse linear systems have been briefly reviewed,
namely direct and iterative methods.
In the present section we intend to compare the performance of
six sparse direct and one iterative linear solvers, in the
context of the gradient method for water distribution network
analysis. The algorithms compared are:

- The envelope ("skyline") method.
- Minimum degree algorithms.
- Quotient tree algorithms.
- One-way dissection method.
205
- Nested dissection method.
- Multifrontal method.
- Preconditioned conjugate gradient method.
Numerous comparisons between different algorithms have been

presented in the literature , though for different kinds of
problems and of limited scope (two or three methods) . George and
Liu (1981) carried out an extensive comparison, geared towards
finite element problems, although their comparison concentrated
mostly on direct methods, excluding the frontal and multifrontal
approaches.
We have taken the comparison carried out by George and Liu
(1981) as a starting point, and we have extended it to include
the preconditioned conjugate gradient, recognised as the best
iterative algorithm available, and also a multifrontal method,
which corresponds to the routine MA27 of the Harwell Library. In
so doing, we are covering the widest spectrum of linear solvers
possible, which is something unavailable in the literature for
such problems, particularly for the case of water distribution

networks.
This has led to the implementation of seven different versions
of the gradient method, each one with a different linear solver.
A bank of nine network examples has been set up , ranging from 48
to 4,900 nodes, and the corresponding analysis problems have been
solved, keeping track of the execution time, number of iterations
of the gradient method and storage required.
206
Table 5.1. presents the main characteristics of the examples
used in the numerical comparison, while Table 5.2 presents the
results, including the execution times of the tested programs in
the Amdahl 5860 mainframe computer at Newcastle University; these
Table 5.1. Main data corresponding to examples used in the

comparison between different linear solvers in
the gradient method for pipe network analysis .
NETWORK DATA
Data file Branches Nodes Sources Equations CV PRV PSV
net5.dat 74 48 2 46 0 0 0
[1]
prv5.dat 74 48 4 44 1 2 2
[2]
bofn.dat 298 266 26 240 0 0 0
net50.dat 420 225 2 223 0 0 0
[4]
net51.dat 760 400 2 398 0 0 0
[4]
net52.dat 1200 625 4 621 0 0 0
[4]
net53.dat 1740 900 4 896 0 0 0
[4]
net54.dat 4900 2500 4 2496 0 0 0
[4]
net56.dat 9660 4900 4 4896 0 0 0
[4]
Notes:
[1] : Data from real network, published by Chin et al. (1978),
"Solution of water networks by sparse matrix methods ",
Int. J. Num. Meth. Eng., 12, 1261-1277, (1978).
[2] : Based on the data of previous example, with pressure
regulating devices added.
[3] : Real main network of Bogota, Colombia. Source: G. Gonzalez
"The gradient method", M. Sc. Dissertation, Civil Eng.
Department, University of Newcastle-upon-Tyne, 1987.
[4] : Synthetic network. Network generated as a square mesh with
SQRT(Nodes) vertices per side.
207
Table 5.2. Comparison of execution times and number of
iterations required by different linear solvers
in the gradient method for pipe network analysis.
EXECUTION TIME (secs.) AND NUMBER OF ITERATIONS (*)

Data file
ICCG RCM QMD RQT 1WD ND MA27
net5.dat 0.286 0.112 0.106 0.166 0.442 0.113 0.115
(7)(*) (7) (7) (7) (7) (7) (7)
prv5.dat 8.624 2.374 2.479 4.504 17.547 2.472 2.877
(232) (232) (232) (232) (232) (232) (232)
bog10.dat 18.524 2.152 2.492 2.332 3.220 1.824 1.785
(25) (18) (24) (18) (18) (17) (19)
net50.dat 3.698 0.924 1.812 1.121 1.452 0.907 0.884
(7) (7) (7) (7) (7) (7) (7)
net51.dat 12.126 2.873 10.036 3.310 3.642 2.684 2.234
(9) (9) (9) (9) (9) (9) (9)
net52.dat 36.428 7.670 37.677 8.658 10.965 6.836 5.269
(12) (12) (12) (12) (12) (12) (12)
net53.dat 63.985 13.471 14.898 14.747 11.655 8.598
(11) (11) (**) (12) (11) (11) (11)
net54.dat 627.275 118.850 - 130.774 116.656 95.363 55.146
(17) (17) (17) (17) (17) (17)
net56.dat 2321.573 424.489 - 461.167 386.721 326.817 156.421
(17) (17) (17) (17) (17) (17)
Notes:
(*) : Time is CPU-time in the Amdahl 5860 mainframe computer at
Newcastle University; in rounded brackets the number of
iterations required (gradient method iterations) for the
solution of the water distribution network analysis problem
is shown.
(**): The program failed to find a solution for the first
iteration after running for 970.0 secs., the program was
interrupted and cancelled. No attempt was made with the
larger examples NET54.DAT and NET56.DAT.
Linear solvers:
ICCG : modified conjugate gradient method, with Incomplete
.Choleski factorization ( Kershaw factorization).
RCM : envelope ("skyline") method with Reverse Cuthill-McKee
ordering.
QMD : Quotient Minimum Degree algorithm.
RQT : Refined Quotient Iree algorithm.
1WD : one-Eay Dissection method.
ND : Nested Dissection algorithm.
MA27 : Harwell subroutine MA27.
References:
ICCG : based on: Ajiz and Jennings (1984) and Kershaw(1978).
RCM, QMD, RQT, 1WD and ND : George and Liu (1981).
MA27 : Harwell Subroutine Library, Atomic Energy Research
Establishment, (AERE), U. K.
208
execution times are, in general, average times of at least a
couple of runs and correspond to the time needed for the solution
of the network (i.e. the non-linear problem) excluding the

in p ut/output of data and results. The number of (non-linear)
iterations of the gradient method, for the analysis of the
networks, is also presented in Table 5.2, in rounded brackets.
Table 5.3. presents the summary of the storage requirements,
both in terms of main storage (that needed for the explicit
handling of the non-zero values) and overhead storage, which
accounts for all the extra auxiliary vectors and pointers needed
to implement the algorithm.
Figures 5.2 and 5.3 show the results of the comparison, in
terms of CPU execution time and total storage, respectively. The
storage shown in Table 5.3 and Fig. 5.3 refers to that required
by the linear solver itself; it does not include that of the non-
linear solver. Thus, for comparative purposes the results are
representative of the relative merits of the different algorithms
for the solution of the linear systems.
Another important point in the comparison, is the fact that
most of the test examples are synthetically generated networks.
This is a factor which may be contributing to the good
performance of the dissection methods, which have been originated
for use with regular finite element meshes. This is a point that
should be explored in the future, if data from larger real

networks becomes available.
209
Table 5.3. Comparison of storage requirements (Bytes) for
different linear solvers used with the gradient
method for pipe network analysis.
Method Data file name and number of equations [2)

[1] NET5.DAT PRV5.DAT 80610.DAT hET5O.DAT NET51.DAT hE152.DAT NET53.DAT NET54.DAT NET55.DAT
46 44 240 223 398 621 896 2496 4896
2,576 [3] 2,464 13,440 12,488 22,288 34,776 50,176 139,776 274,176
ICCG 1,748 [4] 1,672 9,120 8,474 15,124 23,598 34,048 94,848 186,048
------
4,328 [5] 4,136 22,560 20,962 37,412 53,374 84,224 234,624 460,224
1,912 1,248 15,664 22,072 50,800 95,224 161,528 715,624 1,925,944

P01 1,656 1,564 8,640 8,028 14,324 22,356 32,256 89,856 176,256
----- ----- ------ ------ --
3,568 2,832 24,304 30,100 65,124 117,580 193,724 805,480 2,102,200
1,760 1,752 9,072 37,336 113,392 237,496 - [6) - -

3,310 3,190 17,394 18,906 37,866 60,256 - - -
----- ----- ------ ------
5,070 4,942 26,466 56,242 151,258 297,752 - - -
1,472 1,424 8,(:• 14,876 46,128 58,448 96,576 402,048 1,050,128

ROT 3,342 3,198 17,290 16,442 32,830 45,854 66,184 184,504 362,024
----- ----- ------ ------ ------ - ------- ---------
4,814 4,622 25,378 31,338 78,958 104,302 162,760 536,552 1,412,152
1,912 1,840 7,512 12,472 24,992 42,928 67,696 237,384 546,088

1WD 2,852 2,728 15,042 14,148 25,196 39,262 56,608 157,280 307,960
-----
4,764 4,568 22,554 26,620 50,1:: 82,190 124,304 394,664 854,048
1,912 1,736 8,624 17,800 37,192 62,848 97,288 327,552 720,144

ND 3,660 3,490 19,140 18,878 33,930 53,016 77,054 215,202 425,324
----- ------ ------ ------ ----- -------
5,572 5,226 27,964 36,678 71,122 115,864 174,342 542,754 1,145,468
1,296 1,248 5,904 17,056 33,296 57,656 95,068 336,::: 728,608

ti127 3,864 3,696 20,160 18,732 33,432 52,164 75,264 209,664 411,264
5,160 4,944 26,064 35,7:: 66,728 109,820 170,352 546,552 1,139,872
Notes:
[1] : Refers to the linear solver algorithm:
ICES : modified Conjugate Gradient method, with Incomplete Choleski factorization ( Kershaw factorization).
RCM : envelope ("skyline') method with Reverse Cuthill-McKee ordering.
(16) : Datient Minimum Degree algorithm.
ROT : Refined Ouotient Tree algorithm.
1WD : one-Way Dissection method.
ND : Nested Dissection algorithm.
MA27 : Harwell subroutine MA27.
[2] : Corresponds with data sets in Table 5.1.
[3] : Upper figure corresponds to the primary storage (Bytes).
[4] : Corresponds to the overhead storage (Bytes).
[5] : Corresponds to the total storage: primary 4. overhead storage (Bytes).
[6] : No solution for the first iteration after 970.0 secs., program was interrupted and cancelled. No attempt was made with
the larger examples NET54.DAT and NET56.DAT.
.c
4-)
••••n
•••••n
• n••••••
1t11111111 1 111111111 1 11 11111111111111111111 11111110

0 0 0 0 0
0 0 0 0
In 0 If) 0 1.0
(soas) awn. uonnoaxa ndo
211
i
0)
C
0
i
I- (N 0 0 47-;
0 < 0 0
c .1-- 0 0
0
D
0-
0 CD
*
lt
11)
% Lf-
1
%
%
0
1
. 1
a)
%. 1 ..0
0
%
.. 1
E
n
Ct
.
. 0z n
.
.
1
1
0
0
.
. 1 d- as
.
. 1 .4
4-)
. 1
% 1 e4.4
.
0
r.t4
o o o o o
o o o o o
IL) o in o to
N C\1 ,-- .—
(sal.Kq)i) abwols iDloi
212
There is an additional point that is worth noting, which is
that the presented execution times are not "exactly" a function
of the number of nodes of the network (or the number of unknowns
in the linear systems). The first four examples illustrate this
point: the presence of pressure regulating valves in the data set
PRV5.DAT, which is basically the same network as in NET5.DAT,
produces a sharp increase in the execution time; this is due to
the increase in non-linear iterations needed in order to produce
the balance in the regulating devices (232 iterations). The point
is also illustrated in the third and fourth examples, where a
real network of 266 nodes corresponding to the main distribution
system of the city of Bogota (Colombia) requires between 1.6 to 5
times the execution time required for a similar synthetic network
(the latter ratio relate to the case when the preconditioned
conjugate gradient is used as the linear solver). In summary, the
presence of pressure regulating devices requires the network to
be solved more times than when no pressure regulating valves are
present, and the hydraulic characteristics of the network (mainly
represented by the distribution and connectivity of high and low
resistance pipes) produces problems resulting in slower
convergence.
The results in Table 5.2 show that, despite the problems
mentioned in the previous paragraph, the relative difficulty of
similar sized problems does not change the relative performance
of the methods, i.e. "good" methods are still good when a more
difficult problem has to be solved. Thus, we can rely on the
synthetic data sets for determining the relative performance of
213
the algorithms, while the "true" performance of the algorithms in
real networks is something which will depend very much on the

physical characteristics of the system (i.e. presence of
regulating valves, distribution of high/low resistance pipes,
etc.).
5 2.4. Selection of the most efficient method.
One of the main conclusions to be drawn from the comparison
carried out is in terms of the relative efficiency of direct
methods when compared to the iterative algorithm (ICCG). For
large problems, say of more than 200 nodes, a direct solver
always produces faster results, the exception being the quotient
minimum degree (QMD) algorithm. In addition, the larger the
problem, the greater the relative efficiency of the best direct
algorithm in comparison with the iterative scheme ( the MA27 code
produces a solution about 15 times faster than the preconditioned
conjugate gradient algorithm, for the network of 4900 nodes ).
As it can be seen form Fig. 5.2, the execution times produced
with most of the direct solvers increase at a nearly linear rate
with the size of the problem; the preconditioned conjugate
gradient shows the strongest non-linear behaviour in this
respect. The total storage (see Fig. 5.3) behaves linearly for
most of the best algorithms, e.g. preconditioned conjugate
g radient, dissection methods and multifrontal algorithms.
In summary, the conclusion is that when selecting a linear
solver, if storage is available, a direct method is the right
decision. The multifrontal method is the best choice from the
214
execution time viewpoint, followed by the dissection algorithms
of George and Liu (1981). The nested dissection is about 100%
more time-consuming for large problems than the multifrontal
method, while the one-way dissection is about 120% more time-
consuming than the multifrontal algorithm. From the storage point
of view, the preconditioned conjugate gradient is the best
choice, requiring about a half of the storage of its closest
rival, the one way dissection algorithm.
Even in a limited storage environment, like in the present
generation of 8 bit microcomputers, a direct solver can still be
the best choice. Our current implementation of the gradient
method with the one-way dissection linear solver, in a 512 Kbytes
RAM microcomputer, allows us to solve networks of up to 1000
nodes in a fourth of the time required by the preconditioned
conjugate gradient. Because of the reduction in the storage
needed for the preconditioned conjugate gradient, networks of up
to 1200 nodes can be solved in a 512 Kbytes microcomputer. The
arrival of more powerful personal microcomputers, which is
already taking place, makes the case for direct solvers even
stronger.
215
In this chapter, we have compared the performance of the

gradient method for water distribution network analysis when used
with a wide variety of linear solvers, including sparse direct
methods and one iterative method (preconditioned conjugate
gradient).
The main conclusion is that, when enough storage is available,
the multifrontal approach (routine MA27 of the Harwell Library)
provides the fastest solution. If storage is very limited, the
preconditioned conjugate gradient is the best choice. A good
compromise between storage and speed is obtained with the one-way
dissection algorithm of George and Liu (1981), particularly if
the maximum size of the network is between 1,000-2,000 nodes.
216
CHAPTER SIX
CALIBRATION OF WATER SUPPLY DISTRIBUTION SYSTEMS:
A NEW COMPUTER-BASED EXPLICIT METHOD
6.1. Introduction.
Due to the inherent complexities of the mathematical model for
describing the steady and unsteady state of a water distribution
network, computers have been used to ease the modelling process
since the early sixties, when these machines became available for
practical engineering applications.
To build up the mathematical model, a number of simplifying
assumptions are needed, in order to produce a mathematically
tractable problem. Additionally, most of the parameters required
by the mathematical model are subject to estimation errors.
Consequently, we have to recognise that the mathematical model as
a whole is just a rough approximation of the reality.
As a result, if we compare the predicted variables of the model
with the corresponding measured variables from the real network,
we shall find that agreement between both quantities is rarely
reached. If the model is used to represent the reality, something
has to be be done in order to improve the quality of the model
results with respect to the real world.
A calibration stage is then needed before the model can be used
to represent the reality for any practical purpose, the
217
calibration being responsible for guaranteeing that the results
predicted by the model and the corresponding variables in the
real network match within a reasonable degree of accuracy, for
all the possible range of operating conditions in which the model
is going to be applied.
The purpose of this chapter is to present a calibration
procedure for water distribution networks, which can be carried
out with the help of the computer, and used interactively by
experienced engineers dealing with the water distribution system.
Section 6.2 deals with the identification of most of the causes
of uncertainty in the modelling process, while Section 6.3
summarises the nature and accuracy of the measurements normally
available in a typical water distribution network. Section 6.4
contains the problem formulation. Section 6.5 presents a review
of the existing methods for calibration and Section 6.6 compares
the conveniences and shortcomings of a deterministic and a
probabilistic approach to the calibration problem. Section 6.7
introduces the rationale behind the proposed method, whereas
Section 6.8 presents the method itself. In Section 6.9 different
techniques applicable to the problem of estimating the unmeasured
piezometric heads in a water distribution network are presented.
Section 6.10 illustrates the application of the proposed method
to some test examples, while in section 6.11 the calibration
results are compared and discussed. Finally, some conclusions are

drawn in Section 6.12.
6.2. Uncertainties in water distribution modelling.
First of all, we shall constrain the scope of this chapter to
the steady-state flow exclusively. The unsteady state problem is
a completely different problem from the mathematical viewpoint,
and its modelling is usually carried out having different
objectives in mind, as compared with the aims of traditional
network analysis. This involves a first cause of possible error
in the model, since it is more adequate to understand the real
behaviour of the network as in a continuous unsteady state; this
is due to the fact that demands are always changing, pumps are
often switched on and off, valves are moved, etc., generating a
wide range of transient phenomena, some of them with a slow
response and some with a fast response as well.
In modelling the steady-state flow, some of the uncertainties
and simplifications involved are concerned with
6.2.1, Mathematical model for the flow/head loss phenomena.
It is assumed that a mathematical non-linear relationship
approximates efficiently the flow-head loss phenomena in the
network elements (pipes, pumps, valves, etc.). The Hazen-Williams
and the Darcy-Weisbach formulae are widely accepted and both of
them can be modelled as:
= aij Qij n + a ii (1)
where hij and Qij are the head loss and flow between nodes
and "j", respectively; aij and aij are the characteristic

parameters and n is a real exponent. All of them are formula-
dependent.
219
The non-linear model represented by equation (1) is valid for
turbulent flow (i.e. when Reynolds number is greater than 4,000-
6,000). For laminar flow (i.e. Reynolds number less than 2,000),
linear models in the flows (or in the velocities) like those

proposed by Hagen in 1839 and Poiseuille in 1840 are recognised
as a more adequate representation of the physical reality. The
region with Reynolds numbers between 2,000 and 4,000 is often
considered as a "transition zone", and no explicit model is
available for describing the flow in that zone. Generally, either
the turbulent or laminar model is applied in that transition
zone, depending on the state (laminar or turbulent) prevailing
prior to reaching the transition zone.
In other words, the mathematical description of the physical
phenomenon itself has some degree of ambiguity, which is normally
considered as non relevant for practical purposes on the grounds
that, in civil engineering a p plications, most of the flows are
turbulent, which is not always true.
The coefficients aij and Bij in equation (1) are the
characteristic parameters of every branch of the network. In the
case of pipes Oi 3 =0 and aij = Lij/(0 Ci j 1.852 Dij4.87037kj for the
Hazen-Williams formula, with 0 being a units-dependent constant.
In the case of valves, Bij z 0 and aij has to be obtained from the
characteristic curve of the valve (supplied by the manufacturer
or determined in a laboratory). Similarly, in the case of pumps,
both aij and Bij are needed.
In a real network, the characteristic coefficients aij and Bij
are assumed to be known. In practice, a number of factors lead us
220
to the uncertainty in the determination of the characteristic
parameters; some of these factors are:
a) Pipe resistance:
The roughness of a pipe, represented by Cii in the case of the
Hazen Williams formula, is difficult to determine in practice.
Field tests are needed for the estimation of the roughness of an
old pipe [i.e. hydrant tests and others described by Walski

(1984)]; for a new pipe, the corresponding value can be
determined in a laboratory and is usually supplied by the

manufacturer.
Furthermore, the pipe resistance changes slowly with time,
then, in essence, we are dealing with a variable coefficient,
which is assumed to be constant for practical reasons.
The older the pipe, the greater the uncertainty associated with
its resistance.
b) Diameter:
Assuming that at least the nominal diameter and corresponding
materials of the pipes are known (which is not always the case in
the real world), different classes of pipes can be found,
especially when we are studying a system of old pipes; in this
case, it is usual to find different effective diameters for the
same nominal diameter. In addition, deposition of mineral salts
and the action of corrosion processes can reduce the effective
diameter of a pipe with a known nominal diameter or change the

shape of the net flow section.
221
c) Length:
This variable is usually determined from a map of the city
under study (scales 1:5000 or 1:10000 are usual), where the pipes
of the network have been drawn; as a result, the lengths are
quantities subject to errors of measurement.
d) Valve settings:
The valve characteristic parameters are also subject to errors
unless a periodical checking procedure is performed, which can
include laboratory tests to re-determine the characteristic
parameters. Furthermore, the degree of aperture of a valve in a
real network is not known precisely and it is quite common to
find some valves closed when a field inspection is performed.
Again, we have to look at the characteristic parameters as
uncertain parameters.
0 PUMP characteristic values:
Similarly to the case of valves, the real values of the
characteristic parameters of pumps are not precisely known,
unless laboratory tests are carried out. Also, it is known that,
due to the normal wear during the operation of the pumps or due
to abnormal cavitation processes, the values of the

characteristic parameters can change with time.
f) Minor losses:
It is an accepted practice in water supply modelling that minor
losses produced by the presence of special fittings like t-
joints, elbows, reductions, etc. are neglected, due to their
small values in comparison with the head loss produced by
friction in the pipes. This generally accepted practice can fail
in some special cases when short pipe lengths and important
222
fittings are present. For these cases, the introduction of an
equivalent extra length of pipe, which produces a head loss
approximately equal to that corresponding to the fittings, is

sometimes considered an acceptable solution. To cope with these
cases, an adjustment of the length of the pipes should be
allowed, supporting the treatment of the length as a variable

subject to errors of measurement, as noted previously in c).
Consequently, it is no longer possible to consider the

characteristic parameters aij and B ij as deterministic, and a
probabilistic model seems to be more adequate to explain their

uncertainties and errors.
6.2.2. Water consumption.
The demands of water from the customers (residential houses,
commerce, industries, etc.) are assumed to be concentrated in the
nodes of the network model, instead of along the pipes, as
happens in the real system. In addition, the spatial distribution
of the consumptions is something which is difficult to establish

precisely. The normal p rocedure for that purpose involves a
sectorization of the area served, which leads to an approximation
of the spatial pattern of the demand. This process involves a
great deal of judgement and subjectivity, producing estimates of
the demand whose accuracy is not normally known; See Crabbe et
al. (1982) and Brandon (1984; chapter 8) for details.
Finally, the consumptions are subject to time variation.
Within a day the consumption rate changes hour by hour, following
closely the different activities of the consumers. Within a week,
223
there is also a daily variation, due to the fact that the
consumption pattern changes from week days to the week end. In
the same line, monthly variations, induced by weather changes,
and yearly variations, produced by population changes, are also

acknowledged.
The presence of leakages and unaccounted-for water
consumptions, makes the picture of water consumption even more
blurred.
In summary, both the spatial and temporal distributions of
water consumption are subject to uncertainty, and the way in
which this consumption is modelled (concentrated in the nodes) is
a simplification required to make the problem tractable from the
mathematical point of view. This means that, in practice, nodal
consumptions are a fiction, created by network analysts to be
able to tackle the problem; clearly, nodal demand measurements
are not possible to obtain and only indirect and global
measurements of consumptions are sometimes available. The
practice of measuring individual consumptions on a house by house
basis is common in some countries, but that information usually
incorporates a great deal of uncertainty. Thus, water
consumption is a parameter subject to considerable variability.
6.2_3. Effect of network reduction.
In the early stages of network analysis, when the computations
were carried out manually, the real network was "skeletonized" in
order to reduce the amount of pipes and nodes in the mathematical
model, thus reducing the effort needed to solve the problem. The
224
reduction was obtained via the simple elimination from the model
of those small diameter pipes which were supposed to contribute
in a minor way to the flow-carrying function of the real network.
In essence, network skeletonization contributes some extra
uncertainty into the network model and it will certainly lead to
discrepancies between the model and the real network. As a
result, skeletonization should be used carefully and a compromise
should be struck when deciding which pipes are to be eliminated
from the model, if any pipes have to be eliminated at all.
Eventually, the convenience of skeletonization as a regular
practice should be revised.
6.2.4. Influence of network Pressure.
The nodal demands are generally assumed to be independent of
the pressures in the network, which is clearly not true in the
real world. In fact, the operators usually reduce the pressure in
the network in order to lower peak demands and also to reduce the
amount of leakages, especially during the night.
This factor is not considered explicitly in normal network
modelling practice.
6.2.5. Other factors.
Other problems Can contribute to introducing errors and
uncertainty in a network model; some of them may appear at first
sight as improbable but they do occur. See Crabbe et al. (1982)
and Brandon (1984) for details. Some of these problems are:
225
* Misinterpretation of recorded data: wrong dimensions and units
are reported as a common cause of error.

* Actual connections between pipes are different from those
assumed in the model.
* Valves which may be operating in a completely different state,
e.g. open in the real network, but closed in the model, or vice
versa.
* Pipes by-passing pressure regulating valves and meters.
* Pipe obstructions, due either to the presence of debris or to
untidy maintenance work, may change dramatically the
resistance characteristics of the pipes.
All these factors should be checked before a formal calibration
procedure is started; moreover, if some strange results are being'
obtained with the model, it might be necessary to go back to
these possible error sources.
Clarke et al. (1981) give some examples on how meaningful (or
meaningless) the results of some simple pipe hydraulics
calculations can be when dealing with uncertain data. Walski
(1987) stresses the sensibility of network calculations to
measurement accuracy and also refers to the influence of the fact
that steady state conditions are difficult to obtain in a real
network, especially when pumps and pressure regulating devices
are present. All these warnings are applicable to any calibration
procedure.
6 3. Measurements available in a real network and their
accuracy.
Accepting the previous assumptions, in standard network
analysis applications we first proceed to "estimate" the physical
characteristics of the network (roughnesses, diameters, lengths
of pi p es, characteristic parameters of pumps and valves, etc.)
and the global demand and its spatial distribution for a

particular time horizon. With these data, the raw model, as
defined in Chapter 1, is solved and its results in terms of heads
per node and flows per branch are obtained.
At this point of the modelling process, a new question arises:
How close does our model represent
the behaviour of the real network ?
To answer this question we need additional information from the
real network; the required pieces of information are obtained via
the field measurement of different parameters in the real
network:
* pressures at selected nodes of the network.
* water levels in the reservoirs.
* flows at selected points of the system.
* demands for certain areas and users in the network.
* roughnesses of key pipes of the system (field measurements).
* characteristic coefficients of pumps and valves (laboratory or
field measurements) .
* effective diameters.
* ground level elevations at certain nodes.
* meter calibration results.
227
To simplify the terminology, pressure and water level
measurements will - from now on - be referred to as "piezometric
head measurements", though pressure and level measurements need
to be added to the corresponding ground level to obtain the
piezometric head. Nodal piezometric heads and pipe flow
measurements are generally the basic measurements in the network.
There is a wide range of measurement devices for pressures,
flows and levels; for each device, different accuracies can be
available and, obviously, the higher the measurement quality, the
greater the reliability of the conclusions that can be obtained
with these data.
Although a 1% accuracy seems to be a normal requirement for
most of the flow and pressure measurement devices, we have to
accept the fact that poorer measurements can be available and
that it is a matter of statistical interpretation as to the
quality of the results that can be obtained with such kinds of
measurements. Walski (1984, 1988) presents a detailed review of
the different techniques and equipment for water distribution
network measurements.
Water distribution networks, unlike other public utilities, are
characterised by their lack of adequate measurement systems, both
from the quality and the quantity viewpoints. Normally, we should
not expect to have more than 10% of the nodes of the network with
pressure or level measurements. The amount of flow measurements
could be even less, but generally we do have flow measurements at
key points like reservoirs (inlet and outlet pipes), pumps,
228
important consumers (or sectors of the network), etc.
Although nowadays it is possible to find a variety of
measurements devices able to record or transmit the measurements
to a central command post, in many of the existing networks the
presence of a human being is still required to record the
measurement. A mix of telemetry with man-made measurement systems
is perhaps the most common situation. The ability to handle data
with different qualities, is something which seems to be
indispensable if realistic comparisons between the results of
network models with reality are to be carried out.
Due to meter or communication line malfunction, the
measurements themselves may be subject to error and, quite
possibly, some of them may become unavailable during a certain
time interval. A robust calibration procedure should be able to
cope with such problems and eventually reject or correct those
measurements, and avoid the spreading of such errors into the
modelling of the rest of the network.
6.4. Problem formulation.
Having information on the real behaviour of the network at
selected points, a new question appears, since, in general,
measured and modelled variables do not coincide:
What can be done to improve the agreement
between the model and the real network ?
The answer to the latter question is basically the problem
referred to as "calibration" of a network. A network is said to
229
be calibrated when the results obtained with the model match
field measurements within some pre-specified reasonable
tolerance, for the different operating conditions.
In the previous definition, some key words are relevant.
Firstly, what is meant to be a "reasonable tolerance" and,

secondly, what do we mean by "different operating conditions" ?.
In the first case, we may have different criteria for
establishing a reasonable tolerance, for example we could say
that we want individual measurements to be within a fixed
percentage of maximum disagreement (say, 5%; thus the maximum
flow and head difference should be less than 5%) or, we could ask
for a certain proportion of points (say 80% of the measured
points) to be within a certain percentage of disagreement (say
5%), whereas the rest (say 20%) should be within a bigger error
margin (say 10%). Obviously, different criteria can lead to
completely different problems and, of course, could require
different data quality and quantity. There is no definite answer
to this problem, but the tolerance criterion to be used should be
in accordance with the objectives of the calibration. To make
things worse, the establishment of a tolerance criterion is
problem-dependent; thus, on some occasions, when the network is
characterised by a flat piezometric plane, for example, a more
stringent tolerance criterion may become crucial for a
meaningful analysis
It is essential that the network should be calibrated for all
possible operating conditions. Hence, high and low consumption
rates should be considered, as well as all the possible
230
combinations of operating modes of devices like pumps, regulating
valves, etc. Failure to do so may imply that a network appears to
be calibrated when it is not. Of course, a model which does not
match the behaviour of the real system is of little use for
practical purposes and the importance of the calibration
procedure becomes a key factor in the whole modelling process.
This applies to all kinds of models.
The calibration problem, then, can be formulated as that of
findin g a corrective procedure for the different parameters

involved in the network model, in order to reduce (or eventually
eliminate) the discrepancies between the model and the real
network, for all the possible operating conditions, using a
limited amount of measurements, each one possibly with a
different accuracy. This formulation is adequate for an off-line
calibration procedure, when batches of data measurements are
available for different operating scenarios. By a limited amount
of measurements we mean the amount of measurements we normally
have available in water distribution systems, where we rarely
have as many measurements as parameters needing to be calibrated
and, certainly, we are never in a situation where we may have
redundancy of measurements, except in some very special cases,
like long and important pipelines.
As far as the present work is concerned, we shall restrict
ourselves to finding a solution to the calibration problem in the
previous context.
In general, when a model is to be used on a real-time basis, a
more demanding formulation is needed [see O'Connell (1977;
231
chapter 1)]. These more severe prerequisites can be expressed as:
* The model parameters should be updated whenever new data become
available (adaptive model).
* The calibration procedure should be able to estimate the error
associated with the corrected parameters.
* The calibration procedure should be robust, and able to operate
with interrupted data sets.

* Low computational cost is desirable, in order to facilitate
implementation on microcomputers.
6.5. Review of existing methods for water distribution network
calibration.
We shall concentrate on full scale models of a water
distribution network, these being the most widely used models for
analysis and design purposes; in these models some reduction or
"skeletonization" is permissible, but the model still replicates
the real network both in a geometrical and a hydraulic sense.
Reduced scale models, where only 20-30 % of the original
elements are included in the model are not explicitly considered
here, though some of the existing techniques for full scale
models are also applicable to reduced models. In reduced models,
the network to be studied is no longer equivalent to the real one
in the geometrical sense; this means that both the shape and the
size of the model might be completely different to the real
network, but the main hydraulic variables ( flows at important
pipes, heads at key nodes, levels at reservoirs, flows and heads
232
at pumping stations, etc.) are replicated by the model. Reduced
models are used mainly for control purposes, especially in the
cases where only the behaviour of the main network elements
(reservoirs, pumping stations, etc.) is relevant; these models
are necessary for computational reasons, for instance when very
time consuming optimisation techniques (dynamic programming, non-
linear programming, etc.) are being used. See Gilman et al.
(1973), de Moyer (1973), de Moyer and Horwitz (1975) or Shimauchi
et al. (1985) for details on reduced scale models
In the traditional approach to water distribution network
calibration, of all the possible causes of mismatch between
modelled and measured variables, two of them usually play a major
role: the estimated nodal demands and the estimated resistance
characteristics (pipes, valves, pumps, etc.). This has led to two
different schools of thought, looking at the calibration problem
either from a demand or a resistance point of view. In both
cases, the way to deal with the calibration problem in the early
stages of network analysis and modelling has been a trial-and-
error exercise, where the network model is used as many times as
possible, with different demands and/or pipe resistances until,
hopefully, agreement between model results and measurements is
achieved.
One of the earliest attempts to formalise and solve the
calibration problem is due to Shamir and Howard (1968), where the
network analysis problem (nodal formulation) has been reshaped in
order to include nodal consumptions and element resistances as

unknowns, as well as piezometric heads. In this context, a set of
233
non-linear equations is produced, now containing heads,
consumptions and resistances as unknowns and, in principle, the

non-linear system can be solved, provided that there are as many
equations as unknowns. In practice, this is not always true, and
the selection of the unknowns (specifically when dealing with
consumptions and resistances) has to be done following a set of
rules in order to get a "solvable" problem. Shamir and Howard
(1968) enumerated some of these rules, which have subsequently
been extended by Shamir (1973). However, the authors did not
guarantee that even following these rules, a solution could be
found.
An alternative approach has been the development of some
calibration procedures based on a sensitivity analysis, where the
model parameters (either demands or resistances) are studied
according to their impact in the mismatch between modelled and
measured variables. A correction to the most sensible parameter
is computed in order to force agreement between model and real
world. We shall review some of these algorithms.
Donachie (1974), following a nodal approach for the network
analysis solver, presented a post-analysis calibration procedure,
based on determining the impact of changes in the pipe
resistances on the nodal piezometric heads. The values of the
sensitivities of the nodal piezometric heads with respect to the
changes in the pipe resistances are determined using the Jacobian
matrix obtained in the original uncalibrated system, and these
sensitivities are used to reduce the discrepancies between
modelled and observed pressures. This approach leads to a set of
234
linear equations in the resistance corrections, which is
overdetermined because there are as many equations as
measurements (say M equations) and as many unknowns as pipes (say
NP, with NP > M ). The overdetermination is solved via a least
squares fit, and the resistance to change is that which needing
the smallest correction produces the largest effect in a sum of
squares index. Choosing the most sensitive resistance allowsusto
continue the use of the original Jacobian (obtained during the
network analysis solution) to carry out as many corrections as
necessary, until agreement is obtained. This simplifies the
procedure and reduces the computational cost. The procedure can
be used under several demand conditions, thus choosing the
resistance changes having in mind all the possible consumption
scenarios.
Donachie (1974) also shows how the same procedure can be used
to evaluate the sensitivity of the nodal piezometric heads to
changes in the nodal consumptions.
The main comments that can be made in connection with

Donachie's approach are:
* The method does not use the information on flow measurements
(when available); calibration is carried out using the

pressure measurements only.
* The selection of the pipe resistances to be adjusted is
arbitrary, since the discrepancies between model and reality
need not necessarily be produced by the most sensitive pipes.
In practice, this means that the algorithm produces results
that look like a good calibration, when this may not be the
235
case.
* The method considers that the pressure measurements are exact
and does not allow for an explicit consideration of the quality

(or error) of the measurements.
* The proposed method does not reproduce the pressure
measurements, though it gets smoothed values which are close
enough for practical purposes.
* The method does not give an estimate of the errors associated
with the corrected values of the roughnesses.
Rahal, Sterling and Coulbeck (1980) followed a sensitivity
analysis approach very similar to that of Donachie (1974), to
develop a "static parameter tuning" (SPT) algorithm, since it
referred to a time invariant system. This algorithm allows the
characteristic parameters of pumps to be included in the
calibration.
Rahal and Sterling (1981) extended the previous algorithm for
dealing with the dynamics of an extended period simulation, where
now the time varying nodal consumptions and reservoir levels are
adjusted to match model results with the real network. The new
algorithm is referred to as "dynamic parameter tuning" (DPT), to
mark the difference with their previous "static parameter tuning"
(SPT) method. The DPT algorithm operates in two stages:
I) Calibration of reservoir levels and reservoir inflows.
II) Calibration of nodal demands.
It is assumed that the static analysis model is exact, i.e.
that the SPT algorithm has been previously used to calibrate the
resistance p arameters. Thus, the remaining uncertainties in the
236
dynamic case are the reservoir levels and the nodal consumptions.
The calibration of reservoir levels and inflows (stage I) is

needed due to errors associated with the integration of the set
of differential equations describing the dynamics of the system
which is done through a predictor-corrector scheme, but also
because the SPT algorithm does not reproduce exactly the level
measurements. Stage I is needed only once.
The second stage is based on the determination of the
sensitivities of the nodal pressures and reservoir inflows to
variations in the nodal consumptions; a correction for the nodal
consumption is made by changing that nodal demand which, with the
smallest change, produces the largest variation in the sum of
squares of the errors. This needs to be carried out for one nodal
consumption at a time, and as many times as is needed to keep the
maximum discrepancy between the observed and modelled pressures
under a pre-specified limit.
Most of the comments made for Donachie's algorithm are still
valid for the SPT and DPT algorithms. In the case of the DPT, we
can say that, in addition
* The selection of which nodal demand to adjust is arbitrary,
since the disagreement does not need to be produced by the most
sensitive demand.
* The method is not able to reproduce exactly the level
measurements and inflows in the reservoirs.
Coulbeck (1984) recognised that both the methods proposed by
Donachie (1974) and Rahal et al (1980) " do not make full use of
237
the available information and can lead to results which disagree
with practical experience". He presented a generalisation of the
sensitivity-based algorithms, now including flow measurements and
allowing the adjustment of most of the parameters (pipe
resistance parameters, pumps control parameters, consumer demands
and reservoir levels) simultaneously, rather than one at a time
as previously. Coulbeck's algorithm still assumes that the
measurements are exact. This leads to an overdetermined problem,
since normally there are more parameters to adjust than
measurements; the problem is solved via an explicit minimization
of a weighted sum of squares performance index. The computer
program implemented is interactive and the operator can examine
the variation in the calibrated parameters, weighting factors,
etc. in response to changes in the adjusting parameters.
Although Coulbeck's model removed some of the shortcomings of
previous algorithms, the method is still deterministic, because
no explicit handling of the measurement errors is allowed.
Gofman and Rodeh (1981), following a loop approach for the
network analysis problem and looking for efficient microcomputer
solutions, proposed a deterministic solution for the calibration
problem. They introduced a fictitious device, which they called a
"head generator", whose function is precisely to introduce extra
head-losses (or gains) in order to force agreement between
measured and predicted pressures. To obtain the solution, they
have to add a pseudo-loop for each new pressure (head)
measurement, the specification of which is arbitrary, as the
238
location of the "head generators"; additionally, they found that
the problem is solvable only under certain conditions, otherwise
the new problem either has no solution or has many solutions.
Almost the same arguments as in the previous methods are
applicable here, i.e. lack of capability for handling the error
of the pressure measurements, no use of flow measurements, etc.
Ormsbee and Wood (1986) proposed an explicit deterministic

algorithm for calibration, which is based in the well known
linear theory method developed by Wood and his collaborators at
the University of Kentucky. As each new pressure (head) or flow
becomes available, an extra energy equation can be added to the
original system, allowing for the addition of a new unknown. For
calibration purposes, the algorithm proposed by Ormsbee and Wood
(1986) considers that either a head loss adjustment
(multiplicative) factor for the friction losses or an additive
head loss factor for the minor losses are suitable variables. The
user has to decide which kind of adjustment factor (or a
combination of them) is to be determined. A general head loss
adjustment factor is also possible (i.e. when a single
measurement is available), but a combination of head loss
adjustments and head loss factor (for minor losses) is preferred
by the authors. This approach generates an extended non linear
system of equations, one for each extra energy equation, which is
solved following a similar approach as in the linear theory
method. As a result, both the selection of which head loss is
more relevant (friction or minor losses) and also the specific
location of the pipe where that loss is to be introduced are
239
arbitrary and many solutions are possible, or none at all, as
shown by Gofman and Rodeh (1981).
Walski (1987), in a discussion of Ormsbee and Wood's (1986)
paper emphasises that errors in the pressure measurement data can
lead to "unrealistic solutions", thus the accuracy of the
measurements becomes a critical issue, as far as the credibility
of the calibration is concerned. Also, the fact that steady state
conditions are not common in a water distribution network (see
section 6.2 of this chapter) has an impact on the accuracy of the
measurements and in the discrepancies between modelled and
measured heads. Walski's comments are applicable to All the
calibration methods already reviewed.
Walski (1983, 1984, 1986), within a deterministic approach,
proposed a completely different strategy for calibration. He
recognised that both the pipe resistance parameters and the nodal
consumptions can be the origin of the mismatch between observed
and measured heads and flows. His approach is pot based on re-
formulating the network analysis equations. Using data (existing
or especially obtained for this purpose) from fire flow testing,
where a hydrant in the network is open and the pressures in the
network and the level in a reference reservoir (or a PRV) are
observed, Walski replaces the whole area of influence of the
test, which has to be defined by the user, by an equivalent pipe
and looks for multiplicative correction factors for the Hazen-
Williams roughness coefficient C (say, B) and for the nodal
consumptions (say, A), within the area of influence of the fire-
flow test. With the information provided by the recorded
240
pressures in the hydrant before and during the test, plus the
levels in the upstream reservoir and the flows measured at the
h ydrant, Walski's method leads to a system of 2 equations in the

unknowns A and B, i.e. the corrective factor for the nodal
demands and roughnesses, respectively; the system can be solved
via an elimination process and an explicit formula for A and B
can obtained. On obtaining A and B, all the previously estimated
nodal demands within the area of influence are multiplied by A,
while B multiplies all the Hazen-Williams roughnesses. According
to Walski, the method then automatically determines whether the
nodal consumptions or the roughnesses are the main factors
contributing to the disagreement between observed and measured
variables.
Walski's method operates over sectors of the system (which are
defined based on the area of influence of the fire-flow tests),
one at a time, until the whole area fed by the network is
covered. The method actually is an extension of an existing
practical procedure for determining the roughness of a single
pipe, where now the single pipe under analysis is replaced by an
equivalent (fictitious) pipe representing a sector of the
network. The method is arbitrary, in the sense that the area of
influence of the test is difficult to define precisely, this has
an impact on the selection of the nodes affected by the test and
their water demands. The method also downgrades initially well
estimated demands, due to its averaging features; in fact, the
original total demands may be increased or decreased by the
algorithm. To apply this method, we need a network where it is
easy to find a reference reservoir for a sector of the network,
241
which renders the method better suited for pie-shaped networks
with a central reservoir feeding a nearly circular area. Another
shortcoming is the difficulty of implementing the method on a
computer, in order to apply it systematically. This approach is
eminently practical and relies on hydrant flow tests, which
forces the network to act locally under severe flow increases.
Recently, Walski (1988a), based on a number of practical and real
situations, expressed the view that "it will never be possible to
develop a simple analytical procedure or optimisation technique
that will calibrate a model"; he suggested that possibly expert
systems can become helpful in this regard.
Bhave (1988), following a similar approach to Walski, but now
guaranteeing that the total inflow . remains constant and equal to
the measured inflow (which Walski's method failed to maintain),
proposed a calibration technique which adjusts pipe resistance
characteristics and nodal consumptions simultaneously. Let NT be
equal to the number of nodal pressure measurements plus reservoir
level measurements; then, on taking a unique reference node, we
can link this reference node with any of the pressure measured
nodes and level measured reservoirs, thus creating NZ=NT-1 paths,
each one starting at the reference node and ending at another
measured node. Each measured node and associated path defines NZ
zones and, for each zone "i", a demand adjustment Sq i can be
made, if considered necessary by the user, such that E Sqi 0
(summation over all the NZ zones is zero in order to maintain the
total inflow constant) or a zonal resistance adjustment Bi can be
introduced. Also, in each path, the predicted summation of head
242
losses and the measured total head loss between the reference and
measured node must coincide. This leads to the establishment of
NZ equations in NZ unknowns Bi and/or Sqi, which are to be chosen
by the user; the resulting system of equations is non linear but
Bhave uses a simple linearization scheme based on the fact that
Bi z 1. The corrective factor Bi operates over all the pipes in a
zone and the corrections in the demands per zone are distributed
over the nodes, maintaining the proportions originally assumed
for the nodal consumptions. An iterative procedure is followed
until Bi -> 1 and Sqi -> 0 for all the zones.
Bhave's proposed method has most of the limitations and
arbitrariness of Walski's method and according to its author "the
convergence is rather slow", perhaps due to the linearization
introduced in order to get a set of linear equations in the
unknowns. The method is deterministic and it handles neither flow
measurements nor the error inherent in the pressure measurements.
Lansey (1988) formulated the calibration problem within the
framework of non linear programming, having as an objective
function the sum of squared differences between measured and
modelled heads (he did not consider flow measurements), though a
slightly different objective function based on the ratios between
measured and modelled heads is also discussed. The restrictions
are basically the set of non linear equations relating head loss
and flows per branch, the mass balance per node equations and
additional constraints on the variation of nodal demands,
roughnesses, PRV settings and valve characteristic coefficients.
In order to make the problem tractable from the computational
243
viewpoint, Lansey does not include the set of non linear head
loss/flow relationships, but replaces it in an implicit way,
using the network solver in tandem with the optimisation model.

Lansey provides neither examples nor performance information on
this approach. The formulation presented by Lansey is able to
cope with multiple loads, though it is still deterministic and
uses only pressure measurements. Shimauchi et al. (1985), dealing
with a reduced scale model of the network, followed a similar
approach, this time including a weighting factor in the objective
function to consider measurement errors and including explicitly
the set of non linear head loss/flow relationships in the
constraint set.
Ormsbee and Chase (1988), following a similar approach to
Lansey, but now with a linear objective function minimizing the
absolute value of the ratio between measured minus modelled
variables and measured variables. Although the objective function
is explicitly linear, in fact it is implicitly non linear in the
unknowns (heads, flows, resistance parameters). The minimization
problem is constrained by implicit system constraints (mass and
energy equations) and also by minimum and maximum bounds for the
pressures, flows, roughnesses, nodal demands and slope of the
piezometric head line. As in Lansey's formulation, the solution
of the non linear optimisation problem is quite expensive from
the computational viewpoint and the complex method is recommended
by the authors as an efficient solver, the complex method being
to non linear programming as the simplex method is to linear

p rogramming. The authors applied this method to two cases in the
U.S.A.; one of them corresponds to the city of Arlington (316
244
I
pipes and 197 nodes) and the other corresponds to the Federally
Owned Water Main (FOWM) in Washington D.C. (64 pipes and 55
nodes); both systems link together. The authors did not present
any performance- related information.
6 6 Deterministic versus stochastic aPProach to the solution of
the calibration problem.
All the methods reviewed in the previous section for full scale
water distribution network models are deterministic, in the sense
that most of them are not able to handle explicitly the errors
associated with the measurements and, in addition, they are not
capable of producing estimates of the errors associated with the
estimated parameters.
Nowadays it is recognised that data with large errors can lead
to unrealistic results in the calibration process. On the other
hand, due to the shortage of measurements in water supply
networks, which is typical of these systems, we have to utilise

all the measurement information available, even the low quality
data. In addition, the availability of measurements obtained from
devices of different quality is not unusual in real systems. All
this leads to the need for a calibration method which can
explicitly handle the errors associated with the measurements.
The algorithm then must be able to "weight" the data according to
their errors and finally, produce estimates not only of the
calibrated variables, but also of their associated errors, since
a corrected roughness of C=100±10 is very different from another
of C=100±100. This is consistent with our formulation of the
245
problem, as presented in Section 6.4.
Only the methods based on the non linear minimization of the
discrepancies between observed and modelled variables are close
to this approach [Lansey (1988), Ormsbee and Chase (1988)],
although they do not consider error measurements explicitly; this
is something that can be added relatively easily into the
objective function. Even with that modification these methods are
not able to produce estimates of output errors, with the
additional burden that their computational cost is very high
indeed. The main problem with the non linear optimisation
approach is perhaps due to the fact that it is in essence a
deterministic approach, i.e. it does not truly recognise the
probabilistic nature of the problem. In fact probabilistic
(stochastic) estimation methods for model calibration differ from
the optimisation approach in that they add some additional
requirements to the problem, normally in the sense that not only
must the residuals be minimized but also the variances associated
with the estimates have to be explicitly minimized.
We shall adopt a stochastic approach to the calibration
problem, based on some of the parameter estimation techniques
already available from the statistical and control fields, which
have been successfully used in other problems of water resources
engineering. We believe that only in this context is it possible
to handle efficiently the calibration problem, as formulated in
Section 6.4; this can, eventually, lead to algorithms which can
be implemented in real time at the microcomputer level.
Additional ly , we shall implement some deterministic calibration
246
techniques as well and we will compare them in order to evaluate
their respective advantages and disadvantages:
6.7. The rationale of the ProPosed calibration method.
Although it is clear that the calibration problem has to be
addressed within a dynamic framework, i.e. considering the whole
range of possible operational states for the system (e.g. low
demand and high demand periods), we shall follow the usual
procedure of simulating the extended period operation of the
system via the integration of successive static analyses. The
dynamics of the extended period simulation are described via a
set of differential equations which cater for the storage
variation within each reservoir; the system has as many
differential equations as reservoirs. In following this approach,
it is reasonable to assume that the calibration process also has
to be carried out in two mutually interactive stages:
Static calibration: where it is assumed that the total demand
and its spatial distribution is known at a specific time and
only the pipe roughness can be considered as responsible for
the mismatch between the observed and modelled variables (nodal
heads and pipe flows). In this case the reservoir levels and
valve and pump status are considered fixed and known.
* Dynamic calibration: now the time variation of demands and
reservoir levels are assumed to be responsible for the
disagreements between observed and modelled variables.
We shall concentrate mainly on the static calibration problem.
247
First, let us assume that we know exactly the nodal demands and
their spatial distribution at a certain time instant; this
assumption is based on the fact that the spatial variation of the
demand has been thoroughly studied and estimated, using for
example the methodology recommended by the Water Research Centre
[see Brandon (1984)] for nodal demand allocation and/or using the
information from meter reading at household level [see Harrison
(1988)]. Additionall y , we assume that the total demand is

measured at any time in the system (reservoir outflows, pump
discharges, etc.), so that it can be assumed that the nodal
demands allocated are a correct estimate of the spatial variation
and the total demand (summation of all the nodal consumptions)
must match the total measured demand. A proportionality factor
can be applied in order to force agreement between these total
demands (measured and summation of the estimated ones).
Hence, we are faced with a calibration problem where the pipe
roughnesses are meant to be the main factor contributing to the
model mismatch with reality. In fact, this approach implicitly
considers effective diameter and length variations, as well as
minor losses, as part of the roughness to be estimated. The
roughness is then the only "instrument" to be used to calibrate
the model.
The rationale of the proposed static calibration method can be
explained as follows: let us consider a "network" consisting of a
single pipe joining node "i" and "j" then, in order to estimate
the true value of the roughness at the pipe, we need to measure
the difference between piezometric heads at each node and also
248
the flow in the Pipe. With these measurements, say Ht - HI and
Q i I , and using the known head loss/flow relationship (1), we can
compute the corresponding estimate of the true resistance
parameter (a*) for each pipe as:
aij* = ( Hi* - / (Qij*) n(2)
Ideally, having measurements for all the head losses and flows
per pipe in a larger network, the application of equation (2)
would lead us to compute all the resistance parameters (i.e. the
vector **). Because in practice we do not have measurements of
all the nodal piezometric heads and pipe flows, we have to find
the best estimates of them and use equation (2) to produce our
best estimates of the resistance parameters. The calibration
problem is now reduced to the estimation of nodal piezometric
heads and branch flows, based on a limited amount of available
measurements.
Note that if, in our single pipe system with nodes "i" and "j",
Qij* is not the true value, we would be able to find a different

a-- *
1J which could make the network look like it has been
calibrated when it has not. If we analyse a network consisting of
series-connected pipes, where the initial and final nodes of the
series have measured piezometric heads, even when knowing the
true flows per pipe, the problem of finding the resistance
parameters per pipe has infinite solutions; to break up this
indeterminacy we need to incorporate some additional information.
Once the a- 1J
-* have been estimated, using equation (2), the
computation of the estimated pipe roughness vector (C*) is
straightforward in the case of the Hazen-Williams formula, since
249
e is explicitly related to the estimated characteristic
resistance parameter aij* via the followin g relationship:
Cij* = { Lij / [ 33686.36 Di j 4.87 ] (1/1.852). (3)
When using the Darcy-Weisbach formula a similar (direct)

procedure is needed to compute the friction factor. Having the
friction factor, the effective roughness can also be determined
directly from the Colebrook-White formula.
In practice, some control on the maximum variation allowable
for Cij* is needed, in order to prevent Cij* reaching unrealistic
values, especially in pipes where poor head estimation (of
initial and final nodes) and poor flow estimation coincide. We
allow C-1 J-* to move within an "allowable band" around the initial
roughness estimate, i.e.:
(1-x)C1j(°) S Cij* S (1+x)Cij(°) (4)
where:
C-
1J -(°) : initial roughness estimate for pipe "i-j".
X : maximum variation factor for C.
In the results presented here, the maximum variation factor "x"
has been set to 0.10, thus allowing Cij* to move within 90 and
110 % of the initial roughness estimates.
In theory, when having head and flow estimates near the true
values, this variation factor "x" should be left as wide as
possible (i.e. Cij* is allowed to vary without constraints). In
principle, with "good" head and flow estimates, the calibration
algorithm should produce "good" roughness estimates, irrespective
of the initial roughnesses (then there is a need for "x" to be as
250
larg e as possible).
We shall postpone further discussions on the subject of this
variation factor "x" to subsequent sections of this chapter.
The calibration algorithm can be viewed in a slightly
different way: let us denote the true parameters of the network

by at , Elt and ^t
m (i.e. resistance parameter, nodal heads and
flows per pipe); then, because the head loss/flow relationship
holds for every pipe joining nodes "i" and "j", we can write:
a. .t = h . .t /(Q . .tln
ij lj lj • (5)
where hii t is the true head loss, i.e.: hii t = Hi t - Hit.
Similarly, for the estimated variables, a*, Ii* and Q*, the
following is true (equation 2):

a ij * = hii* /(Qij * ) n(6)
where hii* is the estimated head loss, i.e.: hii* = Hi* - Hi*
Hence, dividing equations (5) and (6):
a.1Jj-* h. .* /(Q.lj* * ) n
---- =
a. .t hiit /(Qii t ) n
2.3
which can be re-arranged as:
* h li
- -*/h--t
1J
(7)
a..ttal, . 4C /on. .tln
13 L'-'13 ''"1J J
[
which can be interpreted in the sense that:
a* ___> at
if h* ___> ht - (5)
and if Q* ___> iit
251
It is not difficult to see that equation (7) also holds for the
case when:
E[h* ] = K ht (9)
and
E[Q* ] = K Qt(10)
where K is a constant.
The latter case implies that the estimated roughness parameters
can be equal to the true parameters when both heads and flow
estimates have exactly the same bias. This is only of theoretical
interest, since in practice the probability of having both
estimators with exactly the same bias is zero.
Clearly, (7) and (8) can be thought of as being the formal
expression of the rationale of our proposed calibration
algorithm.
6.8. An iterative approach for the solution of the static
calibration Problem.
The problem now is twofold: firstly we need to estimate the
piezometric heads within the network, based on a limited amount
of head (pressure and levels) measurements, which we assume are
sufficient to represent the structure of the piezometric plane
"floating" above the real network. The second problem will be
the estimation of the pipe flows.
The problem of estimating the piezometric heads in the
unmeasured nodes can be approached in various ways. We shall
examine the feasibility of using a geostatistical estimation
252
technique known as "Kriging", which is able to explicitly handle
noisy data and to produce an estimate of the error of its own
estimates. We also examine the use of a data approximation
technique known as bi-cubic splines, which is able to cope with
noisy data as well and, finally, a simple deterministic one-
dimensional interpolation technique will be described, which uses
the information of the initially guessed pipe resistances to fit
a broken line to the nodal head measurements.
The problem of estimating the flows is more complex, since the
flow distribution is strongly dependent on the network structure.
We can assume that if the flow distribution provided by the raw

model is close to the true flow distribution, we can use it as
an "estimate" of the true flows in the unmeasured pipes. This
assumption holds if the initial roughness estimates are good in
relative terms (i.e. relative to each other, rather than in terms
of their absolute values), since in that case the flow
distribution corresponding to the raw model will be close to the
true flow distribution. Flow measurements replace the modelled
flows, when available. We have found that this solution leads to
pipe flow estimates in a straightforward manner and, because the
assumption that the modelled flows are close to the true flows is
not exact, an iterative procedure leads to the final flow
estimates.
The basic structure of the proposed algorithm for static
calibration of the water distribution model can be represented in
the schematic shown in Fig. 6.1. The iterative algorithm
converges to the true values of the roughness parameters,
253
Start
i=1
Estimate initial value for C=C(0)

and for the max. variation "x"
\Read Heads and Flow measurements
Estimate unmeasured Heads: le

using any estimation technique, e.g.:Kriging,
one-dimensional interpolation, splines, etc.
Run gradient analysis program and compute

modelled variables: H and Q
Estimate unmeasured flows: Q*

replacing the modelled flows Q by i=i+1
their measurements (if available) 4
otherwise Q* =
"C" correction stage, compute:

ai j * = ( H i * -
-*={1.,-1J1(33686 .36 D-1 JA-0 7'"' 1J 1(1/1.852)
'
If Cij* < (1-x)Cij(°) then Cij = (1-x)Cij(0)

If C ij * > (1+x)C1j (0) then Cij* = (1+x)Cii(o)
assign:
C--
1J 1J
no i.e.:
c(i) <-- c*
yes
( Stop )
Fig. 6.1. Basic iterative static calibration procedure, computing

corrected Hazen-Williams roughness vector C.
254
provided that we input optimal estimates of the nodal piezometric
heads and pipe flows, as indicated in (8). Additionall y , because
the estimation processes (of heads and flows) could be carried
out separately from the calibration itself, it would be possible
to introduce further improvements in the estimation of heads and
flows into the calibration procedure in a straightforward manner.
This means that this static calibration scheme can be used as a
framework for improved head and flow estimation algorithms in the
future.
6.9. Estimating the piezometric heads.
There are two main approaches to estimating the piezometric
heads in the unmeasured nodes of the network, namely we can
either follow a probabilistic or a deterministic approach.
In a probabilistic (stochastic) approach both the measurement
errors and estimation errors are explicitly handled. In this
context, we have attempted to solve the problem with a standard
discrete linear Kalman filter, but because of lack of measurement
redundancy (which characterises these networks) the question of
the observability of the system arises, rendering this approach
whfeasible for general distribution networks. Notwithstanding,
the Kalman filter approach can be the right one in situations

where the observability condition is met, like in long pipelines
or in main distribution systems feeding the distribution networks
of a more complex system (these main distribution systems are not
255
looped networks, in general). Another situation where Kalman
filters have been successfully used in water distribution systems
is when dealing with reduced models, like the "replication model"
(see Gilman et al. (1973), de Moyer (1973) or de Moyer and
Horwitz (1975)] where, because of the model reduction, the
observability problem does not arise; however this is not our
objective, since we are interested in full scale models. We have
tried, unsuccessfully, to overcome the observability problem via
the introduction of pseudo-measurements, following a similar
approach to that used in power systems, but the solution via
pseudo-measurements seems to be more appropriate for solving the
problems of missing data and bad data replacement due to
communication failure in telemetered systems, rather than
overcoming the observability question. Research in this field has
been quite active in recent years, which is reflected in the
literature: Krumpholz et al. (1980), Alvarez and Albertos (1982),
Clements et al. (1983), Lo et al. (1983), Bargiela (1985),
Monticelli and Wu (1985 a,b). An alternative possible estimation
technique which by-passes the observability problem is "Kriging",
a geostatistical estimation technique which originated in mining
engineering and is used nowadays in the water resources field
(hydrology, groundwater, etc.).
Following a deterministic approach, the problem is formulated
here in a one dimensional fashion, and the aim is to fit a broken
line to the set of data points (measured heads) and then
interpolate the unmeasured heads. This can be done either via
least-square fitting techniques or interpolation methods. The
main shortcoming of this approach is the fact that the errors
256
associated with the measurements cant be handled and, as a
result, the errors associated with the estimates cannot be
produced. Nevertheless, we shall explore the effectiveness of an
interpolation method, which uses the information available on
pipe resistances and we shall compare its results with those
produced via a probabilistic scheme.
A third alternative for head estimation, which somehow lies in
between the probabilistic and the deterministic approaches is
represented by the use of splines, basically a deterministic data
interpolation and approximation technique which is also able to
handle noisy data; this latter issue is known as the "statistical
problem" in the splines literature. We shall explore the
feasibility of using splines for head estimation in the final
part of this Section.
6.9.2. Estimating the piezometric heads using Kriging.
6.9.2.1. Introduction.
We shall review the main assumptions underlying the application
of Kriging in the estimation of the unmeasured piezometric heads
in a water distribution network. We also summarise, at the end of
this section, the main steps followed in the estimation of
unmeasured heads using Kriging.
We assume that the piezometric heads in the network can be
represented by a continuous piezometric plane "floating" above
the network, at a distance from the ground level which is
equivalent to the pressure at each node of the system. Recently,
257
Hamberg and Shamir (1988 a,b) have used the same concept for
different purposes, though they extended the idea of a continuous
plane to the flows as well.
The spatial variability of the piezometric plane is not random,
it has an underlying structure which is a function of the
distance (i.e. two head measurements close enough should give
nearly the same result) and of the network characteristics. The
structure of the piezometric plane is such that its higher levels
are defined by the reservoir levels and then it starts to
diminish towards the lower parts of the system. The piezometric
plane then has a "trend" in the space, with a maximum level at
the reservoirs (or source pumps, or PRV's, etc.) and a minimum
level near the borders of the network. This may be valid for a
whole network or a sub-network; for example: in multiple source
systems, for every source we can define a sub-network, which
contains all the pipes and nodes fed from this particular source.
The same subdivision arises when a series of pressure regulating
valves are separating the main network into different pressure
zones, each pressure zone becoming a sub-network.
The piezometric plane actually exists only over the pipes, and
does not exist over the rest of the space but, since we are only
interested in the plane's behaviour over the pipes, there is no
reason to prevent us thinking of the piezometric plane as a
continuous plane. Moreover, the traditional way of modelling
water distribution networks is mainly concerned with the
piezometric heads above each node (including reservoir levels);
thus we restrict ourselves to the values of the plane at some
258
particular points of the space. On the other hand, current
models assume that the head loss between two connected nodes is
linearly distributed; this is only because in the model the
consumptions are arbitrarily concentrated in the nodes and not
along the pipes, as in the real network. As a result, we get a
representation of the piezometric head which is a set of straight
lines, broken at the nodes. This discontinuous model may be far
from the real piezometric head behaviour, since nodal

consumptions are just a fiction, created to produce a
mathematically tractable problem.
On top of that, the continuous plane representation of the
piezometric heads in the network provides an automatic way for
considering the looped nature of most water distribution
networks, since a head increase in a particular node of the
network (due to a booster pump, for example) will have an impact
on surrounding nodes, increasing their heads in a way which is
easily handled by the continuous plane approach. Another
interesting feature of this way of modelling the piezometric
plane is that we can, eventually, use head (pressure)
measurements taken not only over the nodes, but also measurements
taken at any point of the network, to estimate the head over
unmeasured nodes.
Finally, we have to stress that even though we are imagining
the piezometric head as a continuous plane, we are not actually
using it as such, since we are constraining ourselves to the
heads over the nodes only. The continuity assumption helps us to
handle the underlying structure of the piezometric heads in the
259
network.
We are aware that Kriging might not be the best solution when
dealing with networks whose piezometric plane is highly
irregular, e.g. when P.R.V.'s and booster pumps are located in
some pipes of the network.
We assume, for the time being, that the pressure measurements
are located at key points of the network and that they are
numerous enough, so that the main features of the piezometric
plane are contained in the measurement set. We believe that this
is not a very stringent requirement and, to explain it, we can
take as an example the case when we need to draw a topographic
map; in that case, if we fail to take measurements at key points
(i.e. a hill, a hole, etc.), we shall miss them when the time
comes to "reconstruct" the topography in a contour map. In other
words, this means that any estimation method (for the piezometric
heads in our case, or for ground levels in the topographic case)
will give "bad results" if the assumption on the location and
amount of measurements is not met. These concepts are dealt with
in the characterisation stage of the Kriging process, something
which in mining geostatistics in referred to as "structural
analysis", the characterisation being a stage previous to the
parameter estimation itself.
6.9.2.2. Estimation using Kriging.
For completeness, we have included in Appendix B the derivation
of the Kriging estimator equations.
260
Following Appendix B, the main steps in estimating the
unmeasured piezometric heads using Kriging can be summarised as
follows:
a) Read the data corresponding to the "n" piezometric head
measurements (Zi) and their location (xi, i=1, n).
b) Based on the information of the field data points, i.e.
piezometric heads at certain nodes of the network (their location
given by their coordinates projected on the horizontal plane),
the experimental semi-variogram (or simply variogram) is
constructed, via a procedure similar to that presented in Table
B.1. Fig. B.1. represents an example of a resulting experimental
(or estimated) variogram.
c) One of the five analytical models summarised in Table B.2.
is selected as the best representation of the experimental
variogram. The parameters Ao, B o and Co are determined. This has
been done in our case on a trial-and-error basis, plotting the
experimental variogram and the modelled variograms and choosing
that model which follows the experimental variogram closest.
This step is critical, since the structure of the piezometric
head plane is encapsulated in the semi-vario g ram and the quality
of the estimation process will be strongly dependent on the
quality of the variogram model produced at this stage.
d) With the specification, in the data set, of the points where

the estimation is required (say x 0 ), and with the analytical
model fitted to the experimental variogram, the Universal Kriging
linear system of equations is set up and solved for the extended
261
set of unknowns weights X's and P's in equation (62) or (63) of
Appendix B, which are reproduced here as equations (11) and (12):
E
j
ko j k
k
K(Xi-Xj) + E Pk P (Xi) = K(Xi - X0)J
1=1,...,n
and j=1,...,n
E'-'
i
k'i k
(Xi) = P (Xo)] k=1, . .. , m
Equation (11) constitutes a system of (n+m) equations in (n+m)
unknowns: Xo i i=1,...,n and Pic k1,. ..,m, which in expanded form
can be seen as:
K 11 K 12 K 13 • 1 D xol
• • Ir.ln -12— P1k K10
xo2
K 21 K 22 K 23 . . • K in P 22 ••• P2k K20
1: • : . • • :
• • . . :
• •••
'Ikon
Kn1 Kn2 Kn3 • • Knn 1Pn2 — Pnk KnO
.1 0 0... 0 * =
1 (12)
1 1 1. -411
P21
. P 22
. P 2
.3 • • P 2n 9 0 ...
• • •
0
:
-P2 P20
: • • •
• • Pmn 0 0 .. 0 -P m Pm0
Pm1 Pm2 Pm3 • -
where
= K(xi - xj)
and
Pij = Pi(Xj)
This represents a simplification of Universal Kriging, since the
generalised covariance K(x) has been approximated by the
analytical model determined in stage c). The computer program
used for Kriging allows the specification of the order (k) of

the drift polynomial pk (x) representing the trend of the
piezometric plane [see equation (56) in Appendix B]; in our case
we found out by successive trials that a linear drift was
262
adequate.
e) With the values of the unknown weights [Vs in equations
(11) or (12)], the estimates of the piezometric heads at the
unmeasured nodes (x 0 ) can be computed via equation (57) in
Appendix B, which is reproduced here as:
Z0* = E X0 i Zi (13)
where:
Zi represents the set of measured piezometric heads.
X0 i represents the unknown Kriging weights associated with the
data set.
f) Having determined the values of the extended set of unknown
weights [Vs and P's in equations (11) and (12)] , the Kriging
variance is computed using equation (66) of Appendix B, for each
point where the estimation is required (x0):
0' 0 2 = var(Z 0 *-Z 0 )= K(0) + E Pk p k ( x 0 ) - EX 0 i K(xi-x0) (14)

k i
The main steps of the piezometric head estimation process using
Kriging are shown in Figure 6.2. This flowchart is connected with
that shown in Fig. 6.1, in fact, with the exception of the input
data section, Fig. 6 2 can replace the block "estimate unmeasured
heads" in Fig. 6.1.

\\\\
Read field data points:
location and measured
piezometric heads.
Determine the experimental

variogram and fit an
analytical model to it.
Let x o be an unmeasured head

point to be estimated.
Assemble and solve the Universal

Kriging linear system of equations
determining X's and p's:
E X0 .3 K(Xi-x) + E Pk P(Xi) = KOli-X0)]
i=1,.. .,n
E Xo i Pk (xi) = Pk (X0 )] k1,. ..,m
Compute the estimates of the

unmeasured piezometric heads:
Zo* = E X0 1 Zi
and the Kriging yariance
co2=1“°)+ElikPk(Xo)-EXo1K(Xi-xo)
are there
more points to b
estimated yes
no
( Stop )
Fig. 6.2. Estimation of the piezometric heads using Kriging.
264
6.9.3. Estimating the piezometric heads using a deterministic
one-dimensional interpolation method.
Prior to the calibration stage in the modelling of a water
supply distribution network, a raw network model has been built
by assuming the values of the resistance parameters for the
pipes; these assumptions have been made based on the best
information available, i.e. pipe material, age of the pipes and,
eventually, some field or laboratory tests.
The information contained in the raw network model has not been
explicitly used in the Kriging approach for estimating the
piezometric heads, but it is clearly relevant, especially if a
systematic initial roughness determination procedure has been
carried out [like those described by Walski (1984, chapter 8),
for example]. In so doing, we may expect local differences in the
assumed roughnesses, due to unforeseeable events, with respect to
the true roughnesses, but, on average, we may also expect that
they do contain the main features of the roughness
characteristics of the network. Consequently, we should try to
use this information already contained in the raw model, together
with piezometric head measurements, in order to obtain improved
head estimation in the unmeasured nodes. This is precisely what
we do in the deterministic one-dimensional interpolation method.
The main steps in the proposed deterministic one-dimensional
interpolation procedure are:
265
a) Read the data corresponding to the piezometric head
measurements and the initial values for the pipe roughnesses.
b) Run the raw network model (gradient method program) and
determine pipe flows, nodal piezometric heads, and, particularly,
pipe head losses (amount and direction). This is a standard run
of the gradient method for the network analysis problem. The
program is run with the initially assumed roughnesses of the
pipes.
c) Determine the minimum head loss spanning trees, rooted at
each reservoir (if more than one). This gives us the paths
connecting each node with a reservoir, following a minimum head
loss criterion. The shortest path algorithm used is Dijkstra's
algorithm [see Smith (1982) or Deo (1974)], where the distance
between nodes has been replaced by the head loss. The reason why
the pipe head loss has been chosen as the criterion for
determining the spanning tree will become apparent in the
following paragraphs.
d) In the case of more than one reservoir, resolve the linkage
of those nodes which are included in more than one spanning tree,
maintaining their linkage only to that reservoir which leads to
the minimum head loss to the node.
e) For each spanning tree, generate as many paths (one-
dimensional arrangements of successive nodes), starting at a
reservoir and following the spanning tree downstream, as to sweep
all the nodes in that spanning tree. As we proceed downstream,
the initial node of a path may be another measured node (not
266
necessaril y a reservoir) or a previously interpolated node. When
information on the accuracy of the head measurements is
available, we should include the nodes with higher accuracy
first, so that the lower accuracy nodes (and the unmeasured ones)
are handled with the best information available, thus minimising
the error of the interpolation.
The minimum head loss criterion has been used to generate the
spanning tree, instead of'other criteria like maximum flow or
maximum head loss, because it produces paths having a flat
piezometric line, thus minimizing the error in the interpolation.
f) Having the paths which cover all the nodes of the network,
perform the one-dimensional interpolation in order to force
coincidence at the measured nodes between modelled and measured
piezometric heads, thus obtaining interpolated heads for all the
unmeasured nodes. This point is explained in more detail in the
next section.
The whole one-dimensional deterministic interpolation process
is summarised in Fig. 6.3. This flowchart is also connected with
that shown in Fig. 6.1, in fact, with the exception of the input
data section and the estimation of initial roughnesses, Fig. 6.3
can replace the block corresponding to the estimation of
unmeasured heads in Fig. 6.1.

Read field data points: Read initial values for
location and measured pipe roughnesses:
piezometric heads. C's
Run the network analysis

program (raw model), determine
the head losses per pipe: h
4,
Generate the minimum head loss
spanning trees rooted at each
reservoir and generate the
paths linking each node with a
measured or already interpolated
node
Interpolate the piezometric

heads of the unmeasured nodes
within this path
are there
more paths to be
interpolated yes
no
( Stop)
Fig. 6.3. Estimation of piezometric heads using the one-

dimensional deterministic interpolation scheme.
6.9.3.2. The one-dimensional interpolation.
For the one-dimensional interpolation, we take each one of the
paths (generated via the minimum head-loss spanning trees) at a
time.and we compare the modelled piezometric line for every path

with the corresponding head measurements. This can be represented
as in Fig. 6.4.
Because each path starts in a measured node (or in an already
estimated node), we search along the path for the next measured
node, defining a reach between two known head nodes [see Fig.
6.4, where there are three reaches]. Having selected a reach in
the path, we then compute the difference between the modelled
piezometric heads corresponding to the initial and final node of
the reach (DIF1 in Fig. 6.4) and also that corresponding to the
observed (measured) piezometric line (DIF2 in Fig. 6.4) and we
seek the coincidence of both modelled and observed piezometric
lines at the measured nodes (initial and final nodes of the
reach). To do so, we force the coincidence of modelled and
observed piezometric lines at the initial node and spread the
head difference DIF=DIF1-DIF2 along the reach in proportion to
the length, so that coincidence will be achieved at the final
node of the reach as well; the intermediate unmeasured (modelled)
nodal heads are modified accordingly, becoming the "estimates"
produced by this interpolation algorithm.
When a final measured node is not available, as in the last
reach of Fig. 6.4, the modelled piezometric line nf this last
reach is moved parallel in order to force agreement in the
initial node, since this is the only information available.
269
Piezometric
Hood
(1.4)
Measured piezometric head (known)

Modelled piezometrle line (known).
DIF1 Measured piezornetric head (known)

DIF2
Agi piezanervic line (unkn.
OD- Length (1.1)

0 500 1000
1 3 5 9 16 20 31 19 Node Number
reach 1 reach 2 reach 3
Fig. 6.4. Minimum head loss path of nodes, before the head
interpolation has been carried out.
PiezcmetrIc
Hood
(H)
Measured plezometric hoods (known)
Length (14)
500 1000
1 3 5 9 16 20 31 19 Node Number
Fig. 6.5. Minimum head loss path of nodes after the head
interpolation has been carried out.
270
The result of this one-dimensional interpolation procedure, for
a particular path, after having gone through all the possible
reaches, is shown in Fig. 6.5, where the adjusted modelled
piezometric line coincides with the observed piezometric line at
the measured nodes, while still keeping most of the information
on the initially assumed roughnesses of the pipes.
The algorithm relies on the fact that the piezometric head has
been measured at "key" nodes of the paths, for example:
- reservoirs (level measurements): these measurements are
considered as indispensable.
- end points of the paths.
- intermediate points, between reservoirs and end points,
especially if relevant changes in the piezometric line take
place, due to higher consumptions or changes in the pipe
characteristics.
The algorithm is deterministic, in the sense thatais not able
to consider the error of the head measurements explicitly and, as
a result, is not able to produce an estimate of the error in the
interpolated (unmeasured) piezometric head.

6.9,4, Estimating the piezometric heads using bi-cubic splines.
In the previous sections we have been dealing with two
different ways of estimating the piezometric heads: Kriging and a
one-dimensional interpolation method. The first one is a
stochastic method, which allows us to handle explicitly the
errors associated with the measurements and also can give an
estimate of the error associated with the estimated piezometric
heads at the unmeasured nodes. The interpolation method is a
deterministic approach and, as such, it does not allow us to
handle the errors either in the measurements or the estimates.
In the Kriging-based method the structure of the piezometric
head plane is encapsulated in the semi-variogram and the approach
is valid when there is some correlation in the spatial
distribution of the piezometric plane (i.e. its variation is not
random).
In the interpolation method we make use of our a priori
knowledge of the physical characteristics of the network. We
blend this knowledge with the information provided by the head
measurements, to produce an estimate of the piezometric heads in
the unmeasured nodes.
The traditional way of representing the modelled piezometric
heads has been through a set of broken straight lines, with the
heads being computed at the nodes only and being linearly
interpolated elsewhere. The one-dimensional interpolation
technique already presented follows more closely this way of
272
representing the piezometric heads. It can be argued that, as far
as the estimation of the piezometric heads is concerned, Kriging
and the one-dimensional interpolation represent two completely
opposite approaches. So, perhaps another approach halfway between
Krigin g and the one-dimensional interpolation scheme can

offer a further alternative.
In the search for such an approach, and having in mind the need
for an explicit handling of the errors in both the measurements
and the estimates, bi-cubic splines came to our attention,
particularly because of their reported ability to represent large
and complex geodetic systems and terrain following problems
[Anthony and Cox (1987)] , which are somehow similar to our
piezometric head estimation problem, especially when pumps and
regulating valves are present within the network. Splines have
also been used in parameter estimation [Lainiotis and Desphante
(1974)]. The possibility of obtaining the errors associated with
the estimates puts the splines somewhere in between the two
methods we have used so far: Kriging and the one-dimensional
interpolation method.
6.9.4.2. Bi-cubic splines estimation.
Appendix C includes the derivation of the bi-cubic splines
approximation equations and also includes the estimation of the
error associated with the estimates. The main steps in the
estimation of the piezometric heads using bi-cubic splines are
the following:
273
a) Read the field data points representing the measured nodal
piezometric heads:
fxr, Yr, f (xr, Yr1 r=1, 2, ..., m
where xr and yr represent the horizontal projection of the
coordinates of the measured point and f(xr, yr) is the

corresponding value of the observed piezometric head. Also,
the position of possible discontinuities in the spline polynomial
or in its derivatives must be specified, within the range of the
independent variables; this is done through the specification of
the location of the "interior knots", both in the X and Y-
directions
X-direction knots: Xi1=1,2, ..., h
Y-direction knots: p J- j=1,2, ..., k
The exterior knots, needed for continuity reasons, are
automatically placed by the computer program used.
b) With the data information, compute the values of the B-
splines Mi(xr) and Ni(yr) using the recursive scheme represented
by equations (4) and (5) in the Appendix C, set up the normal

e quations and solve them for the unknown vector of "weights" r ,
i.e., from equation 14, Appendix C, solve
[AT A] r = AT f (15)
where:
A : (mx(h+4)(k+4)) matrix obtained from the observation

equations (equation 12, Appendix C):
h+4 k+4
E E rij Mi(xr) Nj(yr) = f(xr, yr) = fr (16)
1=1 j=1 r = 1,2, ..., m
274
r : (h+4)(k+4)xl column vector of unknowns, with components rij
in the previous equation.
f : mx1 column vector, given by the components fr in the

observation equations.
c) For each one of the "n" pre-specified nodes where the
piezometric head estimation is required:
(x l , y l ) 1 = 1, 2, ..., n
compute the B-splines Mi(xl) Nj(yi), using the recursive
equations (4) and (5) of Appendix C, then introduce them into
equation (11) of Appendix C to compute the splines approximates

of the piezometric head at the unmeasured nodes:
h+4 k+4
s(xl, yl) = E E rij Mi(xl) Nj(yi) (17)
i=1 j=1
d) If required, compute the covariances and variances of r,
using equations (26) and (27) of Appendix C and use them to
determine the variances of the spline estimates.
The piezometric head estimation process using bi-cubic splines
is shown in Fig. 6.6. This flowchart is connected with that shown
in Fig. 6.1, in fact, with the exception of the input data
section, Fig. 6.6 can replace the estimation of unmeasured heads
block in Fig. 6.1.
The computer software used is that contained in the NAG-
Library. Step b) is carried out with the subroutine E02DAF, while
step c) is performed using the subroutine E02DBF.
275
Read field data points:
location and measured
piezometric heads.
Assemble the normal linear

system of equations:
[A T A] r = A T f
and solve it for the vector r
with components rij
,
t
Let x=(x,y) be an unmeasured
head point to be estimated.
1 ,
Compute the estimate of the
unmeasured piezometric head:
s(x,y)= E E rij
M(x)N(y) 4
i j
and the variance, if needed
are there
more points to be
estimated yes
?
no
CStop )
Fig. 6.6. Estimation of piezometric heads using bi-cubic splines.
276
6.10. Examples of applications of the proposed calibration
method.
6.10.1. Description of the networks used as examples.
Six examples, based on two basic networks subjected to
different operating conditions, will be used to carry out a
comparison of the performance of both the proposed explicit
calibration method and the different piezometric head estimation
techniques.
i) Example A:
This example uses a small network, made up of 16 pipes, 11
nodes and 1 reservoir, under a low demand condition. The network
structure is shown in Fig. 6.7, whereas Table 6.1 shows the
corresponding numerical data. Table 6.1 a) includes the true

Hazen-Williams roughness considered, both in this example and in
the following two examples. Table 6.1. b) shows the nodal data.
Note that the nodal consumptions for this example (example A) are
the same as that for the third (example C), and that they
represent a low demand condition.
ii) Example B:
The same network used in example A is subjected to higher nodal
consumptions, as defined in Table 6.1. b). The objective is
to study the performance of the calibration algorithm under
increased demand conditions.

iii) Example C:
The same basic network used in example A will be used with a
worse initial assumption for the Hazen-Williams roughness, as
defined in section 6.10.4.
iv) Example D:
We use this time a 180 pipe, 100 nodes network, arranged in a
square grid network of 10 nodes per side. The geometrical
characteristics of this network are such that the length and the
diameter of every pipe is the same (150 m. and 100 mm.,
respectively). As far as the nodes are concerned, the true demand
is 1.0 (1/s) for all non-reservoir nodes and the ground level is 10
m. for all the nodes. This basic network, with minor changes, is
used under 3 different confi gurations, leading to examples D, E
and F.
65 m
.1.
6 1
9 e:f.‘
n w 14
. 0
Fig. 6.7. Network for Examples A, B and C.
278
Table 6.1. Network data for Examples A, B and C.
a) Pipe data:
LENGTH DIAMETER HAZEN-WILLIAMS

NO. NODE NODE (M) (MM) TRUE "C'S"
1 1 2 500.0 200.0 80.0
2 2 3 650.0 150.0 85.0
3 2 4 400.0 150.0 90.0
4 2 5 575.0 150.0 95.0
5 4 3 500.0 100.0 100.0
6 4 5 400.0 100.0 105.0
7 3 6 500.0 100.0 110.0
8 4 7 500.0 100.0 115.0
9 5 8 500.0 100.0 120.0
10 7 6 500.0 100.0 125.0
11 7 8 400.0 100.0 130.0
12 6 9 500.0 100.0 135.0
13 7 10 500.0 100.0 140.0
14 8 11 500.0 100.0 145.0
15 10 9 500.0 100.0 150.0
16 10 11 400.0 100.0 155.0
•
b) Nodal data: only nodal demand changes between examples A, B

and C.
NODE TYPE DEMANDS (L/S) GROUND RESERVOIR

EXAMPLES LEVEL LEVEL
NO. A and C B (M) (M)
1 1 0.0 0.0 50.0 65.0
2 0 0.0 0.0 45.0
3 0 1.5 2.0 41.0
4 0 1.5 2.5 41.0
5 0 1.5 2.5 41.0
6 0 2.0 3.0 36.0
7 0 2.0 3.0 36.0
8 0 2.0 4.0 36.0
9 0 1.5 4.5 31.0
10 0 1.5 4.5 31.0
11 0 1.5 5.0 31.0
Note: Type = 1 for reservoirs, 0 for unknown head nodes.
279
In Example D only one reservoir feeds the whole network and is
connected to the upper left-hand corner (node 1); the reservoir
level is 100 m. The network is shown in Fig. 6.8.
v) Example E:
Using the same basic network as in example D, we now consider
that four reservoirs are feeding the network and they are
connected to each corner of the square (nodes 1, 10, 91 and 100);
the reservoir levels are: 100.0, 90.0, 90.0 and 110.0 m.,
respectively. This network is presented in Fig. 6.9.
vi) Example F:
Basically the same network is employed, as in Example E, with
some minor changes to introduce a line of 4 pressure reducing
valves, separating the network into high and low pressure zones.
The network has 178 links (4 of them pressure reducing valves,
the rest are pipes), 104 nodes and 4 reservoirs (connected at
nodes 1, 55, 64 and 100, with reservoir levels: 110.0, 100.0,
50.0 and 40.0 m., respectively). The node and link labelling has
been changed with respect to the previous two examples, in order
to get a consecutive numbering for the higher and lower pressure
zones. The P.R.V.'s are located at links 175, 176, 177 and 178
and all of them have a minimum resistance parameter (a) equal to

5.0x10 -5 (i.e. when valves are fully open) and the outlet head is
set to 50.0 m. This network is shown in Fig. 6.10.

100.0M
1J=
1
0.111 I I I 1). I I I I lig I lard I I 1111 I I I
0:0 30 2:031=1=0
0•
. D,SIMITIN411
1
crairaterglimill
ess o e
57
75
Rind
1795
G
N
'211
IS. 114
e
133
0
on ma
CD
171
Fig. 6.8. Network for Example D.
281
Fig. 6.9. Network for Example E.
282
110.0 M. 50.0 M.
Fi g . 6.10. Network for Example F.
283
Due to the uniformity of the data for the last 3 network
examples, their numerical data are not presented in detail here.
6.10.2. Defining the "true" network characteristics.
In order to test the ability of the calibration algorithm to
estimate the "true" roughnesses of the example networks, we
defined a set of true roughnesses for each one of them:
a) For examples A, B and C (smaller networks) the true Hazen-
Williams roughnesses were set to 80, 85, 90, ..., 160 for the 16
pipes (i.e. starting with C=80, we incremented C by 5 for each
successive pipe), as shown in Table 6.1 a).
b) For examples D, E and F (bigger networks) the true Hazen-
Williams roughnesses of all pipes was set to 140.
6.10.3. Defining the measurements for the network examples.
With these "true" roughnesses, and with the physical
characteristics of the network defined in 6.10.1., we ran the
network analysis program (gradient method) to determine the
steady state flows and piezometric heads of the "true" systems.
From the results of these runs we obtained our sets of
"measurements", i.e. heads and flow measurements at selected
nodes and pipes of the networks.
To investigate the effect of flow measurements taken in
different pipes within the networks, two sets of flow
measurements were produced. Table 6.2 shows the measurement data
considered for Examples A B and C, while Table 6.3 corresponds to
Examples D, E and F.
284
Table 6.2. Measurement data for Examples A, B and C.
a) Piezometric head (m) measurement data:
Example A and C Example B
Node Piez. Head Node Piez. Head
1 65.000000 1 65.000000
4 62.818108 4 56.614697
9 60.988101 9 47.368164
11 60.988960 11 47.364705
b) Flow (1/s) measurement data:

1
1st set of flow measurements 2nd set of flow measurements
I
Examples A and C Example B 1 Examples A and C Example B
Pipe Flow Pipe Flow I Pipe Flow Pipe Flow
I
3 5.7045 3 11.82781 3 5.7045 3 11.8278
11 -0.1018 11 0.4583' 4 4.9642 4 10.3282
I
Note: flowrates are considered positive when they flow from the
initial into the final node of the pipe (see Table 6.1.a)
Table 6.3. Measurement data for Examples D, E and F.

a) Piezometric head (m) measurement data:
Example D Example E Example F
Node Piez. Head Node Piez. Head Node Piez. Head
1 100.000000 1 100.000000 1 110.000000
4 28.042592 4 90.087208 4 84.022801
7 21.145426 7 89.670934 6 81.718164
10 20.224128 10 90.000000 19 85.304132
31 28.042592 31 90.087208 22 83.234221
34 24.315165 34 89.797302 24 81.517016
37 20.819741 37 89.696615 37 84.756624
40 20.131486 40 89.707143 40 82.859427
56 20.665754 56 89.782311 42 80.841693
61 21.145426 61 89.670934 55 100.000000
64 20.819741 64 89.696615 58 82.891918
70 19.875863 70 90.882238 60 79.946825
78 19.876248 78 91.207603 64 50.000000
91 20.224128 91 90.000000 76 46.744788
94 20.131486 94 89.707143 88 45.060264
97 19.875863 97 90.882238 100 40.000000
100 19.770666 100 110.000000 101 50.000005
102 49.999780
103 50.000154
104 49.999025
285
Table 6.3. Measurement data for Examples D, E and F.
(Continuation).
b) Flow (l/s) measurement data:
First set of flow measurements

Pipe Flow Pipe Flow Pipe Flow
1 49.5000 1 17.3758 1 28.5634
5 8.9960 5 1.7104 5 5.8167
10 49.5000 10 17.3758 6 28.3576
11 20.6499 11 7.0481 7 11.5337
20 20.6499 20 7.0481 12 11.9710
38 0.8997 38 0.9625 33 4.0455
53 3.6985 53 -0.4622 48 5.5255
79 5.5638 79 0.7851 50 1.3063
86 8.9960 86 1.7104 83 -11.6611
114 1.6112 114 -5.2377 92 6.6734
121 2.1334 121 -3.3762 118 5.2651
143 2.9574 143 -1.1708 146 4.5898
151 1.0259 151 -9.5725 147 5.0332
156 1.6824 156 -1.7850 159 5.2148
Second set of flow measurements

Pipe Flow Pipe Flow Pipe Flow
1 49.5000 1 17.3758 1 28.5634
2 27.8501 10 17.3758 4 7.3102
4 12.4275 11 7.0481 6 28.3576
10 49.5000 19 3.0247 12 11.9710
13 4.4079 51 1.9283 84 -7.8444
20 20.6499 130 -3.3762 89 9.5888
47 1.3852 149 -3.1766 94 -22.2490
77 2.4315 151 -9.5725 106 -2.4649
86 8.9960 158 -4.8944 107 -7.8444
96 1.4620 170 -10.0725 151 4.4264
99 4.0688 171 -24.5171 171 7.8931
108 5.2679 172 3.0247 172 10.2300
120 2.6334 173 0.9625 173 12.6191
122 1.4826 180 -24.5171 174 13.0494
Note: flowrates are considered positive when they flow from the
initial into the final node of the pipe, which is
indicated by an arrow in Figures 6.8, 6.9 and 6.10. Flows
are negative otherwise.
286
6 10.4. Defining the initially assumed roughnesses for the
network examples.
In order to simulate the lack of knowledge of the true
roughnesses of the pipes prior to the calibration process, we
assumed that on average we knew the values of the C's and we
proceeded to generate a new set of C's, with the same mean as the
true values, but with some added Gaussian noise. This was done
with a pseudo-random number generator program, and the
corresponding values are presented in Table 6.4 (for examples A
and B), for two different degrees of uncertainty in the
roughnesses. Table 6.5 shows the initial roughnesses considered
for example C, which have been obtained for the initially assumed
C's for example A (low uncertainty) , multiplied by a factor
1.2., i.e. only in this particular example, we assume that the
initial C estimates are not coinciding - on average - with the
true values, but they have been over-estimated by 20%. Table 6.6.
presents the initial C's assumed for examples D, E and F, for low
and high degree of uncertainty in their estimates. These values
have been generated by adding some Gaussian noise to the true

C=140.
for testing the calibration algorithm.
For Examples A and B and low and high uncertainty.
Uncertainty in C
Low High
Pipe
Values of C
1 81.4482 73.7684
2 89.6747 88.9714
3 95.4802 96.5187
4 98.6072 95.5956
5 103.7588 101.0750
6 116.2208 129.6719
7 116.7340 120.4833
8 121.4841 124.6930
9 126.4144 129.4729
10 130.4846 131.5323
11 133.3617 129.8191
12 137.0585 130.6980
13 148.0070 154.5090
14 152.2471 157.1059
15 153.0364 148.7907
16 162.2371 167.0742
Mean 122.8909 123.7362
Variance 596.1455 702.9258
For the true C's:
Mean 122.5000 122.5000
Variance 566.6667 566.6667
288
for testing the calibration algorithm in Example C.
.
Pipe Values of C
1 97.738
2 107.610
3 114.576
4 118.328
5 124.511
6 139.465
7 140.081
8 141.781
9 151.697
10 156.582
11 160.034
12 164.471
13 177.608
14 182.696
15 183.643
16 194.684
.nn
Mean 147.219
Variance 860.341
For the true C's:

Mean 122.500
Variance 566.667
289
Table 6.6. Initial values of C (Hazen-Williams) used for testing
the calibration algorithm. For Examples D, E and F,
and for low and high uncertainty.
C values * C values * C values *

Pi pe Pipe Pipe
Uncertainty in C Uncertainty in C Uncertainty in C
Low High Low High Low High
1 136.777 135.4418 61 139.265 138.9603 121 143.384 144.7855
2 139.713 139.5943 62 133.002 130.1041 122 138.545 137.9428
3 140.412 140.5830 63 141.288 141.8213 123 140.936 141.3244
4 138.835 138.3520 64 139.622 139.4660 124 139.777 139.6850
5 138.987 138.5668 65 136.080 134.4567 125 141.905 142.6946
6 144.962 147.0180 66 141.210 141.7110 126 143.297 144.6623
7 141.353 141.9132 67 140.447 140.6316 127 136.971 135.7164
8 141.134 141.6030 68 142.781 143.9334 128 137.310 136.1954
9 141.058 141.4969 69 137.335 136.2316 1ZG 13.g . q4€ t3q.eeT4
10 140.356 140.5029 70 139.742 139.6348 130 136.762 135.4204
11 138.820 138.3316 71 143.479 144.9200 131 141.459 142.0632
12 137.920 137.0585 72 141.429 142.0208 132 134.405 132.0869
13 142.089 142.9547 73 138.294 137.5873 133 1414317 141.8623
14 141.535 142.1709 74 135.402 133.4972 134 136.535 135.0998
15 138.680 138.1339 75 142.383 143.3697 135 143.593 145.0811
16 141.480 142.0926 76 140.139 140.1965 136 140.819 141.1581
17 138.061 137.2573 77 139.836 139.7682 137 136.643 135.2530
18 143.167 144.4793 78 142.321 143.2820 138 138.641 138.0779
19 138.037 137.2233 79 141.605 142.2702 139 141.706 142.4125
20 137.395 136.3165 80 137.548 136.5322 140 138.438 137.7913
21 140.941 141.3304 81 138.622 138.0510 141 131.478 127.9477
22 140.602 140.8510 82 139.878 139.8282 142 132.876 129.9255
23 136.461 134.9955 83 142.079 142.9401 143 142.037 142.8810
24 142.466 143.4878 84 141.535 142.1711 144 133.725 131.1257
25 137.879 137.0008 85 143.508 144.9613 145 140.406 140.5746
26 140.731 141.0338 86 138.512 137.8961 146 144.464 146.3133
27 145.442 147.6956 87 137.641 136.6643 147 138.508 137.8897
28 140.408 140.5774 88 142.027 142.8665 148 144.708 146.6582
29 143.092 144.3722 89 143.049 144.3116 149 138.238 137.5084
30 141.982 142.8027 90 134.054 131.5905 150 137.330 136.2236
31 137.353 136.2564 91 137.938 137.0846 151 145.478 147.7466
32 143.962 145.6036 92 142.915 144.1229 152 142.393 143.3842
33 138.493 137.8690 93 137.323 136.2145 153 141.975 142.7928
34 142.970 144.1997 94 137.420 136.3516 154 138.677 138.1289
35 140.924 141.3075 95 141.167 141.6508 155 139.800 139.7175
36 140.138 140.1959 96 141.969 142.7846 156 136.602 135.1939
37 139.260 138.9536 97 142.954 144.1778 157 134.132 131.7016
38 141.594 142.2545 98 136.841 135.5324 158 141.742 142.4635
39 141.886 142.6677 99 142.205 143.1184 159 135.740 133.9759
40 141.829 142.5865 100 136.540 135.1062 160 137.425 136.3581
41 142.539 143.5908 101 137.082 135.8732 161 139.667 139.5290
42 141.195 141.6895 102 134.688 132.4885 162 137.828 136.9281
43 143.073 144.3462 103 145.144 147.2742 163 139.878 139.8268
44 142.760 143.9030 104 140.969 141.3699 164 139.190 138.8543
45 142.281 143.2259 105 142.018 142.8537 165 141.531 142.1658
290
Table 6.6. Initial values of C (Hazen-Williams) used for testing
the calibration algorithm. For Examples C, D and E, and
for low and high uncertainty. (Continuation).
C values * C values * C values *

Pipe Pipe Pipe
Uncertainty in C Uncertainty in C Uncertainty in C
Low High Low High Low High
46 133.313 130.5432 106 136.143 134.5457 166 138.747 138.2282
47 140.610 140.8628 107 137.139 135.9537 167 137.079 135.8693
48 138.165 137.4045 108 144.336 146.1318 168 141.363 141.9278
49 140.115 140.1627 109 139.179 138.8386 169 139.669 139.5317
50 140.800 141.1319 110 143.868 145.4707 170 138.230 137.4969
51 142.525 143.5712 111 137.100 135.8987 171 140.582 140.8232
52 143.163 144.4729 112 137.777 136.8568 172 144.035 145.7070
53 138.079 137.2839 113 142.443 143.4547 173 141.078 141.5238
54 142.135 143.0193 114 141.579 142.2325 174 140.617 140.8721
55 140.616 140.8708 115 141.920 142.7149 175 136.865 135.5667
56 138.652 138.0942 116 141.818 142.5717 176 143.407 144.8179
57 140.453 140.6406 117 139.663 139.5230 177 143.506 144.9576
58 134.676 132.4708 118 142.057 142.9091 178 137.298 136.1791
59 141.788 142.5281 119 140.220 140.3106 179 138.035 137.2212
60 140.349 140.4934 120 142.158 143.0526 180 138.745 138.2249
(*) Because example F has only 174 pipes, the first 174 values
of this Table should be considered.
Summary:
*******
For Examples D and E:
Low High
Mean 139.959 139.9418
Variance 7.499 14.9984
For example F:
Low High
Mean : 139.9697 139.9572
Variance : 7.4917 14.9832
For the true C's:
Mean 11X000
Variance 0.000 11410!18000
0.0000
____ \ n
291
6.10.5. Defining a perturbed set of nodal demands.
To study the behaviour of the calibration algorithm when having

an incorrect estimate of the nodal demands, we took the true
nodal consumptions presented in Table 6.1. b) (for examples A, B
and C) and we added some Gaussian noise, producing the sets of
perturbed nodal demands presented in Table 6.7. For examples D, E
and F, where the nodal consumption is 1.0 (l/s) for all the
nodes, we also added some Gaussian noise to this consumption,
generating the new perturbed nodal demands shown in Table 6.8
Table 6.7. Modified nodal demands (us), used to study the impact
of bad demand estimation on the calibration algorithm.
Examples A B,and C.
Demands (1/s)
Node Examples A and C Example B

1 0.0 0.0
2 0.0 0.0
3 1.2016 1.6148
4 1.4734 2.4580
5 1.5382 2.5603
6 1.8754 2.7842
7 1.8917 2.8123
8 2.5305 5.0610
9 1.6252 4.8254
10 1.6049 4.7726
11 1.5980 5.2684
Table 6.8. Modified nodal demands (us), used to study the impact
of bad demand estimation on the calibration algorithm.
Examples D, E and F.
Node Demand Node Demand Node Demand Node Demand

1 0.0 26 0.8612 51 1.0524 76 1.1560
2 0.7890 27 1.0479 52 1.1653 77 1.0091
3 0.9812 28 1.3562 53 1.2071 78 0.9893
4 1.0270 29 1.0267 54 0.8743 79 1.1519
5 0.9237 30 1.2024 55 1.1398 80 1.1051
6 0.9337 31 1.1297 56 1.0403 81 0.8395
7 1.3249 32 0.8267 57 0.9118 82 0.9098
8 1.0886 33 1.2594 58 1.0297 83 0.9920
9 1.0742 34 0.9014 59 0.6515 84 1.1361
10 1.0693 35 1.1944 60 1.1170 85 1.1005
11 1.0233 36 1.0605 61 1.0228 86 1.2297
12 0.9228 37 1.0091 62 0.9519 87 0.9026
13 0.8638 38 0.9516 63 0.5419 88 0.8456
14 1.1368 39 1.1044 64 1.0843 89 1.1327
15 1.1005 40 1.1235 65 0.9753 90 1.1996
16 0.9136 41 1.1197 66 0.7434 91 0.6107
17 1.0969 42 1.1662 67 1.0792 92 0.8650
18 0.8730 43 1.0782 68 1.0292 93 1.1909
19 1.2074 44 1.2012 69 1.1821 94 0.8248
20 0.8715 45 1.1807 70 0.8256 95 0;8311
21 0.8295 46 1.1493 71 0.9831 96 1.0764
22 1.0616 47 0.5622 72 1.2278 97 1.1289
23 1.0394 48 1.0399 73 1.0935 98 1.1934
24 0.7683 49 0.8799 74 0.8883 99 0.7932
25 1.1615 50 1.0075 75 0.6990 100 1.1444
6.11. Comparison of the calibration results using different head
estimation techniques.
6.11.1. Main objectives.
The main purpose of this section is to evaluate the performance
of the proposed calibration algorithm under different conditions.
We also aim at establishing the advantages and disadvantages of
the different head estimation techniques introduced in previous
sections: Kriging, deterministic one-dimensional interpolation and
bi-cubic splines. Basically, what we are trying to find out is
which head estimator has a better response, in the sense of being
able to provide a better estimate of the true parameters and
state variables of the network. Thus, we shall compare not only
the estimates of the pipe roughnesses, but also the flows per
pipe and nodal piezometric heads.
6.11.2. Study cases.
To study the effect of different scenarios of measurement
availability and different degrees of certainty in the initially
assumed values of roughnesses and nodal consumptions, on the
performance of the calibration algorithm, the following study
cases have been defined:
i) Case I.
In this case we assume that all the measurement information
available is provided by head measurements only. The purpose of
this case is to study the response of the calibration algorithm
to different levels of availability of flow measurements, this
294
case establishing the lowest possible level in terms of flow
measurement data.
The amount of head measurements has been set to 4 in networks
A, B and C (1 level measurement plus 3 pressure measurements), 15
head measurements in networks D and E and 20 head measurements in

network F (4 level measurements and 16 pressure measurements,
including inlet and outlet pressure measurements at the
P.R.V.'s). Even though the proportion of number of measurements
to number of nodes may appear high in the smaller networks (A, B
and C) due to their small size (11 nodes), the proportion has
been reduced to more realistic values in the case of networks D,
E and F , although increasing the number of measurements, in view
of the complexity of the networks under study (number of

reservoirs and presence of P.R.V.'s).
The selection of the location of pressure measurements has been
done having in mind that it is convenient to have them evenly
distributed over the network, in order to give a representative
picture of the variation of the piezometric plane. Although we
recognise the importance of pressure measurement placement in the

p erformance of any head estimation technique, we did not follow a
formal measurement placement technique; this is partly justified
by the fact that the networks are fairly regular both in their
geometrical and physical characteristics and also in their
spatial distributions of the nodal demands. In the case of
Example F, with P.R.V.'s, we followed the same principle,

treating each pressure zone as a regular sub-network.
295
ii) Case II.
To the same amount and distribution of piezometric head
measurements used in Case I, we have now added 2 flow
measurements for networks A, B and C, and 14 flow measurements
for networks D, E and F.
The flow measurements are located in exactly the same pipes for
examples A, B and C, and also for networks D, E and F. This
corresponds to a first set of flow measurements, which have been
placed rather arbitrarily, especially as far as examples E and F
are concerned.
iii) Case III.
To study the effect of a more rational flow measurement
placement, the same amount of flow measurements utilised in Case
II has been redistributed , this time by studying the flow
distribution patterns for each network model. Flow measurements
were re-placed in order to get a better picture of the
distribution of the maximum flows, particularly those leaving the
reservoirs or passing through P.R.V.'s. Again, as in the case of
pressure measurement placement, we did not apply a formal
procedure for flow measurement placement.
iv) Case IV.
To investigate the impact of poor initial estimates of the
pipes' roughnesses, we used the example networks, with the head
and flow measurements as in case III, this time with worse
estimates of C, as given in the previous section 6.10.4.
296
v) Case V.
In order to evaluate the behaviour of the calibration algorithm
under perturbations in the spatial variation of the nodal
demands, we used the example networks with the perturbed nodal
demands given in section 6.10.5.
6.11.3. Calibration exercise.
The calibration of each example network (Examples A, B, C, D, E
and F) will be performed, using the proposed algorithm, under the
five different Cases already described, which represent variable

degrees of knowledge of the true values of the main network
characteristics.
The overall calibration process (represented in Figure 6.1.),
of each network and for each one of the five Cases studied, has
involved the following main stages:
a) Estimation of the unmeasured piezometric heads.
Based on the piezometric head measurements available for each
example network, the unmeasured piezometric heads were estimated
using Kriging and bi-cubic splines, resulting in two sets of
estimated heads for each network. Since the measured and
estimated heads do not depend either on the initial roughnesses
assumed, or on the flow measurements, or on the quality of the
nodal consumption data, these two sets of head estimates are the
same for all the five Cases studied, and they depend only on the
network under study. For the one-dimensional interpolation
method, and because the proposed algorithm uses the information
contained in the raw model, based on the a priori information
297
assumed to be available, the head estimates may be different from
case to case. Because Cases I, II and III are based on the same
initial assumptions for pipe roughnesses and nodal consumptions
(i.e. same raw model), and their differences consist only in
different availability of flow measurements, Cases I, II and III
have the same head estimates, but those estimates are different
from those for Cases IV and V, where the initial assumptions
(pipe roughnesses and nodal consumptions, res p ectively) have
changed. This difference becomes clearer in section 6.11.4.3.
The piezometric head estimation stage is carried out only once
during the calibration process, as shown in Fig. 6.1.
b) Estimation of the unmeasured flows.
Using the initially assumed roughnesses and nodal consumptions,
run the network analysis program for the raw model corresponding
to each network and determine the pipe flows. If flow
measurements are available at some pipes, replace the modelled
flows by their corresponding measurements. This leads to a set of
modelled/measured flows for each network and each study case.
0 Calibration stage.
Run the calibration program, using the set of
estimated/measured heads and modelled/measured flows, and
determine the calibrated roughnesses (Hazen-Williams C's).
For each example and study case, piezometric head estimates
produced with Kriging, the one-dimensional interpolation method
and bi-cubic splines have been used, leading to different
298
calibration results, corresponding to each one of the piezometric
head estimation techniques used.
6.11.4. Results of the calibration.
In this section a "performance index" and a "global success"
index are defined for evaluation purposes; later on, the
numerical results are presented, both for the head estimation and
calibration procedures, including a discussion of them.
6.11.4.1. Definition of the "performance index".
Because the calibration exercise generates a large amount of
numerical information, which is impossible to display in full
detail here, we have included in Appendix D a summary of the main
statistics associated with such results. Even then, further
condensation of the results is necessary, in order to quantify
the results, particularly those concerning the performance of the
calibration algorithm and the head estimation techniques. For
those purposes, a performance index has been used to quantify the
degree of successfulness of the calibration and estimation
processes. Such a performance index is defined as follows:
(xt _ xi)2 _ (xt _ x)2

Rx2 = (18)
(Xt - Xi)2
where:
X : represents the value of the variable being assessed (i.e.
the average head, the variance of the heads, etc.).
Xt : represents the true value of that variable.

X i : represents the initial estimate of the value of that
variable.
299
This performance index Rx2 , which is represented in Fig.
6.11, has the following properties:
i) Rx2 = 0, if no improvement results from the calibration
process (i.e. when X=Xi).
ii) Rx2 =1, if the true value has been determined (i.e. when
X=Xt).
iii) Rx2 = 0.75, if the variable X= 0.5*(Xt+Xi), i.e. if it is
halfway between the initial and true values.
iv) Rx2 > 0, whenever some improvement has been obtained.
v) Rx2 < 0, when the calibration process has actually
deteriorated the initial assumptions.
6.11.4.2. Definition of a "global success" index.
The performance index already defined allows us to quantify, in
a non-dimensional way, the behaviour of a procedure with respect
to a single parameter (i.e. either with respect to the average or
the variance, etc.). In order to consider ex p licitly the ability
of the procedure being assessed (Kriging, one-dimensional
interpolation or bi-cubic splines) to improve both the average of
the variable, its variance, the maximum residual and the variance
of the residuals simultaneously, a "global success" index is
defined as simply the percentage of cases when the algorithm
improves all those parameters, over all the tested cases. Hence,
if a procedure has a good (positive) "performance index" with
respect to the average, but a negative "performance index" with
respect to the variance of the residual, this procedure will not
count, as far as the "global success" index is concerned.
300
Rx2
1.0
0.0 X
Fig. 6.11. The performance index Rx2.
Thus, the "performance index" is associated with the
improvement achieved by a certain procedure with respect to a
single parameter, whereas the "global success" index is related
to the consistency of the procedure.
6.11.4.3. Piezometric head estimation results.
Because the piezometric heads estimated using the one-
dimensional interpolation method make use of the raw network
model (based on an initial assumption of the roughnesses) for
determining the minimum head loss paths, which subsequently are
used in the interpolation itself, their head estimates are
different according to which initial roughness estimates have
been considered (i.e. they are different according to which of
the five study cases we are referring to). Because Cases I, II
and III are based on the same initial roughness estimate, their
head estimates obtained with the one-dimensional interpolation
301
are exactly the same, but they differ from the head estimates for
cases IV and V. This does not happen when Kriging or bi-cubic
splines are used as head estimators, where the head estimates are
the same for all the five study cases.
The head estimation results for example A, with the one-
dimensional interpolation results corresponding to cases I, II
and III, and for the three different head estimation procedures,
are presented in Table 6.9., while in Figure 6.12. the residuals
of the initial nodal piezometric heads (example A) are compared
graphically with those produced by the one-dimensional
interpolation method. Both Table 6.9. and Fig. 6.12 show how the
initial residuals are reduced by the head estimation process.
Due -to space constraints, the corresponding detailed results
for networks B, C, D, E and F are omitted although a summary of
the main statistics relating them is presented in Table 6.10.
Note that, from now on, when dealing with the residuals, we are
referring to the absolute value of the residuals, as specified in
the notes corresponding to the respective Table.
Table 6.11 presents a summary of the main statistics associated
with the piezometric heads estimated when the one-dimensional
interpolation procedure is used, for each example and under the
different study cases. Table 6.11 shows that the differences in
the estimated heads under different initial assumptions for pipe
roughnesses and nodal consumption are not significant on average,
though locally (maximum difference) they can be of the order of 1

or 2 meters (as in examples B and D, Table 6.11).
302
Table 6.9. Comparison between estimated piezometric heads and
its residuals, for exam p le A.
NODE TRUE INITIAL INTERPOLATION [2] KRIGING SPLINES

[13
VALUES ESTIMATES RESIDUALS ESTIMATES RESIDUALS ESTIMATES RESIDUALS ESTIMATES RESIDUALS
1 65.000000 65.000000 0.000000 65.000000 0.000000 65.000000 0.000000 65.000000 0.000000

2 63.485885 63.361335 0.124550 63.432510 0.053375 63.740849 -0.254964 66.255523 -2.769638
3 62.765635 62.635732 0.129903 62.763824 0.001811 62.858039 -0.092404 23.035473 39.730162
4 62.818108 62.689977 0.128131 62.818085 0.000023 62.818108 0.000000 62.818108 0.000000
5 62.811073 62.680875 0.130198 62.756592 0.054481 62.849605 -0.038532 38.813053 23.998020
6 61.217718 61.126266 0.091452 61.213028 0.004690 61.868119 -0.650401 35.249859 25.967859
7 61.220484 61.127675 0.092809 61.210495 0.009989 61.833739 -0.613255 51.912286 9.308198
8 61.221935 61.128740 0.093195 61.208405 0.013530 61.862176 -0.640241 45.592621 15.629314
9 60.988101 60.897367 0.090734 60.988068 0.000033 60.988101 0.000000 60.988101 0.000000
10 60.988599 60.898532 0.090067 60.981339 0.007260 60.972607 0.015992 32.488891 28.499708
11 60.988960 60.899024 0.089936 60.988953 0.000007 60.988960 0.000000 60.988960 0.000000
[1] : NODES 1, 4, 9 AND 11 HAVE PIEZOMETRIC HEAD MEASUREMENTS.

[2] : THE VALUES SHOWN IN THIS TABLE CORRESPOND TO THOSE FOR CASE STUDIES 1, II AND III.
Fi g . 6.12. Comparison between the initial residuals and those

obtained after the one-dimensional interpolation
procedure has been applied, for example A.
303
Table 6.10. Summary of the comparison between different
piezometric head estimation procedures.
EXAMPLE PROCEDURE AVERAGE VARIANCE STAN.DEV. MAXIMUM AVERAGE VARIANCE RATIO AVERAGES
RES.[1] RES.[2] RES. [3] EST/TRLE (4]
TRUE 62.136954 1.780894 1.334501

INITIAL 62.040502 1.810364 1.345498 0.130198 0.096452 0.001502 0.998448
A INTERPOL. 62.123754 1.766190 1.328981 0.054481 0,013200 0.000473 0.999788
KRIGING 62.343664 1.602447 1.265878 0.650401 0.209617 0.088996 1.003327
SPLINES 49.376625 229.028789 15.133697 39.730162 13.263900 227.144583 0.794642
TRUE 53.068512 35.740184 5.978309

INITIAL 52.717053 35.932102 5.994339 0.500262 0.351459 0.023382 0.993377
B INTERPOL. 53.068356 35.250739 5.937233 0.204610 0.046578 0.004287 0.999997
KRIGING 53.933735 32.590245 5.708787 2.744054 0.865900 1.714367 1.016304
SPLINES 42.850260 215.365555 14.675338 36.471455 10.671146 173.663753 0,807452
TRUE 62.136954 1.780894 1.54501

INITIAL 62.878059 0.938416 0.968719 1.066052 0.741105 0.149334 1.011927
C INTERPOL. 62.201541 1.641180 1.281085 0.243991 0.062440 0.009006 1.001039
OGRE 62.343664 1.602447 1.265878 0.650401 0.209617 0.068996 1.003327
SPLINES 49.376625 229.028789 15.133697 39.730162 13.263900 227.144583 0.794642
TRLE 23.469569 92.318030 9.6007

INITIAL 22.684429 93.275043 9.657901 1.141310 0.785140 0.009442 0.966546
D INTERCOL. 23.391606 90.859916 9.532047 0.868389 0.082309 0.037531 0.996678
KRIGING 24.378280 154.233956 12.419096 23.095598 1.389888 17.833472 1.038719
SPLINES 23.891365 117.496713 10.839590 13.247301 0.924584 4.667879 1.017972
TRLE 90.580945 6.435581 2.536845

INITIAL 90.565063 6.444084 2.538520 0.098798 0.019310 0.000276 0.999825
E INTERPOL. 90.577184 6.438101 2.537341 0.090795 0.008939 0.000276 0.499958
KRIGING 91.072444 10.506110 3.241622 5.765531 0.571530 1.411968 1.005426
SPLINES 90.781541 8.434977 2.904303 3.721798 0.408525 0.429774 1.002215
TRUE 68.544348 376.086468 19.392949

INITIAL 68.415341 370.880707 19.258263 0.376937 0.165889 0.014416 0.998118
F INTERPCC. 68.528829 374.797523 19.359688 0.289966 0.041197 0.003383 0.999774
KRIGING 69.7681:4 390.903668 19.771284 8.006676 1.406806 2.713941 1.017855
SPLINES 56.852004 847.694714 29.115197 35.664425 12.608786 180.405342 0.829419
NOTES:
===
(1] : MAXIMUM ABSOLUTE VALUE RESIDUAL = MAX.(ABS(TRUE-ESTIMATED)1
[2] : AVERAGE OF THE RESIDUALS = (SUMMATION OF ABSOLUTE VALUES OF THE RESIDUALS)IN
[3] : VARIANCE OF THE ABSOLUTE VALLE OF THE RESIDUALS = ( SSAVR - N I AVRtt2 1 / (N-2)
SSAVR : SUMMATION SQUARES ABSOLUTE VALUES RESIDUALS
AVR : AVERAGE ABSOLUTE VALUE RESIDUALS
N : AUMBER OF DATA POINTS
(4] : RATIO AVERAGE ESTIMATED VALUES/AVERAGE TRUE VALLES
(5] : ALL THE VALUES SHOW IN THIS TABLE FOR THE ONE-DIMENSION-1 INTERPOLATION PROCEDURE
CORRESPOND TO THOSE FOR CASE STUDIES I, II AND III.
304
Table 6.11. Estimated piezometric heads for Cases I, II, III, IV
and V, using the one-dimensional interpolation
procedure.
EXAMPLE C AVERAGE VARIANCE WITH RESPECT TO CASE I

A
RATIO MAXIMUM
AVERAGES DIFFERENCE
62.123754 1.766190
II 62.123754 1.766190 1.000000 0.000000
A III 62.123754 1.766190 1.000000 0.000000
IV 62.108222 1.686334 0.999750 0.142227
V 62.110307 1.798450 0.999784 0.103516
53.068356 35.250739
II 53.068356 35.250739 1.000000 0.000000
III 53.068356 35.250739 1.000000 0.000000
IV 53.102352 33.923882 1.000641 0.545105
V 52.815367 36.779845 0.995233 1.221238
62.201541 1.641180
II 62.201541 1.641180 1.000000 0.000000
III 62.201541 1.641180 1.000000 0.000000
IV 62.149066- 1.653343 0.999156 0.244263
V 62.185013 1.755175 0.999734 0.243790
23.391606 90.859916
II 23.391606 90.859916 1.000000 0.000000
III 23.391606 90.859916 1.000000 0.000000
IV 23.387610 90.408397 0.999829 0.798309
V 23.227227 89.780054 0.992973 1.732925
90.577184 6.438101
II 90.577184 6.438101 1.000000 0.000000
III 90.577184 6.438101 1.000000 0.000000
IV 90.581279 6.431283 1.000045 0.044312
V 90.564589 6.428866 0.999861 0.108506
68.528829 374.797523
II 68.528829 374.797523 1.000000 0.000000
III 68.528829 374.797523 1.000000 0.000000
IV 69.260476 375.312015 1.010676 0.153580
V 69.221385 373.855982 1.010106 0.258896
Figure 6.13 presents a graphical comparison of the head

estimates for network E; only 4 equidistant transverse sections
of the network are shown, joining nodes 1 to 10, 31 to 40, 61 to
70 and 91 to 100 (see Figure 6.9). Figure 6.13 shows how close
305
the one-dimensional interpolation method "follows" the true
piezometric plane. Also evident from Fig. 6.13 is the oscillatory
behaviour of the heads estimated using bi-cubic splines, which
overestimates and underestimates the heads in different parts of
the network. The heads estimated by Kriging seem to overestimate
the true piezometric plane in this example.
Table 6.12. a) presents the performance indices corresponding
to each head estimation procedure: deterministic one-dimensional
interpolation (cases I, II and III), Kriging and bi-cubic
splines. In general, we concentrate our attention on the
behaviour of the estimation processes with respect to 4
parameters: the average of the piezometric head for all the nodes
of the network, its variance, the maximum residual (in absolute
value) and the variance of the absolute value of the residuals.
In Table 6.12. b) the performance index has been averaged across
the six examples, producing a unique "average index", for each
one of the four parameters already mentioned. Table 6.12. c)
shows the frequency in which the estimation procedure improved
the initial head estimates
Tables 6.10. and 6.12. show clearly that, on the whole, the
deterministic one-dimensional interpolation method effectively
improves both the average and the maximum residual of the
estimated heads; both parameters have been improved in 100 % of
the tested cases, as shown in Table 6.12. c).
Table 6.10. shows that, as far as the average of the head
estimates is concerned, this parameter is always within ± 0.4 %
of the true values, for the one-dimensional interpolation
306
•
04
0
rn
.p
-1-)
z
-P o
Ca
rn
1:1
• 0
P
-
C.)
Ca
•ri (1)
• in
fat
cci
o
cr)
fr-4
sq
g
B MW *Pau10Mci lowligr3
sq
3 2 •
p0914 VILZW pa4Du.ing3
-
z
1,1111/1111111111111TIII1IIIIITITIIIIII11
• q sq q
0
;
- 8 • 8 2 --2
61300 14 0 1.4Kuove ld Polou4163
Ivlisiv“Dilivvi.vvvi.mivroviv,11,11111 J ill •
?
v. S
Y; ; SY ; ;
VP DOH 441U/OZeld Pept14153
307
Table 6.12. Performance indices for the different head estimation
procedures, for examples A, B, C, D, E and F.
a) Performance indices.
PERFORMANCE INDEX
EXAMPLE PROCEDURE AVERAGE VARIANCE MAXIMUM VARIANCE
RESID. RESID.
INTERPOL. 0.981 0.751 0.825 0.901
A KRIGING -3.593 -35.666 -23.955 -999.999
SPLINES -999.999 -999.999 -999.999 -999.999
INTERPOL. 1.000 -5.504 0.833 0.966
B KRIGING -5.060 -268.385 -29.088 -999.999
SPLINES -844.286 -999.999 -999.999 -999.999
INTERPOL. 0.992 0.972 0.948 0.996
C KRIGING 0.922 0.955 0.628 0.645
SPLINES -295.459 -999.999 -999.999 -999.999
INTERPOL. 0.990 -1.321 0.421 -14.800
D KRIGING -0.340 -999.999 -408.497 -999.999
SPLINES 0.711 -691.198 -133.725 -999.999
INTERPOL. 0.944 0.912 0.155 0.000
E KRIGING -956.712 -999.999 -999.999--999.999
SPLINES -158.527 -999.999 -999.999 -999.999
INTERPOL. 0.986 0.939 0.408 0.945
F KRIGING -88.996 -7.101 -450.198 -999.999
SPLINES -999.999 -999.999 -999.999 -999.999
NOTE: A VALUE OF -999.999 INDICATES THAT THE INDEX
IS ACTUALLY LESS THAN OR EQUAL TO -999.999
b) Average indices.
AVERAGE INDEX
PROCEDURE AVERAGE VARIANCE MAX.RES VAR.RES
INTERPOL. 0.982 -0.542 0.598 -1.832
KRIGING -175.630 -385.032 -318.518 -833.225
SPLINES -549.593 -948.532 -855.620 -999.999
c) Frequency of improvement.
FREQUENCY OF IMPROVEMENT (%)
PROCEDURE AVERAGE VARIANCE MAX.RES VAR.RES
INTERPOL. 100.000 66.667 100.000 66.667
KRIGING 16.667 16.667 16.667 16.667
SPLINES 16.667 0.000 0.000 0.000
308
procedure; the worst case corresponds to example D. Table 6.12.
b) shows that the average performance index (averaged across all
the examples) is 0.982, which is very close to an optimal index
of 1.0.
As far as the maximum residuals are concerned, their

improvement when using the one-dimensional procedure is reflected
in an average performance index of 0.598 (Table 6.12. b). while
Table 6.10. shows that when compared with Kriging and splines,
these values (produced with the one-dimensional interpolation)
are at least 2.6 times smaller (for example C, 0.243991 against
0.650401 for Kriging).
Even though, by and large, the variance of the estimated
piezometric head and the variance of the absolute value of the
residuals deteriorated with respect to the initial estimates
(those corresponding to the raw model), as shown in Table 6.12.
b), both parameters actually improved in 66 % of the examples
studied, when the one-dimensional interpolation method was
applied. Moreover, Table 6.10 shows that, with the exception of
example D, the variances of the residuals computed using the one-
dimensional interpolation method are about 10 times smaller than
the values obtained with Kriging (example C). Table 6.12. a)
suggests that the deterioration in the variance of the heads and
that of the residuals with respect to the initial conditions,
occurs in networks with a steeper piezometric plane (i.e.
examples B and D); the difference in the results between examples
D and E is significant in this respect.
309
The results for the estimated heads using Kriging and the bi-
cubic splines are, in general, worse than the initial estimates,
althou gh Kriging improved the initial estimates in one example

(network C, Table 6.12 a).
6.11.4.4. Calibration results.
For the six networks specified in Section 6.10 and for the five
cases detailed in Section 6.11.2, the calibration exercise has
been carried out, and the results are tabulated in Appendix D. We
refer to "initial" heads and flows as those heads and flows
computed before the calibration program has been run, i.e. their
values have been determined using the initial estimates for the
pipe roughnesses. Similarly, we refer to "calibrated" C's,
"calibrated" heads and "calibrated" flows, as the corresponding
variables after the calibration program has been run, i.e.
"calibrated" heads and flows refer to the values of heads and
flows determined with the calibrated model, rather than with the
raw model.
The results in Appendix D are ordered in the following way:
a) Calibrated heads results:
For each one of the six examples, and for each study case
considered, we include a summary of the following statistics:
average head, its variance and standard deviation, the maximum
residual, the average absolute value of the residuals, the
variance of the absolute value of the residuals and the ratio
between the estimated and true averages. This represents 6 Tables
(Tables D.1, D.2, D.3, D.4, D.5 and D.6). In order to quantify
310
the performance of the calibration algorithm, for each one of the
above 6 Tables we have prepared another Table with the
corresponding values of the performance index defined in section
6.11.4., for the average calibrated heads, their variance, the
maximum residual and the variance of the absolute value of the
residuals; this generates 6 more Tables (D.7, D.8, D.9, D.10,
D.11 and D.12). Each one of the last 6 tables (with the
performance index) also include an "average index", averaging
across the five study cases (mid part of the tables) and a
summary with the frequency (as a %) of the cases when the
calibration algorithm actually improved the initial heads.
Figure 6.14 presents a graphical comparison of the calibrated
heads for network E (case I), for 4 equidistant transverse
sections, which corresponds with Fig . 6.13 already shown
(comparing the estimated heads).
Tables 6.13., 6.14., 6.15. and 6.16. summarise the behaviour
of the "performance index" for the five cases considered, and its
variation example by example (double entry tables), for the
average calibrated piezometric heads, their variance, the maximum
residual and the variance of the residuals, respectively.
In Tables 6.13. to 6.16. averages have been computed across the
different examples (last columns of the Tables), as well as
averages corresponding to each head estimation procedure, across
all the study cases (last rows of the Tables). Also, a "global
average" has been computed, across all the examples and study
cases, which is shown in the bottom-right corner of the Tables.
311
0 el-I
• rl 0
O $4
O +5 0.1
•,-i 0 0
0 4-) E 0
.0 be 0
a
1 -4
os
r-I
(n
0
N
a)
•,-4
4-)
c.)
> c C4 a) •r 0
•ffi $4 c f:14
(a
- 0 to 0 ...4
O ..-, 4-) ,-i ICJ a)
c.,) 14
0-1 E- W I-I
0
I f,4 0 OA
cn 4-)
a)
(1:i
0
a)
, r $.4 >
cC) , t
r * .0 0
1 g
›-• •1 I •,-4
cr) :1 ,-4
(J $4 I-I
n I
(•) 43
cHEn
o -4.
(i)
cri
U
,.
0
o
P 0
4-,
Ca
•r
U)
Cd
01 Ti
a
S
0 a)
-• 0 .0
/
di
n ,-1
(D
./
vt
2
sp oott 3t4oLuoza4 ostructwo
I,,''' 111111111111TM
et
; ; i it 6
so c•H 31-q•uxizaid PewigiD3
- g
- z
-S
; a a
lIlIllIlIlil
a
II i
LI, a 2
0 n.
sPo pH *t l i owoze tal Palem el.P0
_.
.
,..N
0
111.1110 11111101011110111/101.10110.11III-
3 2 ; ;^
- - -
sP
i 2
o• 3 1UX49
H
-
in
; ;
14 PelcuclUD3
ort
6
312
Table 6.13. Summary of the comparison for the calibrated
piezometric heads: average heads.
PERFORMANCE INDEX: AVERAGE
CASE PROCEDURE A B C 0 E F AVERAGE
INTERPOL. 0.981 1.000 0.992 0.990 0.929 0.985 0.980

I KRIGING -3.584 -5.049 0.924 -0.144 -999.999 -77.643 -180.916
SPLINES -999.999 -999.999 -999.999 0.951 -13.299 -999.999 -668.724
INTERPOL. 0.976 1.000 0.993 0.989 0.933 0.960 0.975

II KRIGING -4.176 -5.122 0.917 -0.195 -999.999 -78.716 -181.215
SPLINES -999.999 -999.999 -999.999 0.936 -10.081 -999.999 -668.190
INTERPOL. 0.982 1.000 0.992 0.992 0.938 0.927 0.972

III KRIGING -4.109 -5.161 0.913 -0.155 -999.999 -80.021 -181.422
SPLINES -999.999 -999.999 -999.999 0.910 -52.001 -999.999 -675.161
INTERPOL. 0.922 0.991 0.999 0.989 1.000 0.866 0.961

IV KRIGING -4.119 -5.065 0.856 -0.183 -999.999 -79.329 -181.307
SPLINES -999.999 -999.999 -999.999 0.910 -37.897 -999.999 -672.831
INTERPOL. 0.925 0.441 0.993 0.878 -0.664 0.726 0.550

V KRIGING -3.413 -2.960 0.913 0.940 -999.999 -75.728 -180.041
SPLINES -999.999 -999.999 -999.999 0.262 -42.912 -999.999 -673.774
IINTERPOL. 0.957 0.886 0.994 0.968 0.627 0.893 0.888 1

AVERAGE KRIGING -3.880 -4.671 0.905 0.053 -999.999 -78.288 -180.980
I SPLINES -999.999 -999.999 -999.999 0.794 -31.238 -999.999 -671.740
NOTE
tilt
A VALUE OF -999.999 INDICATES THAT THE INDEX IS ACTUALLY

LESS THAN OR EQUAL TO -999.999
313
piezometric heads: variance of the heads.
PERFORMANCE INDEX: VARIANCE
CASE PROCEDURE A B C D E F AVERAGE
INTERPOL. 0.758 -5.502 0.973 -1.324 0.886 0.938 -0.545

I KRIGING -35.488 -267.833 0.956 -999.999 -999.999 -4.976 -384.557
SPLINES -999.999 -999.999 -999.999 -774.217 -999.999 -994.999 -962.369
INTERPOL. 0.810 -5.209 0.973 -1.273 0.897 0.903 -0.483

II KRIGING -41.100 -271.631 0.951 -999.999 -999.999 -5.225 -386.167
SPLINES -999.999 -999.999 -999.999 -762.062 -999.999 -999.999 -960.343
INTERPOL. 0.741 -5.592 0.972 -1.485 0.881 0.868 -0.602

III KRI6IN6 -40.483 -273.701 0.949 -999.999 -999.999 -6.201 -386.572
SPLINES -999.999 -999.499 -999.999 -760.159 -999.979 -999.999 -960.026
INTERPOL. -9.810 -88.752 0.974 -2.944 0.758 0.821 -16.492

IV KRI6IN6 -40.622 -268.612 0.946 -999.999 -999.999 -6.041 -385.721
SPLINES -999.999 -999.999 -999.999 -760.038 -999.999 -999.999 -960.006
INTERPOL. 0.653 -29.432 0.998 -5.156 0.646 0.640 -5.275

V KRIGING -34.004 -159.964 0.951 -999.999 -999.999 -5.181 -366.366
SPLINES -999.999 -999.999 -999.999 -845.138 -999.999 -999.999 -974.189
INTERPOL. -1.370 -26.898 0.978 -2.436 0.813 0.834 -4.679

AVERAGE KRIGING -38.340 -248.348 0.951 -999.999 -999.999 -5.525 -381.877
I SPLINES -999.999 -999.999 -949.999 -780.323 -949.999 -999.999 -963.387
NOTE
Mt
314
Table 6.15. Summar y of the comparison for the calibrated
piezometric heads: maximum residual.
PERFORMANCE INDEX: MAXIMUM RESIDUAL
CASE PROCEDURE A B C 0 E F AVERAGE
INTERPOL. 0.829 0.833 0.948 0.420 0.145 0.407 0.597

I KRIGING -23.431 -29.029 0.634 -408.581 -999.999 -450.616 -318.504
SPLINES -999.999 -999.999 -999.999 -133.729 -999.999 -999.999 -855.621
INTERPOL. 0.821 0.833 0.949 0.418 0.160 0.338 0.587

II KRIGING -24.508 -29.165 0.624 -408.782 -999.999 -452.981 -319.135
SPLINES -997.999 -999.999 -999.999 -133.860 -999.999 -999.999 -855.642
INTERPOL. 0.832 0.833 0.947 0.437 0.184 0.238 0.579

III KRIGING -24.346 -29.244 0.619 -408.587 -999.999 -459.157 -320.119
SPLINES -999.999 -999.999 -999.999 -133.739 -999.999 -999.999 -855.622
INTERPOL. -1.257 -1.246 0.967 -0.173 -0.692 -0.491 -0.482

IV KRIGING -24.296 -29.057 0.479 -408.550 -999.999 -458.137 -319.927
SPLINES -999.999 -999.999 -999.999 -133.742 -999.999 -999.999 -855.623
INTERPOL. 0.310 -3.819 0.939 -1.448 -0.508 -0.413 -0.823

V KRIGING -23.672 -24.767 0.466 -385.586 -999.999 -449.428 -313.831
SPLINES -999.999 -999.999 -999.999 -115.761 -999.999 -999.999 -852.626
INTERPOL. 0.307 -0.513 0.950 -0.069 -0.142 0.016 0.092 1

AVERAGE KRIGING -24.051 -28.252 0.564 -404.017 -999.999 -454.064 -318.303
I SPLINES -999.999 -997.999 -999.999 -130.166 -999.999 -999.999 -855.027
NOTE
Mt
315
piezometric heads: variance of the absolute value of
the residuals.
PERFORMANCE INDEX: VARIANCE OF THE RESIDUALS
CASE PROCEDURE A 13 C D E F AVERAGE
INTERPOL. 0.915 0.966 0.996 -14.333 0.014 0.745 -1.749

I KRIGING -999.999 -999.999 0.660 -999.999 -999.999 -999.999 -833.222
SPLINES -999.999 -999.999 -999.999 -999.999 -999.999 -999.999 -999.999
INTERPOL. 0.918 0.969 0.997 -14.135 0.007 0.943 -1.717

II KRIGING -999.999 -999.999 0.644 -999.999 -999.999 -999.999 -833.225
SPLINES -999.999 -999.999 -999.979 -999.999 -999.999 -999.999 -999.999
INTERPOL. 0.935 0.966 0.996 -13.308 0.022 0.925 -1.577

III KRIGING -979.999 -999.999 0.638 -999.999 -999.999 -999.999 -833.226
SPLINES -999.999 -999.999 -999.999 -999.999 -999.999 -999.999 -999.999
INTERPOL. -12.950 -8.383 0.998 -34.845 -3.058 0.762 -9.579

IV KRIGING -999.999 -999.999 0.612 -999.999 -999.999 -999.999 -833.231
SPLINES -999.999 -997.999 -999.999 -999.999 -999.999 -999.999 -999.999
INTERPOL. 0.019 -21.146 0.996 -382.031 -2.101 0.291 -67.329

V KRIGING -999.999 -999.999 0.259 -999.999 -999.999 -999.999 -833.289
SPLINES -999.999 -999.999 -999.999 -999.999 -999.999 -999.999 -999.999
-16.390 1
I
INTERPOL. -2.033 -5.326 0.997 -91.730 -1.023 0.773
AVERAGE KRIGING -999.999 -999.999 0.563 -999.999 -999.999 -999.999 -833.239
SPLINES -999.999 -999.999 -999.999 -999.999 -999.999 -999.999 -999.949
NOTE

LESS THAN OR EQUAL 10 -999.999
The results presented in Tables 6.13 to 6.16 follow the same
pattern as the results corresponding to the estimated piezometric
heads (see Table 6.12, for example), in the sense that the
calibrated heads produced using the one-dimensional interpolation
method are the best, both in the average and in the maximum
residual sense. Indeed Table 6.13. shows clearly that, except in
example E (case V), the calibrated heads computed using the one-
dimensional interpolation scheme always improve the initial (raw
model) heads; according to this Table the addition of flow
measurements into the calibration process is not having a
positive impact on the average of the calibrated heads (Tables
6.14., 6.15. and 6.16. show that the same holds for the head
variance, maximum residual and variance of the residuals as
well). We shall discuss this point later on. Table 6.13 also
shows the impact of worse initial estimates of the roughnesses
and nodal consumptions (Cases IV and V, respectively), the latter
being the more sensitive parameter.
The calibrated heads computed using Kriging and bi-cubic
splines do not improve the initial heads, except in a few
examples (example C for Kriging and D for the splines, as shown
in Table 6.13.).
Example C demonstrates that the calibration algorithm works
well even when the average of the initial roughnesses has been
overestimated ( by +20 % in this example). In this particular
example, the head variances, the maximum residual and the
variance of the residuals are always improved when the one-
dimensional interpolation or Kriging procedures are used.
317
Table 6.14 shows that, on average, the calibration algorithm
does not reduce the initial variance of the piezometric heads
(global performance average index = -4.679), although when the
one-dimensional interpolation procedure is used, improvement is
achieved in approximately 66 % of the cases. It is also apparent
from Table 6.14 that the calibration algorithm failed to improve
the initial variance in examples B and D, which are those with
the steeper piezometric planes. In general, worsening the initial
pipe roughnesses and the nodal consumptions (cases IV and V,
respectively), makes the variance of the heads even worse; the
effect of the worse initial roughnesses seems to be more
important than the initial nodal consumptions. Of the three
procedures tested for head estimation (one-dimensional
interpolation, Kriging and splines), the one-dimensional
interpolation is the procedure which performs best, as far as
reducing the variance of the calibrated heads is concerned.
Table 6.15. shows that, on average, the calibration algorithm
improves the initial maximum residuals when the one-dimensional
interpolation procedure is used to estimate the heads,
particularly when the mean initial roughness and the nodal
consumptions have been well estimated (i.e. in Cases I, II and
III). Example C is a different situation, since the calibration
algorithm improves the maximum residuals of the heads even with
worse initial roughnesses and worse nodal consumption estimates.
From Table 6.15 it is also clear that the application of Kriging
leads to improvements in the variance of the residuals only in
example C, whereas the application of splines for head estimation
318
does not produce any improvement at all.
As far as variance of the residuals (absolute values) is

concerned (Table 6.16), the results of the calibration algorithm
are, on average, worse than the initial estimates, although the
use of the one-dimensional interpolation procedure reduces the
variance in almost 66 % of the cases. As before, in the case of
the maximum residuals, better results are obtained when both the
initial roughnesses and nodal consumptions are well estimated
(cases I, II and III).
In summary, the calibration algorithm improves the initial

average piezometric heads as we/1 as the maximum residual (gLotNal
average performance indices of 0.888 and 0.092, respectively).
The results for the variance of the heads and the variance of the
absolute value of the residuals are, by and large, negative
(global average performance indices of -4.679 and -16.390,
respectively), even though improvement with respect to the
initial heads is achieved in 66 % of the cases, particularly when
better initial estimates of the pipe roughnesses and nodal
consumptions are available.
In looking at the behaviour of the calibrated heads more

broadly, it is possible to find from Tables 6.13 to 6.16 that, in
15 out of 30 cases, the algorithm improved simultaneously the
average heads, their variance, the maximum residual and the
variance of the residuals, i.e. we have a "global success index"
(as defined in section 6.11.4.2) of 50 %. This has happened for
example A (Cases I, II, III, and V), example C (Cases I, II, III,
IV and V), example E (Cases I, II and III) and example F (Cases
319
I, II and III). The corresponding "global success index" when
Kriging was used is 16.67 % (5 cases out of 30) and 0.0 % for
splines.
b) Calibrated flows results:
Following the same sequence as for the heads, Appendix D

includes a set of 12 more tables, with the result summaries
(Tables D.13 to D.18, for examples A to F, respectively) and the
performance indices (Tables D.19 to D.24).
Fig. 6.15. represents the pipe flows for example E, Case I, for
the same 4 transverse sections, as in Figs. 6.13 and 6.14., when

the estimated and calibrated piezometric heads were compared.
Similarly to the case of the calibrated piezometric heads, four

two way tables summarise the behaviour of the performance index
across the five cases and six examples used for the comparison.
Tables 6.17, 6.18, 6.19 and 6.20 refer to the average calibrated
flows, their variance, the maximum residual and the variance of
the residuals, respectively.
i i
I /
- 31
0Z
41.
In
.-1
cd
r1111111IIIIIIII.1/7111.1111WII
4
q 0
r. $1.
3 3 ; 3
I
(sil) 2.01.4 rancuquD9
1
I
E
_g
o q
4
I
I
OM &"14 Pal cOQIIDO
tr
to
—6
o
4
—:
i
(s/I) "%o ll PaIDAIPO
IIIIIIIIIIIII 1W 111.11"111III wiiii..-r-rp.

2 3 3 3 ?pg a
...
(11/0 sieou pormo3
321
flows: average flows.
CASE PROCEDURE A AVERAGE
INTERPOL. -0.004 0.008 -0.023 -19.250 0.453 -0.192 -3.168

I KRIGING -184.733 1.000 -31.332 -0.563 -999.999 0.478 -202.525
SPLINES -72.902 -42.610 -140.271 0.437 -459.738 -999.999 -285.847
INTERPOL. 0.773 0.866 -12.538 -999.999 -4.690 0.958 -169.105

II KRIGING -2.597 0.891 -0.931 -999.999 -999.999 0.995 -333.606
SPLINES -29.806 -1.351 -46.695 -999.999 -333.037 -999.999 -401.815
INTERPOL. 0.661 0.994 0.900 -999.999 1.000 0.968 -165.913

III KRIGING 1.000 0.984 0.996 -999.999 -35.633 -0.804 -172.243
SPLINES -52.206 -0.861 -116.581 -999.999 -37.607 -999.999 -367.876
INTERPOL. 0.124 0.923 0.967 -999.999 0.998 0.980 -166.001

IV KRIGING 0.999 0.999 0.991 -999.999 -34.361 -0.858 -172.038
SPLINES -63.255 -7.648 -5.318 -999.999 -20.644 -999.999 -349.477
INTERPOL. 0.937 0.548 0.988 -164.766 0.909 0.876 -26.751

V KRIGING -999.799 -999.999 0.342 -999.499 -83.941 -0.079 -513.946 1
SPLINES -999.999 -999.999 0.091 -999.999 -82.577 -999.999 -680.414
INTERPOL. 0.498 0.668 -1.941 -636.803 -0.266 0.718 -106.188

AVERAGE KRIGIN6 -237.066 -199.225 -5.987 -800.112 -430.787 -0.054 -278.872
SPLINES -243.633 -210.494 -61.755 -799.912 -186.721 -999.999 -417.086
NOTE
tits
LESS THAN OR EOUAL TO -999.999
322
flows: variance of the flows.
INTERPOL. -0.244 -0.050 0.771 -999.999 0.268 -0.127 -166.563

I KRIGING -177.615 0.920 -29.866 -999.999 -445.618 -999.999 -442.029
SPLINES -40.567 -14.722 -148.540 -999.999 -247.806 -999.999 -408.605
INTERPOL. 0.901 0.894 -77.063 -999.999 0.694 -35.067 -184.940

II KRIGING -142.343 -0.129 -278.370 -999.999 -371.842 -999.979 -465.447
SPLINES -162.665 -117.440 -533.205 -999.999 -426.801 -999.999 -540.018
INTERPOL. 0.877 0.940 -1.074 -999.999 1.000 -33.502 -171.960

III KRIGING -76.926 -0.104 -243.327 -999.949 -463.270 -999.999 -463.938
SPLINES -155.384 -158.095 -535.163 -999.999 -343.325 -999.999 -531.994
INTERPOL. 0.975 0.999 1.000 -999.999 0.998 -39.651 -172.613

IV KRIGING -81.717 -1.145 -6.542 -999.999 -537.835 -999.999 -437.873
SPLINES -142.358 -472.515 -17.798 -999.999 -999.999 -999.999 -605.445
INTERPOL. 0.978 0.956 0.994 -999.999 0.965 -37.067 -172.196

V KRIGING -999.999 -999.999 0.033 -999.999 -242.848 -999.999 -707.135
SPLINES -573.619 -999.999 0.811 -999.999 -368.592 -999.999 -656.899
INTERPOL. 0.697 0.748 -15.074 -999.999 0.785 -29.083 -173.654

AVERAGE KRIGING -295.720 -200.091 -111.615 -999.999 -412.283 -999.999 -503.284
I SPLINES -214.918 -352.554 -246.779 -999.999 -477.304 -999.999 -548.592
NOTE

323
flows: maximum residual.
PERFORMANCE INDEX:MAXIMUMRESIDUAL
CASE PROCEDURE A B C 0 E F I AVERAGE
INTERPOL. -0.136 0.005 -2.638 -52.814 -1.701 -0.091 -9.563

I KRIGING -170.771 -0.154 -6.784 -936.003 -271.042 -135.710 -253.411
SPLINES -67.387 -26.108 -33.078 -364.938 -395.403 -999.999 -314.485
INTERPOL. 0.307 0.022 -2.586 -66.950 -0.834 0.037 -11.667

II KRIGING -16.335 -0.404 -3.860 -999.999 -225.047 -178.260 -237.318
SPLINES -68.515 -19.608 -25.706 -346.077 -999.999 -310.745 -295.108
INTERPOL. 0.616 0,022 -2,870 -39,519 -1,019 -0.054 1 -7.138

111 1B161NB -15.565 0.376 -6.084 -999.999 -999.999 -84.489 -350.960
SPLINES -60.444 -22.171 -23.[15 -s75.ar-fa.141 -4imez
AVEAPB,I. -10.655 0.005 0.798 -54.663 -2.326 -0.067 -11.151
IV KRIGING -15.290 -7.038 -0.007 -999.999 -999.999 -84.659 -351.165
SPLINES -34.568 -69.540 -2.484 -582.368 -999.999 -467.876 -359.472 1
INTERPOL. 0.376 -16.042 0.275 -349.430 -8.595 -3.983 -62.900
KRIGING -108.782 -91.807 0.201 -999.999 -834.416 -87.983 -353.798
SPLINES -252.898 -999.999 -2.654 -575.688 -919.051 -481.720 -538.668
INTERPOL. -1.898 -3.198 -1.404 -112.675 -2.895 -0.832 -20. 484

AVERAGE KRIGING -65.349 -19.805 -3.307 -987.200 -666.101 -114.220 -309.330
'
SPLINES -96.763 -227.525 -17.411 -488.865 -857.280 -545.584 -372.238
NOTE
slat
LESS THAN OR HUAI TO -999.999
324
flows: variance of the absolute value of the
residuals.
CASE PROCEDURE A 8 C D E F I AVERAGE
INTERPOL. 0.278 -0.003 -18.641 -999.999 -2.970 -2.424 -170.627

KRIGING -999.999 0.478 -73.242 -999.999 -999.999 -999.999 -678.793
SPLINES -999.999 -551.471 -999.999 -999.999 -999.999 -999.999 -925.244
INTERPOL. 0.817 0-306 -15.137 -999.999 -2-77.1 -4.9h2 -L49.I9t

KRIGING -165.491 0.221 -24.126 -994.999 -999.999 -999.999 -531.565
SPLINES -999.999 -179.195 -999.999 -999.999 -999.991 -991.111 JW113.5M
INTERPOL. 0.917 0.081 -28.460 -745.078 -1.390 -0.681 -129.102

KRIGING -148.833 0.655 -50.106 -999.999 -999.999 -999399 -533.047
SPLINES -999.999 -397.106 -999.999 -997.999 -999.999 -999.999 -899.517
INTERPOL. -84.968 0.157 0.974 -999.999 -10.028 0.014 -182.308

IV KRIGING -206.090 -64.831 0.109 -999.999 -999.999 -999.999 -545.135
SPLINES -895.254 -999.999 -7.726 -999.999 -999.999 -999.999 -817.163
INTERPOL. 0.662 -129.818 0.605 -999.999 -59.700 -141.496 -221.624

KRIGIN6 -999.999 -999.999 -0.116 -999.999 -999.999 -999.999 -833.352
SPLINES -999.999 -999.999 -11.237 -999.999 -999.999 -999.999 -835.205
INTERPOL. -16.459 -25.855 -12.172 -949.015 -15.292 -29.110 -174.650

AVERAGE KRIGING -504.082 -212.695 -29.496 -999.999 -999.999 -999.999 -624.378
I
SPLINES -979.050 -665.554 -603.792 -999.999 -999.999 -999.999 1 -874.732
1
NOTE
tttt
325
Contrary to what we have seen in the case of the estimated and
calibrated piezometric heads, the behaviour of the calibrated
flows is more variable and needs a more detailed review. Table
6.17 shows that, despite the fact that the global performance
index is negative when the one-dimensional interpolation scheme
is used (i.e. -106.188), most of the cause of this apparent bad
performance is attributable to the example D, which will be
analysed in more detail in the following paragraphs. If we remove
example D, the global performance index becomes -0.0646
indicatin g only a slight deterioration of the initial average
flow.
As it is possible to see from looking at the averages across
all the examples (last column) of Tables 6.17 to 6.20, the values
of the calibrated flows computed using the one-dimensional
interpolation procedure are always better than those
computed using Kriging and splines.
In looking at Tables 6.17 to 6.20 as a whole, it is possible to
find out that in examples A (Cases II, III and V), B (Cases II,
III and IV) and C (Cases IV and V), the average flow, its
variance, the maximum residual and the variance of the residuals
improved simultaneously with respect to the initial estimates.
This represents 7 cases out of 30, i.e. a "global success index"
(as defined in section 6.11.4.2) of 26.7 % . The corresponding
"global success" indices for Kriging and splines are 0.0 % in
both cases.
The behaviour of the performance index for example D deserves
further attention, which has to be complemented with the
326
corresponding Table D.16, from Appendix D. Basically, what is
happening in this particular example is that the initial flows
are quite close to the true values (this is a consequence of
assuming that we know the mean pipe roughnesses which, by the
way, is one of the reasons why we use the raw model flows as
estimates of the true flows); this can be seen from Table D.16,
where the true and average calibrated flows differ only by 5x10-6
(1/s). This implies that a calibrated flow which on average is
onl y 3.9x10 -5 (1/s) different with respect to the true value (as
in Case I) is worse than the initial estimate, producing a large
negative performance index; in practice, a difference of this
magnitude is irrelevant.
On the other hand, if we review the definition of the

performance index [equation(16)], it can be seen that the index
"blows up" when the true and initial values of the variable being
assessed are very close to each other [ i.e. X t -X i Is 0 in
equation (16), and the denominator (X t-Xi ) 2 becomes even closer
to zero]. This indicates that the "performance index" should be
always checked carefully for the case where the initial and true
values are too close.
In general, unless the measurement information corresponds

exactly to the true heads and flows, any algorithm should
deteriorate the initial estimates when starting too close to the
true solution.
As can be seen from Table D.16 (for example D), all the
calibrated flows computed either with the one-dimensional
327
interpolation, Kriging or splines give flows which, on average,
reproduce the true values within a ± 3 % margin, although from
the maximum residual and variance points of view, the one-
dimensional interpolation seems to be much better than the

others.
The incorporation of flow measurements, represented by cases II
and III, with respect to case I (when only head measurements were
considered) did not seem to produce a great impact on the
calibrated flows. For example, as far as the average flows are
concerned (Table 6.17), and leaving example D aside, for the
reasons already explained concerning the ill-conditioning of the
performance index near the true values, Examples C and E did not
improve the calibrated flows when the first set of flow
measurements was considered in the calibration algorithm (i.e.
passing from case I to II in Table 6.17); something similar can
be observed in Table 6.19 with respect to the maximum residuals.
Some improvement can be detected in relation to the variance of
the residuals (Table 6.20) from case I to II, though the
performance index is still negative. When considering the second
set of flow measurements, improvement can be noticed in all the
examples (except D), for the average flows, and some smaller
improvements can be detected in the flow variance and residuals.
This behaviour of the calibration algorithm with respect to the
availability of flow measurements perhaps reveals the need for a
more systematic and rational approach for flow measurement

p lacement, something which has not been studied here.
328
Thus, in summary, as far as the estimation of the true flows is
concerned, which is currently done iteratively at the same time
as the estimation of the roughnesses is carried out, the results
given by the calibration algorithm are not as successful as in
the case of the piezometric heads. This seems to suggest that
perhaps a different flow estimation procedure should be
implemented, instead of the iterative scheme used in the present
work.
c) Calibrated Hazen-Williams roughnesses:
As before, we include in Appendix D 6 tables with the summary
of the results (Tables D.25 to D.30) and another batch of 6
tables with the results corresponding to the performance indices
(Tables D.31 to D.36).
Figure 6.16. presents the behaviour of the calibrated C's, for
example E and case I.
Summarising the Tables presented in Appendix D, Tables 6.21,
6.22, 6.23 and 6.24 condense the results of the performance index
for the average C's, their variance, the maximum residual and the
variance of the absolute value of the residuals.
In Figure 6-16 a) the calibrated C's using the one-dimensional
interpolation scheme are compared with the initial estimates
(shown by the asterisks in Fig. 6-16) and with the true values
(C=140 for all the pipes, represented by the horizontal line).
The values of the roughnesses (C's) corresponding to pipes with
consecutive numbering have been joined with a continuous line in
order to visualise their variability; it should not be forgotten
329
that we are dealing with a discrete variable. Fi g ures 6-16 b) and
c) correspond to the calibrated C's obtained using Kriging and
splines, respectively.
The main statistics associated with Figure 6-16 are summarised

in Table D-29 (case I).
Figure 6-16 a) shows clearly that the C's produced using the
one-dimensional interpolation tend to "follow" the initial
estimates in some pipes. It is also clear from that Figure that
the average calibrated roughnesses are close to the true values,
though their dispersion (variance) is larger than that of the
initial estimates.
For the C's produced using Kriging and splines Fig. 6-16 b) and
c) and Table D-29 show clearly that they are, on average,
underestimating the true roughnesses. What seems to be even
worse with the roughnesses produced with Kriging and splines is
the fact that they are "jumping" from one border to the other of
the "allowable band" [determined by the maximum variation factor
x=0.10, see equation (4)], but they never get close to the true
values. This explains why, in general, the variances of the
absolute value of the residuals are always smaller (when using
Kriging and splines) than those produced using the one-
dimensional interpolation; roughl y speaking, the residuals
(absolute value) produced by using Kriging and splines are
consistently wrong, while those C's produced using the one-
dimensional interpolation are not. In this case we have to look
330
co
to
(I)
0)
r--1
0 1-1
ed
4.)
0 C/1 *el
1-3 C
.•••••n
CD 1-4 E4
CO
0 t
0 n r *
›"‘ 1
E
.• oa
• 0
•
0
0 In 0 In 0 o In o
ID in in a PI PI
5,0 - UO2014 p211DJCIIID3 puo
0 0
-co -co
• ••• •
•
...nn••••
.••nnn•
•0
n111
o 0
0
.s-i
0 t
-o .0
- E E
3
:co
- _o
o co b:
0
4-)
•el •
03
-0 0 _o
a
• el
(NI 0
• r•I
•
filTfl .1,8 0 ;III 1•11111“111111•IIIII 0
o 0 o o In 0 In 0 In
••••
to In * PI PI CNI C-4 ID in in •P •t• Pn
41,
0
S.0 swomiJA -uazoil pa l oicth oo puo lomui 5,0 SU.101. 111A- u o il pa l oig n i oo puo
331
C's: average C's.
INTERPOL. -62.709 -51.879 0.679 -384.064 0.014 -18.400 -86.060

I KRIGING -210.664 0.902 0.633 -999.999 -999.999 -999.999 -534.854
SPINES -773.546 -564.898 0.824 -999.999 -999.999 -999.999 -722.936
INTERPOL. -70.898 -112.730 0.764 -594.284 0.264 -239.267 -169.359

II KRIGING -242.359 0.902 0.633 -999.999 -999.999 -999.999 -540.137
SPLINES -773.546 -564.898 0.824 -999.999 -999.997 -999.999 -722.936
INTERPOL. -65.936 -85.120 0.679 -33.414 -0.543 -306.269 -81.767

III KRIGING -242.769 0.902 0.638 -999.999 -999.999 -997.999 -540.204
SPLINES -773.546 -564.898 0.824 -997.999 -999.999 -999.999 -722.936
INTERPOL. 0.916 -8.867 0.672 -351.862 -39.720 -28.531 -71.232

IV KRIGING -17.207 -12.650 0.634 -999.999 -999.999 -999.999 -504.870
SPLINES -67.284 -47.362 0.786 -999.999 -999.999 -999.999 -518.976
INTERPOL. -30.780 -31.573 0.604 -2.518 -999.999 -183.203 -207.912
1
V KRIGING -230.244 -19.121 0.585 -999.999 -999.999 -999.999 -541.463
SPLINES -773.546 -395.178 0.819 -999.999 -999.999 -999.999 -694.650
1 INTERPOL. -45.881 -58.034 0.680 -273.228 -207.997 -155,134 I -123.266

0.625 -999.999 -999.999 -999.999 -532.306
I I
IAVERAGE KRIGING -188.649 -5.813
SPLINES -432.293 -427.447 0.815 -999.999 -999.999 -999.779 -676.487
NOTE
Itti
332
C's: variance of the C's.
CASE PROCEDURE A B C D E F AVERAGE
INTERPOL. -13.645 -22.196 -0.984 -131.976 -128.856 -48.132 -57.632

I KRIGING -14.007 -12.729 0.673 -585.184 -499.395 -247.525 -226.361
SPLINES -24.279 -10.740 0.892 -641.875 -494.328 -484.876 -275.868
INTERPOL. -12.253 -24.106 -0.099 -161.234 -122.212 -52.456 -62.060

II KRIGING -8.062 -12.729 0.673 -601.478 -465.098 -255.237 -223.655
SPLINES -24.279 -10.740 0.892 -576.977 -472.896 -468.641 -258.773
INTERPOL. -12.957 -18.550 -0.986 -100.423 -125.500 -44.957 -50.562

III KRIGING -7.990 -12.729 0.689 -604.710 -385.901 -354.426 -227.511
SPLINES -24.279 -10.740 0.892 -591.182 -498.385 -497.128 -270.137
INTERPOL. -1.648 -1.292 -0.134 -85.504 -47.949 -14,705 -25.205

IV KRIGING 0.992 -6.325 0.727 -165.657 -107.576 -91.712 -61.592
SPLINES 0.720 0.987 0.867 -158.133 -129.673 -137.148 -70.397
INTERPOL. -145.424 0.927 -3.776 -447.881 -249.762 -159.499 -167.569

Y KRIGING -10.255 -19.584 0.580 -603.746 -413.693 -350.050 -232.791
SPLINES -24.279 -5.757 0.915 -595.019 -524.876 -522.272 -278.548
I
INTERPOL. -37.185 -13.043 -1.196 -185.404 -134.856 -63.950 -72.606
AVERAGE KRIGING -7.865 -12.819 0.668 -512.155 -374.333 -259.790 -194.382
SPLINES -19.279 -7.398 0.892 -512.637 -424.031 -422.013 -230.745
/
NOTE
Mt
333
C's: maximum residual.
PERFORMANCE INDEX: MAXIMUM RESIDUAL
CASE PROCEDURE A 0 C 0 E F I AVERAGE
INTERPOL. -6.704 -6.704 -0.807 -3.863 -4.500 -4.737 -4.552

I KRIGING -6.704 -7.806 -0.066 -4.737 -5.466 -5.466 -5.041
SPLINES -6.704 -6.704 0.800 -4.737 -4.737 -5.466 -4.591
INTERPOL. -6.704 -6.704 -0.798 -4.719 -4.500 -4.737 -4.694

II KRIGING -6.704 -7.806 -0.073 -4.737 -5.466 -5.466 -5.042
SPLINES -6.704 -6.704 0.800 -4.737 -4.737 -5.466 -4.591
INTERPOL. -6.704 -6.704 -0.808 -3.863 -4.500 -4.737 -4.553

III KRIGING -6.704 -7.806 -0.075 -4.737 -5.466 -5.466 -5.042
SPLINES -6.704 -6.704 0.800 -5.466 -5.466 -5.466 -4.834
INTERPOL. 0.119 0.230 -0.062 -1.759 -1,750 -1.552 -0.796

IV KRIGING -0.293 -1.753 0.286 -2.663 -3.250 -3.250 -1.820
SPLINES -0.293 -0.293 0.566 -3.250 -3.250 -3.250 -1.629
I
INTERPOL. -7.806 -7.227 -1.438 -5.466 -4.673 -4.737 -5.224
V KRIGING -6.704 -7.806 -0.457 -4.737 -5.466 -5.466 -5.106
SPLINES -6.704 -6.704 0.800 -4.737 -5.466 -5.466 -4.713
INTERPOL. -5.560 -5.422 -0.782 -3.934 -3.985 -4.100 -3.964

AVERAGE KRIGING -5.422 -6.596 -0.077 -4.322 -5.023 -5.023 -4.410
I SPLINES -5.422 -5.422 0.753 -4.585 -4.731 -5.023 -4.072
NOTE
MI
LESS THAN OR EQUAL TO -999.999 .
334
C's: variance of the absolute value of the residuals.
INTERPOL. -273.295 -200.036 -14.342 -148.606 -195.419 -96.340 -154.673

I KRIGING -90.406 -42.248 -6.634 -28.590 -7.868 -23.788 -33.256
SPLINES -31.983 -34.296 0.931 -11.530 -14.659 -7.782 -16.553
INTERPOL. -261.133 -257.706 -5.482 -174.161 -204.031 -102.722 -167.539

II KRIGING -47.609 -42.248 -6.638 -20.929 -11.185 -28.545 -26.192
SPLINES -31.983 -34.296 0.931 -12.697 -8.501 -8.346 -15.815
I
INTERPOL. -268.537 -190.796 -14.355 -150.180 -200.832 -92.442 -152.857
III KRIGING -47.169 -42.248 -6.104 -17.321 -9.159 -22.785 -24.131
SPLINES -31.983 -34.296 0.931 -11.248 -17.220 -10.806 -17.437
INTERPOL. -0.770 -0.801 -0.278 -31.279 -50.605 -22.705 -17.740

IV KRIGING -1.719 -7.621 -0.354 -11.312 -8.416 -12.585 -7.002
SPLINES -1.047 -1.318 0.755 -9.002 -9.477 -7.886 -4.663
INTERPOL. -282.156 -139.664 -27.710 -127.279 -218.845 -174.175 -161.638

V KRIGING -64.730 -42.482 -12.893 -18.142 -12.487 -21.099 -28.639
SPLINES -31.983 -35.257 0.933 -14.938 -12.089 -8.499 -16.972
INTERPOL. -217.178 -157.801 -12.433 -126.301 -173.947 -97.677 -130.889 1

AVERAGE KRIGING -50.327 -35.369 -6.525 -19.259 -9.823 -21.760 -23.844 I
SPLINES -25.796 -27.893 0.896 -11.883 -12.389 -8.664 -14.288
NOTE
stiR

at the average residuals, which in the case of the one-
dimensional interpolation are about one half of those for the
cases of Kriging and splines. When the variance of the residuals
is computed (not that of their absolute values) the situation
reverses, in sense that the minimum variance is obtained when the
one-dimensional interpolation is used. We have used the absolute
value of the residuals (instead of just the residual) because it
implies a more stringent test.
Tables 6.21 to 6.24 show that, with the exception of the
example C, the calibration algorithm, in general, does not
improve the initially assumed pipe roughnesses. The "global
success" index is now zero (0 %) when the one-dimensional
interpolation and Kriging have been used to estimate the heads,
and 16.7 % when the bi-cubic splines have been used for that
purpose, thus conflicting with previous results in favour of the
one-dimensional interpolation.
As in the case of the calibrated flows, a closer look to the
detailed results is needed, and also a review of the way in which
the whole calibration exercise was planned is required.
In carrying out the calibration exercise, it was assumed that
the initial estimation of pipe roughnesses, considering the pipe
material, pipe age, laboratory data, etc., was good enough to
capture the true average roughness of the network, and we were
expecting only "local" differences in the roughnesses. This means
that, in general, the average initial estimates of the pipe

roughnesses are nearly the same as the true values; the only
exception to this is example C, where the initial C's were
336
overestimated by 20% with respect to the true values. This can be
seen from Tables D.25 to D.30, where the ratio between estimated
and true averages of the C's are shown in the last column of
those Tables; on averaging those values across the different
study cases (I to V), we obtain Table 6.25.
Table 6.25 shows that the initial roughness estimates, except

example C, are always within 0.46 % of the true values. This
implies that again, as in the case of example D for the
calibrated flows, we may find ill-conditioned behaviour in the
performance index (although this time it does not seem to be as
bad as in the case of the flows). On the other hand, as already
noted, the way to improve initial estimates too close to the true
values is by using better estimates and/or more measurements (in
this case both of the heads and the flows, since the calibrated
roughnesses are computed from these two variables), which is not
the present case, specially from the flows point of view. As a
result, some deterioration of the initial C's is not a surprise,
although, on average, Table 6.25 is showing that the calibrated
C's produced when using the one-dimensional interpolation
procedure are better than the originally estimated ones,
including example C (i.e. 1.0115 against 1.0354, last row of
Table 6.25).
Example C illustrates that, on average, the proposed

calibration algorithm improves the initial roughness estimates of
C when starting from poor initial estimates. Table 6.21, for
example C, shows that improvement was obtained when any of the
head estimation procedures was used.
337
Table 6.25. Summary of the ratios between average calibrated
roughnesses and average true roughnesses.
Ratio averages calibrated/true roughnesses

Example Initial Interpolation Kriging Splines
A 1.0046 0.9885 0.9525 0.9123
B 1.0046 0.9726 1.0109 0.9277
C 1.2039 1.1150 1.1249 1.0876
D 0.9997 0.9987 0.9740 0.9749
E 0.9997 0.9955 0.9485 0.9613
F 0.9998 0.9986 0.9495 0.9565
Average 1.0354 1.0115 0.9934 0.9700
Table 6.24 shows that the splines-based results produce the

least variance of the absolute value of the residuals, but this
should be considered similarly to the analysis made of the
results shown in Fig. 6-16, in the sense that a least variance of
the residuals estimate is not decisive when the average, the
variance and the maximum residuals of the estimate are the worst.
The results obtained for example C help us to review the

discussion on the "maximum variation factor" for C [see equation
(4)], used to control the change in the calibrated C's. Indeed,
because the variation factor "x" used in the calibration exercise
was x=0.10 (i.e. 10 % of variation), and because the initial
roughnesses for example C were overestimated by + 20 %, the
algorithm is performing the best it can, within the specified
constraints ( i.e. ± 10 % ) thus, the average estimates are
always halfway between the initial and true roughnesses.
338
To study the effect of the "maximum variation factor" on the
results of the calibration exercise for example C, we ran the
calibration program again, this time allowing a ± 30 % of
variation for C (i.e. x=0.30 in equation (4)] and the results can
be summarised as follows:
- As far as the calibrated piezometric heads are concerned, the

results are nearly the same as with x=0.10, though a little bit
worse.
- The new calibrated flows are much worse than when x=0.10.
- The calibrated roughnesses using the one-dimensional
interpolation procedure are much better from the average C point
of view, although from the variance, maximum residual and from
the variance of the residuals standpoint they are worse.
The value x=0.10 used in the calibration exercise was
determined by successive trials. We found that, in so doing, the
initial estimates lead, by and large, to reasonable calibrated
roughnesses, although this value of x depends on the quality of
the initial roughness estimates; poorer initial estimates would
need a larger x than better initial estimates. Because the
quality of the initial roughness estimates is not known, a trial
and error solution might be necessary to determine "x".
On the other hand, the factor "x" can be used to spot pipes or
zones of the network where problems with the qualit y of the
estimators exist. In fact, when the calibration algorithm gives a
C* equal either to the minimum or maximum allowable (the computer
339
program displays a message in that case), it means, unless there
is some physical evidence to the contrary, that either the
initial C, or the head or flow estimates are wrong, requiring a
closer look at the pipe or zone of the network. This clearly
deserves more attention in the future.
Another alternative to control the variation of C is not
through the initial roughness estimates (( 0 )), as in equation
(4), but as a function of the "current" C within the iterative
calibration procedure.
Eventually, the calibration process should be carried out in
stages, with a variable "maximum variation factor" (x), starting
with a small value and relaxing it as soon as values of C lying
in the border of the allowable band are produced and provided
that the there is some additional evidence to support the new
roughnesses. Some way of linking the value of "x" to the quality
of the estimates (not available at the moment) should be devised.
In summary, we believe that the results for the calibrated C's
are highly dependent on the flow estimation, an aspect that
deserves more attention in the near future. On the other hand,
the way in which the calibration exercise was planned, means that
the performance index for the calibrated C's tends to give a
rather disappointing picture, perhaps worse than the real one.
Example C has been included to show the behaviour of the
calibration algorithm when the calibration exercise is carried
out in another way.
340
6.11.5. Summary of the results.
We summarise the results in terms of the main variables
involved in the whole calibration process: piezometric head
estimates, calibrated nodal piezometric heads, calibrated pipe
flows and calibrated C's.
As a summary of the discussion presented in section 6.11.4,
Table 6.26 recapitulates the results for the "global success"
index, for the different head estimation procedures tested.
a) Piezometric head estimation.
According to the results obtained using different methods for
estimating the unmeasured heads (see Tables 6.10 and 6.12) the
best aPProach seems to be the deterministic one-dimensional

interpolation method. All the evidence available strongly
suggests that the deterministic one-dimensional interpolation
method should be used for head estimation purposes, instead of
Kriging and splines.
Table 6.26. Summary of the "global success" index.
Global success index (%)

Variable
1-D interpolation Kriging Splines
Calibrated heads 50.0 16.7 0. 0

Calibrated flows 26.7 0.0 0.0
Calibrated C's 0.0 0.0 16.7
341
The main reasons for the success of the one-dimensional
interpolation scheme, seems to be attributable to the fact that
it is the only method which explicitly considers the initially
assumed shape of the piezometric plane. Kriging and splines-based
estimation algorithms seem to be smoothing the estimated
piezometric plane, whereas the interpolation method has the
ability to keep the main structure of the initial piezometric
plane (raw model) throughout the whole estimation procedure.
There is an additional advantage of the interpolation method

over Kriging and splines, which is not clearly apparent from the
examples tested. The one-dimensional interpolation method seems
to be more appropriate for handling the head estimation problem
when sharp and local variations of the piezometric plane take
place; this is the case when booster pumps, high minor losses, or
special regulating valves are installed in the network. In these
cases the smoothing effect of both Kriging and splines should
lead to a poorer performance of those methods.
The head interpolation method relies not only on the

piezometric head measurements, as do the other methods, but also
on the initial estimates of the pipe roughness characteristics,
since it uses the raw model to determine the minimum head loss
spanning tree, which allows us to generate the paths used in the
one-dimensional interpolation procedure. This emphasises the need
and the importance of a systematic study of the network
resistance characteristics (according to pipe material, age,
etc.), prior to the head estimation, if a successful network
calibration is to be achieved; this study should be able to
342
determine accurately at least the relative value of the
roughnesses of the different pipes.
As a result of this calibration exercise, it is believed that
the one-dimensional interpolation procedure for the estimation of
the unmeasured heads could be improved a bit further. The
improvement could be achieved by re-estimating the heads once a
first run of the calibration program has been made; the new head
estimates being computed with the pipe head losses of the
calibrated model, not with those of the raw model. The calibrated
model may have a flow (and head loss) distribution slightly
different from that of the raw model, in certain pipes or zones
of the network. Perhaps this can help to reduce both the variance
of the head estimates and that of the residuals.
b) Calibrated piezometric heads.
As expected from what has been noted in section 6.7, and from
the results presented in section 6.11.4, the calibrated
piezometric heads follow closely the estimated piezometric heads
(compare Figures 6.13 and 6.14). Indeed, the use of the one-
dimensional interpolation head estimation method again leads to
the best results, as confirmed in the summary of Table 6.26.
As can be seen from Tables 6.13 to 6.16, the results using
Kriging and splines are clearly unacceptable.
Further improvements should be achieved, as a by-product of the
improvement suggested for the head estimation procedure already
mentioned but, in general, the proposed calibration algorithm
seems to be satisfactory with respect to the piezometric heads.
343
0) Calibrated pipe flows.
It has been already noted that the performance of the
calibration algorithm must be assessed in terms of the agreement
between the true and calibrated roughnesses; however, the
assessment should also extend to the pipe discharges, since
flows, together with piezometric heads, are the main state
variables of the networks.
In general, as far as the calibrated flows are concerned, the
calibration algorithm performs better when using the
deterministic one-dimensional interpolation procedure, as shown
in Table 6.26. This re-affirms previous conclusions concerning
the superiority of the deterministic one-dimensional
interpolation method over Kriging and splines.
As shown in Table 6.26, the "global success" index with respect
to the flows ( 26.6 %) is about one half of that corresponding to
the heads, and this is not good enough for our purposes,
especially because the flows are subsequently used to estimate
the roughnesses. Consequently, it is our belief that the weak
point within the proposed calibration procedure is precisely the
estimation of the unmeasured flows, which is currently done
iteratively, using the information contained in the raw model.
Whilst performing this calibration exercise, an alternative
approach has emerged, which avoids the use of the raw model by
estimating the unmeasured flows from the topology of the network.
The basic idea is to relate the flow measurements with the co-
tree "chords" of the network (i.e. the independent set of flows,
344
as defined in Chapter Two, section 2.3.7, Fig. 2.1), and to
comp ute the rest of the flows (i.e. the dependent ones) directly
from the flow measurements. However, this has yet to be tested
d) Calibrated Hazen-Williams roughnesses.
As summarised in Table 6.26 the success of the proposed
calibration algorithm with respect to the roughnesses is
relatively small, specially in comparison with the success in the
piezometric heads.
We believe that, due to the rationale behind the proposed
calibration algorithm (section 6.7 of this chapter), the best way
to improve the calibrated roughnesses is by previously improving
the estimation of the unmeasured flows. In theory, if we include
more flow measurements, the performance of the calibration
algorithm improves; in the extreme, if the estimation of heads
and flows provides all the true nodal heads and pipe flows, the
calibration algorithm is able to compute all the true pipe
roughnesses. This holds irrespective of the initial rouahnesses
assumed in the raw model.
Additionally, and due to the fact that the head and flow
estimation are carried out as independent processes, some
consistency check should be implemented in the future, in order
to detect situations where both estimates (heads and flows) are
not reasonable for some pipes. For example, if for pipe joining
0 •
nodes 1 and "j", the head estimator Hi > Hj and the flow
estimator gives a flow from node "j" to "i". A decision on which
estimate to adopt could be based on the quality of the estimates
345
(when available). Undoubtedly, this should improve the
performance of the calibration algorithm.
A new computer-based water distribution network static

calibration algorithm has been proposed, which is based on the
estimation of unmeasured nodal piezometric heads and on the
estimation of unmeasured pipe flows.
Three different piezometric head estimation techniques have

been proposed and tested: a geostatistical interpolation
procedure known as Kriging, a deterministic one-dimensional
interpolation scheme and a third procedure based on a bi-cubic
splines fitting.
For the flow estimation, the raw network model, obtained with
initial estimates of the roughnesses, provides the estimates of
the flows for the unmeasured pipes. Because those estimates, when
merged with the flow measurements, are neither balanced nor
compatible with head estimates, an iterative algorithm
approximates those flow estimates to the true ones ( as shown in
Fig. 6.1.), also giving the estimates of the C's.
The results of systematic testing of the calibration algorithm,

with a set of six examples, subjected to five different
conditions, show that:
i) The best piezometric head estimator is the deterministic

one-dimensional interpolation procedure.
346
ii) Following on from i), the calibration algorithm produces
calibrated piezometric heads which follow closely the estimated
head results, and which are superior to those produced using
Kriging and bi-cubic splines.
iii) The calibrated flow results produced using the one-
dimensional interpolation procedure for the estimation of
unmeasured piezometric heads are better than those using Kriging
and splines, but not good enough, and an alternative new approach
for unmeasured pipe flow estimation has been suggested for future
implementation.
In the case that with an alternative pipe flow estimation
procedure an iterative calibration algorithm is still needed,
like the one proposed in this work, the possibility of updating
only the flows (by-passing the updating of the heads) in the
gradient method should be explored. This could save considerable
computer resources, since the head updating is the most
computationally expensive step of the gradient algorithm.
iv) The calibrated roughnesses produced by the calibration
algorithm do not seem to be fully satisfactory. This can be
explained partly because the pipe flow estimation procedure
produces flow estimates which are, in general, not fully
satisfactory. The improvements suggested both for the nodal
piezometric heads and pipe flow estimators should have a positive
impact on the calibrated roughnesses.
347
To summarise, we believe that the proposed calibration
algorithm has shown to be the adequate framework for solving the
the water distribution static calibration problem, but clearly
extra efforts should be made in the future to improve it. The
deterministic one-dimensional interpolation procedure for
unmeasured piezometric head estimation has been identified as the
best piezometric head estimator. An alternative approach for pipe
flow estimation is required for improving both the calibrated
flows and roughnesses.
In our opinion, one of the main advantages of the proposed
calibration algorithm, with the one-dimensional interpolation
head interpolator, lies in the fact that it allows us to
incorporate the physical and engineering knowledge of the
network, existing prior to the calibration, thus breaking up the
inherent ill-definition of the calibration problem. The ill-
definition comes from the fact that, normally, there is
insufficient measurement information to estimate reliably all the
parameters.
Further intensive testing of the proposed algorithm and the
suggested improvements should be carried out, with data from real
networks.
Some effort should be made to estimate the estimation errors
associated with the head estimates produced by the one-
dimensional interpolation scheme, as well as errors in the flow
estimates, in order to be able to produce an estimate of the
error in the calibrated C's.
346
Some way of testing the quality of head and flow estimates,
prior to the calibration process, should be devised, since there
is no point in carrying on with the calibration when the head and
flow estimates are not satisfactory.
The problem of the design of the piezometric head and flow
measurement system has not been studied here, but the need for a
rational measurement placement algorithm has became clear, as a
pre-requisite for a successful calibration.
Also, the problems of bad data detection and replacement need
to be addressed, particularly if the algorithm were to be used in
telemetered networks in the future.

CHAPTER SEVEN
FURTHER EXTENSIONS OF THE GRADIENT METHOD
7.1. Introduction.
In this chapter we deal with some additional extensions to the
gradient method, particularly an extended period simulation
version of the program, which is useful, for example, in
determining the best operating policy from a set of alternatives.
We also introduce in this chapter the latest extension of the
gradient algorithm, which consists of modelling the nodal
consumptions as a linear function of the pressures, thus
providing a more realistic model for the demands. We re-derive
the gradient algorithm for the new conditions, and we briefly
discuss the implications that this approach can have for extended
period simulation and calibration purposes
7.2. Extended Deriod simulation version of the gradient method.
As a natural development of the gradient method, an extended
period simulation algorithm has been implemented, which allows us
to simulate the behaviour of a water distribution system over a
p eriod of 24-48 hours, for example.
As far as the simulation algorithm is concerned, the approach
followed is fairly standard, since it follows approximately the
same pattern as other simulation algorithms [Stephenson (1985),
Coulbeck and Orr (1983), Rao and Bree (1977)]. The main objective
350
has been the study of the performance of the gradient method in
such applications, particularly with some of the implementation
features discussed in Chapter Five.
The program makes use of the output of a demand forecasting
routine, with the total demand of the system discretised hour by
hour, or in shorter periods if necessary. Demand forecasting has
not been covered in the present work. The program computes the
nodal demand, using the standard demand profiles (residential,
commercial, industrial demands, etc.), and demand allocation
data, which must be supplied by the user, together with the
physical and topological data on the network.
The inflows at the reservoirs are also needed as input, with
the corresponding geometrical data of each reservoir:
level/volume relationship parameters, maximum and minimum levels,
initial water levels, etc. The inflows can be input either as
flows per hour or through profile curves associated with a
reference maximum flow; thus, for gravitational sources (constant
flow) only one flow is needed, while for pumped sources an on/off
diagram with the maximum flow is enough.
Starting from a known initial state, represented by some nodal
consumptions and reservoir levels, the problem is to model the
variation of reservoir levels (or storage) with respect to time,
due to the differences between network consumptions and inflows.
This is done by considering that, for each reservoir "i", the
volume variation (denoted by the differential d Volt) in a time
interval "dt" is equal to the difference between inflow ( QINi )
351
and the outflow ( Qi ), i.e.:
d Voli = QINi - Qi (1)

dt
or
d hi QINi - Qi
(2)
dt (d Voli/d hi)
where hi : is the water level at the reservoir "i".
On knowing the geometry of each reservoir, we can relate the
volume with the water level, through a polynomial function as:
Voli = ai hi 2 + bi hi + ci (3)
where ai, bi and ci : are constants, dependent on the geometrical
characteristics of the reservoir .
Hence, from (3), we get:
d Volt
= 2 ai hi + bi (4)
d hi
and the water level variation (2) is given by:
d hi QINi - Qi
= fi(t, hi) (5)
dt 2ai hi + bi
For the particular case of a prismatic reservoir (cylindrical,
rectangular, etc.) the coefficient ai = 0 and bi represents the
transversal section, thus:
Voli = bi h i + c i(6)
and
d Voli = bi
(7)
d hi
which, when introduced into (2), gives:
352
d hi QINi - Qi
- fi(t, hi) (8)
dt bi
The problem now reduces to integrating equation (5), or
equivalentl y , equation (8) for the case of prismatic reservoirs.

Because (5) holds for each reservoir, the problem is actually to
integrate a set of NS differential equations like (5).
On introducing the superscript "k" for identifying the time
step, the water level variation in a finite time interval St can
be approximated as:
h 1 (k + 1) _ hi( k)
fi(t, h i ( k )) (9)
St
where the approximation holds because the right hand side
should correspond to an average value of the function between
time steps "k" and "k+1".
This approximation allows us to implement a simple integration
scheme corresponding to Euler's integration method, which is,
from (9), as follows:

h(k) z h i ( k ) + St f i (t, hi(k)) (10)
Because the analytic expression of the functions fi(t,h) are
not known, the network analysis program is used to compute the
reservoir outflows Qi, thus allowing the values of the
functions to be obtained.
The error in the approximation is relatively small, due to the
fact that the function "f" depends on QINi - Qi and, as a result,
oscillates between positive and negative values, thus allowing
for error cancellation. In a small example with two reservoirs,
353
the error accumulated during a 24 hours simulation has been
computed as 5.8 m 3 , in a total inflow of 12,960 m 3 , for a
simulation with a time step of one hour. The error was defined as
the difference between the total inflow, on the one hand, and the
summation of the total consumption plus the storage variation, on
the other hand. This shows that the error due to the simple
integration scheme is not relevant for most practical purposes,
though it can be reduced even further either by reducing the time
step, or by including some more sophisticated integration scheme.
The main flow chart corresponding to the extended period

simulation program is shown in Figure 7.1.
The present extended period simulation implementation should be
taken as a first step towards a more complete simulation program,
able to provide additional facilities such as: graphical output
of user specified variables, variable time step, a more flexible
nodal demand allocation scheme (possibly including the pressure-
sensitive algorithm presented in the next section), different
kinds of switches (pressure, water level and time switches). etc.

( Start
\ Read data
Initialise
time step
k = 0
Allocate total demand

to the nodes
at time "k"
Compute reservoir inflows

QINi at time "k"
Run network analysis program

(gradient method) and determine
reservoir outflows: Qi
Update reservoir levels: hi

using equation (10)
yes
L'="1
Print results
( Stop)
Fig. 7.1. Main flow chart of extended period simulation program.
355
7.3. Extending the gradient method for Pressure-dependent nodal
demands.
Practically all the authors in networks analysis, with the
exceptions of Lam and Wolla (1972 a and b) and Bhave (1981),
consider the nodal demands as independent of the pressure at the
node. Thus, the nodal demands remain the same regardless of
whether or not the pressure is very low or very high. It is a
well known fact that the actual consumptions in the network are
pressure-dependent. Indeed, most network operators reducc the
pressures in order to cope with extreme demands. One of these
cases is clearly the reduction of nocturnal pressures to reduce
leakages.
Thus, we have to understand the traditional way of modelling
the network demands as valid for a certain pressure range (i.e.
around a nominal service pressure). For design purposes this

assumption may be reasonable, because in that case we are
interested in the behaviour of the system at a certain time
horizon, where we have to assume that the pressure is at least at
the nominal service level. Nevertheless, it seems that the
assumption of a constant demand is not fully valid when studying
the behaviour of the network during a period of, say, 24 hours,
where the pressures at the nodes change over a wide range. The
same might be true for calibration purposes, where we try to
match the performance of the network model with that of the real
network, under at least two extreme situations: very high and
very low flows (i.e.: very low and very high pressures), to cover
the widest spectrum of operational scenarios.
356
Lam and Walla (1972 b) proposed a general model to include the
influencehe pressures on the nodal demands:
qi = ai + bi H i c(11)
where q i : consumption at node "i".
Hi : piezometric heads at node "i".
a i , bi and ci : appropriate constants.
Lam and Wolla's approach is close to the pressure/discharge
relationship for irrigation emitters [see Karmeli et al. (1985)]:
qi = ki Pi x(12)
where
emitter operating pressure.
ki : characteristic coefficient of the emitter, dependent on
the nozzle physical dimensions and form.
: discharge exponent, dependent on the flow regime and
pressure/discharge relationship of the emitter, with
typical values of 0.4-0.6.
In general, all the parameters ai, bi and ci in Lam and Wolla's
relationship, and ki and x in the emitter's case, have to be
determined either by field testing or from manufacturers' data
(emitters).
Bhave (1981) followed a completely different approach,
establishing a discrete relationship between nodal demands and
heads, and satisfying the demands if the heads are greater or
equal to a minimum, the nodal outflow being zero otherwise. Bhave
posed the problem in a constrained non-linear optimisation
format, maximising the outflow. To solve the non-linearity, a
357
recurrent algorithm is used, based on the repetitive solution of
a standard network analysis algorithm, verifying if the head at
each node allows a given consumption to be satisfied.
We believe that the right way to approach the problem is
through an explicit demand/head relationship, similar to that
proposed by Lam and Walla (1972 b), allowing the system itself to
find the equilibrium point between the actual nodal outflow and
pressure.
We are aware that this implies a completely different approach
to the traditional network analysis problem, but it seems a much
more realistic approach, in the sense that it is closer to what
actually happens in a real network. Because of that, we expect
that the results of using this approach in extended period
simulation, or for calibration purposes, would be worthwhile,
though this is something that should be fully investigated in the
near future, using data from reAl networks.
The importance of pressure variations on the demands is fully
recognised by Lonsdale (1985) who, upon using information
published by the National Water Council (1980), showed that,
instead of a theoretical relationship depending on the square
root of the pressure (like in the irrigation emitter case) the
discharges actually follow a slightly quadratic relationship, as
shown in Figure 7.2, which may be well represented as a piecewise
linear function, or via a simple linear relationship in the range
of working pressures (say, between 25-60 m). As correctly pointed
out by Lonsdale (1985), the difference is due to various reasons;
first of all, in the case of leakages an increase in pressure
358
produces an increase in the size of the leaking cracks or holes,
then the square-rooted model is no longer valid. The second
reason deals with a more conceptual topic, connected with the way
water consumption is actually carried out by the customers. There
appears to be two different ways in which consumption occurs:
first, on a volumetric basis, i.e.: we may need to fill a cup of
tea or a bath, which produces a demand that is independent of
pressure (the pressure only affecting the timing, or peak flow,
but not the volume); the second consumption is on a time basis,
i.e.: if we are taking a shower or washing a car, we consume more
water if a high pressure is available in the system. Clearly, for
the whole system, the water consumption increases with pressure.
We shall assume that a linear relationship between pressure and
consumption is enough to describe the real demand/pressure
behaviour of the system around the service pressure region. This
relationship is shown in Figure 7.3, and is such that:
a = as - A22 ( 2-2 s ) (13)
where
g : pressure-dependent nodal demand, a (NN-NS)xl column vector.
9.s : demand at the service pressure (%), a (NN-NS)xl column
vector. See Figure 7.3. This is equal to the demand when
nodal consumption is assumed to be independent of
pressure.
A22 : a diagonal (NN-NS)x(NN-NS) matrix of the sensitivities of
the nodal demands with respect to changes in the
pressures:
A22 = diag (-Oqi/Opi) , i=1,2,...,NN-NS (14)
2 : Pressure (NN-NS)xl column vector, which is computed as:
359
Theoretical
w relationship
0
<
<
w Actual
_J
/ / relationship
/
/
/
/
I i r 1 1
0 20 40 60 80 100
PRESSURE (m)
Fi g . 7.2. The effect of pressure on.leakag es. From: National

Water Council (1980).
q A
Demand
(I/s)
Actual demand
Linear model
—4 \----
Constant demand
nn•
p : Pressure (M)
P s P ma.
At
Service pressure
Fig. 7.3. Relationship between actual demand, pressure, adopted

linear model and constant demand.
360
P = H - Z (15)
where:
H : piezometric head (NN-NS)xl column vector.
z : ground level (NN-NS)xl column vector, which is known from
the topography of the system.
Note that, with this demand/pressure model, g = gs either when
A22=0 (demands are not dependent on pressure) or when the
pressure is equal to the service pressure.
Upon introducing (15) into (13) we get:
2 --= a s - A22 H 4. A22 Z + A22 Ps (16)

or
g = % - A22 H (17)
with
ao = as + A22 Z + A22 Ps (18)
Other demand/pressure linear models are also possible, but all
of them lead to a relationship like (17), only with a different
definition for go (equation 18).
Then, recalling that the necessary conditions for the steady
state flow in the network [see Chapter Three, equation (3)], for
the constant demand case, are:
[ A ll : A l2 I [ I
=
[ -A10 11 0 I
A 21 : 0 H a
and introducing the pressure-dependent nodal demands from (17),
we get:
361
{ A 11 A l2 Q -A10 HO
(19)
A 21 : A 22 I I Hgo
We can now re-derive the gradient algorithm for the new

conditions. In fact, on differentiating (19) and assuming that
A 22 does not depend either on the flows or on the heads, we

get:
N A 11 * 1
Al2 dE
: * dQ = (20)
A 21 : A 22 I d HI d g
[
where, as before in the constant demand case:
u ll Ql1n1-1 51/Q1
Q21n2-1
*21 132/Q2
A 11 * = =A11
aNFIQNprINP-1
1-6NP/QNP
where N = (NPxNP) diagonal matrix of the exponents "n" of the

head loss-flow relationship.
In the right hand side of (20), dE represents the head

imbalance at each branch, and the. ((NN-NS)xl) vector dg,
represents the nodal imbalance which, at an iteration 61 •
1
III
when
convergence has not yet been achieved , can be expressed as:
d Q(i) (21)
E = All") Al2 H(i) +A 10 HO
and
d g = A 21 Q (i) 4- A22 H (i) - go (22)
362
The solution of (20), for the flow and head increments, can be
obtained as:
-1
[dCil N All * 1 A 12
= : * (23)
d H A21 1 A22
I [ : : 1
where A11* is evaluated at Q=Q(i).
The inverse of the partitioned matrix in (23) can be computed

as another block-partitioned matrix:
N A 11 * 1 A 12 [ B ll : B 12 1
1
1 = : (24)
[ A21 : A22 B 21 1 B22
On defining:
G = N A11* (25)
the blocks of the inverse can be computed explicitly, according

to Ay res (1974), we get:
B ll = G-1 + G-1 A l2 ( A22 - A21 G-1 A l2 )-1 A21 G-1
B 22 = (A22 - A21 G-1 Al2)-1

(26)
B 12 = - G-1 A l2 (A22 - A21 G-1 Al2)-1
B 21 = - (A22 - A21 G-1 Al2)- 1 A21 G-1
With the partitioning (24), the s y stem (23) becomes:
d Q =B11 dE + B 12 dg
(27)
d H = B21 d+ B 22 dg
1
which, when introducing equations (21) and (22), becomes:
d Q = B il( A llO i4A l211 (i) + A lollo] + 13 12[ A 210 i)+A22H(i)-go] (28)
and
363
d H = B 21 [A 11 Q( i)+A l2H" )+A 100] B 22 (A21Q (i) +A20 (i) -20] (29)
introducing (26) into (29) and considering:

d Q = Q(i) - Q(i1-1)
(30)
and
d if(i) _ li(1+1) (31)
we obtain, after some algebra:
li (3.+1)= [A 22 - A21 G 1 A l2] -1 f A21 G-1 (A ll Q(i) 4 A 10 HO ) -

- (A21 Q(i) - go) (32)
and from (28):
gi(=
i +1) (I- G-
A ll ) Q(i) - G-1 ( Al2 H(1+1) + A 10 Ho) 1 (33)
Equation (32) can be re-ordered into the traditional format

of a linear system of NN-NS equations in the unknown piezometric
heads:
[A21 G- 1Al2-A22] 11(i+1)= f A21 G-1 (A ll Q(i) A 10 HO ) -

- (A21 Q(i) - go) (34)
As in the case of the gradient method with constant demands

[Chapter Three], equations (34) and (33) have to be solved
recursively, in that order. The structure of the equations is
nearly the same as before, except that now the matrix of
coefficients of the linear system in H has its diagonal modified
by A2 2 (the consumption sensitivity matrix) while in the right
hand side the previous nodal demands g are now replaced by go, as
defined in equation (18) of the present chapter. Of course, when
A22 =0 , S o = Ss and we are in the previous case of demands being

independent from the network pressures.
364
At the present time the computer implementation of this
extended algorithm is under way, and its application for extended
period simulation and calibration purposes has to be studied in
the near future. In order to assess the real advantages of this
approach, data from real systems will be required. The actual
sensitivity of the demands with respect to pressure has to be
determined through field tests, either at the level of the whole
network (total demand versus average pressure) or by grouping

nodes on a geographical basis.
365
CHAPTER EIGHT
SUMMARY, CONCLUSIONS AND FURTHER WORK
8.1. Summary.
The present work has been mainly concerned with the development
and testing of a steady state network analysis algorithm, known
as the gradient method, and with the development of a new

automatic calibration procedure.
The gradient method, due to Todini (1979), has been extended in
the present work to include pumps and pressure reducing and
sustaining valves. Those pressure regulating devices have been
modelled using an original physically-based method, which can be
used as a framework for modelling other regulating valves in the
future. Also, a generalised version of the gradient algorithm
which considers the nodal demands as a linear function of the
pressures has been introduced. An extended period simulation
version of the gradient method has also been developed.
An extensive investigation has been carried out to find the
best solver of the linear systems of equations generated by the
gradient method, including direct and iterative methods.
An automatic calibration algorithm has been proposed which
estimates the true pipe resistance parameters, based on estimates
of the unmeasured nodal piezometric heads and unmeasured pipe
flows. For estimating the unmeasured piezometric heads, three

different methods have been proposed and compared: one based on
366
Kriging, another based on bi-cubic splines and a third based on
an original deterministic one-dimensional interpolation
procedure. For the estimation of the unmeasured flows, the raw
(non-calibrated) network model has been used, based on initial
estimates of the pipe roughnesses. The calibration algorithm has
been tested with a set of 6 examples, subjected to five different

measurement availability scenarios.
8.2. Conclusions.
As far as the development of the gradient method is concerned,

the main conclusions are the following:
a) The original gradient method [Todini (1979), Pilati and
Todini (1984)], has been extended to include pumps and pressure
regulating valves. The extended algorithm is now able to handle
most of the devices normally found in water distribution systems.
A comparison with some of the best existing algorithms shows that
the extended version of the gradient method is stable and
efficient, and it can be recommended for modelling water

distribution systems under the most demanding circumstances: for
example, when the network becomes disconnected, when multiple
pressure regulating valves are working simultaneously, in ill-

conditioned examples, etc.
b) An original ph y sically-based algorithm for modelling
pressure regulating valves has been proposed, implemented and
tested with several examples. The algorithm follows closely the
physical behaviour of the regulating valves, modelling them as
variable-resistance devices, thus being completely different in
367
comparison with most of the existing methods. A comparison with
some examples found in the literature demonstrates that the
proposed method is robust and, indeed, has allowed the detection
of problems in the results published by other authors. The
proposed approach for modelling regulating valves provides the
appropriate framework for including other regulating devices in
the near future. Other advantages of the algorithm are such that
neither topological changes within the model nor re-assembly of
the system of equations are needed, as in other existing methods.
The matrix of coefficients of the linear system of equations
remains symmetric, retaining the advantages of such matrices from
the storage and stability point of view.
c) In order to improve the efficiency of the gradient -method,
which relies heavily on the successive solution of symmetric
positive-definite linear systems of equations, an extensive
investigation of its computational performance has been carried
out, using seven of the most efficient linear solvers. A
multifrontal linear solver has been identified as the fastest
method when enough computer memory is available (routine MA27 of
the Harwell Library); if storage is limited, a preconditioned
(modified) conjugate gradient method is the recommended linear
solver. A good compromise between memory and speed is represented
by the one-way dissection method of George and Liu (1981). The
current microcomputer implementation of the gradient method,
using a 512 Kbytes RAM computer, allows the analysis of networks
of up to 1,200 nodes with the preconditioned conjugate gradient
linear solver, whereas the one-way dissection solver allows a
368
maximum of 1,000 nodes, though is four times faster than the
version with the preconditioned conjugate gradient.
d) As a natural development, an extended period simulation
version of the gradient method has been implemented, providing a
framework for further work in this direction.
e) Finally, an attempt has been made to overcome one of the
main limitations of present network analysis techniques, in the
sense that most of them consider a constant nodal demand,
irrespective of the pressures within the system. Indeed, a
generalised version of the gradient method algorithm has been
proposed, which incorporates explicitly the sensitivity of the
nodal consumptions with respect to the pressures, the constant
demand case being a particular case of this more general
formulation.
As far as the development of an automatic calibration algorithm
is concerned, the main conclusions are as follows:
a) A new automatic calibration procedure has been proposed and

tested using synthetically generated data. The method relies on
piezometric head and flow estimates to compute the pipe
resistance parameters. The proposed calibration algorithm is such
that it allows the incorporation of alternative piezometric head
and flow estimators.
b) Three new alternative methods for estimating the unmeasured

piezometric heads in the network have been proposed, implemented
and compared. These methods are based on the application of
Kriging, bi-cubic splines and an original one-dimensional
369
deterministic interpolation method. The results of such a
comparison indicate that the deterministic one-dimensional
interpolation method is the most appropriate one.
c) The iterative method proposed to estimate the unmeasured
pipe flows does not give results which are as good as the head
estimator and, as a consequence, the calibrated roughnesses
obtained are far from the true ones, indicating that further work
is needed. An alternative direct method, using the topological
concepts of tree and co-tree, has been identified as a possible
solution.
8.3._ Further work.
The present work has meant the development of a number of
algorithms and computer programs, covering the areas of network
analysis, simulation and calibration. As is common in the
development stages, and because the programs have been developed
practically from scratch, there is an immediate need to re-write
all of them having in mind their integration into a single water
distribution network modelling program, using a modular structure
both for programs and data files. Also, future developments in
areas like optimum design, unsteady state flow and optimum
operation should be borne in mind.
We also feel that an effort should be made in order to describe
the basis and main features of the gradient method in the
simplest possible terms, thus widening the spectrum of potential
users, particularly engineers in the water industry. It is a
370
well-known fact that many engineers are not very keen on highly
sophisticated mathematical developments and, unfortunately, this
may be the case for the gradient method and its related
techniques. It is quite natural for the users to avoid the
application of methods which they cannot understand, and a bridge
should be built between the present work and future potential
users.
From the particular point of view of each application, the
following work is envisaged in the near future:
8.2.1. Gradient method for network analysis.
a) Re-write the existing software in modular fashion, so that
any other application can make use of it via a simple subroutine
call. Although the arrival of new powerful microcomputers makes
the memory requirements a less important issue, an effort should
still be made in order to allow the integration of the gradient
method with other applications by using the computational
resources in the most efficient way. A single precision version
of the program should be developed in order to reduce the memory
requirements even further.
b) The generalised version of the gradient method, which allows
the inter-dependence between nodal consumption and pressures to
be considered explicitly, should be fully implemented and tested.
Its convenience for extended period simulation and calibration
should be fully investigated and assessed, ideally with data from
real water distribution systems.
371
6.2.2 Calibration.
a) The calibration program should be integrated in the single

network modelling program we have already referred to. At the
moment the head estimator (the deterministic interpolation
algorithm) is not integrated with the calibration program, and
they should also be linked in the near future.
b) So far, we have relied on the network model itself to
estimate unmeasured flows. This implies that the calibration
procedure has to be carried out iteratively. An explicit flow
estimation procedure would improve both the quality and the
performance of the calibration procedure. One alternative, which
has not been developed in the present work, is to use the graph-
theory concepts of tree and co-tree of the network (see section
2.3.7., chapter 2, Figure 2.1), and to identify the flow
measurements with the co-tree, so that the unmeasured flows (in
the tree) can be determined explicitly, following a similar
approach as that of Hamam and Brameller (1971), though with a
different purpose. This might improve the qualit y of both flow
and roughness estimates.
c) Further improvement in the performance of the piezometric
head estimator (the deterministic one-dimensional interpolator)
is possible. An iterative scheme, which involves re-computing the
head estimates after the calibration algorithm has been rwn was
suggested in Chapter 6 and should be implemented and tested in
the future.
372
d) One of the assumptions made in the proposed calibration
method is that the nodal consumptions are known. This hypothesis
should be reviewed, particularly if the gradient algorithm
incorporating pressure-sensitive demands proves to be relevant
for calibration purposes. Eventually, some practical computer-
assisted procedure for demand estimation and allocation should be
developed. The possibility of an on-line demand estimation
procedure should also be kept in mind.
e) We have suggested that techniques used for piezometric head
estimation should be able to quantify the errors associated with
the estimates. In this context, only Kriging and splines are able
to produce this statistical information. Some way of handling
this within the one-dimensional interpolation method should be
devised. The same should be required for the flow estimates, thus
allowing us to estimate the error in the estimated pipe
resistance parameters.
f) Some validation procedure for the head and flow estimates
should be devised, in order to decide whether the estimates are
good enough to be used in the roughness calibration process,
prior to the calibration itself. Also, some consistency check
should be implemented, to avoid conflict between head and flow
estimates.
g) The subject of the design of the measurement s ystem (heads
and flows) has not been explicitly tackled here. This subject is
relevant, since in fact it is a pre-requisite for a successful
calibration, therefore it has an impact on our final product,
which is a calibrated model of the water distribution system.
373
Thus, it seems that something should be done in the future in
this respect. Kriging provides a neat way of determining the
benefit of an additional measurement within the system ,since the
standard error of an estimate at any ungaged point in the domain
is given, but further research is needed on this aspect.
374
REFERENCES
AJIZ M. A. AND JENNINGS A.: (1984), "A robust incomplete Choleski

conjugate gradient algorithm", International Journal for
Numerical Methods in Engineering, Vol. 20, pp. 949-966.
ALVAREZ C. AND ALBERTOS P.: (1982), " On-line observability

determination as a further result of state estimation
algorithms". IEEE Transactions on Power Apparatus and Systems,
Vol. PAS-101, No.4,April.
ANTHONY G. T. AND COX M. G.: (1987a), "The fitting of extremely

large data sets by bivariate splines", in Mason J.C. and Cox
M.G. (Editors): (1987), "Algorithms for approximation",
Clarendon Press, Oxford, 1987, pp. 5-20. Also NPL Re p ort DITC
84/86.
ANTHONY G. T. AND COX M. G.: (1987b), "The National Physical
Laboratory's data approximation subroutine library", in Mason
J.C. and Cox M.G. (Editors): (1987), "Algorithms for
approximation", Clarendon Press, Oxford, 1987, pp. 669-687.
Also NPL Report DITC 71/86.
ARMSTRONG M. : (1984), "Improving the estimation and modelling of

the variogram", in Verly G. et al. (Editors), "Geostatistics
for natural resources characterization", Part 1, 1-19, 1984 by
D. Reidel Pub. Co.
D'AURIAC A. (1947), "A propos de l'unicite de solution dans les

problemes de reseaux mailles", La Houille Blanche, Mai-Juin
1947, pp. 209-211.
AYRES F.: (1974), "Matrices", Schaum's outline series, McGraw

Hill Book Co.
BARGIELA A.: (1985), "An algorithm for observability

determination in water-system state estimation". IEE
Proceedings, Vol.132,Pt. D,No. 6, November, 1985.
BARLOW J. F. AND MARKLAND E.: (1969), "Computer analysis of pipe

networks", Proc. Inst. Civ. Eng., 43, May-Aug 1969, pp. 249-
259
BARNETT S.: (1979), "Matrix methods for engineers and

scientists", McGraw-Hill, London.
BARKER V. A. (Editor): (1977), "Sparse matrix techniques",

Advanced course held at the Technical University of Denmark,
Copenhagen, August 9-12,1976, Lecture Notes in Mathematics
572, Springer-Verlag, New York.
BATHE K. J.: (1982), "Finite elements procedures in engineering

analysis", Prentice Hall
375
BHAVE P. R.: (1981), "Node flow analysis of water distribution
systems", Transportation Engineering Journal, ASCE, Vol. 107,
No. TE4, pp. 457-467.
BHAVE P. R.: (1988), "Calibrating water distribution network

models", Journal of Environmental Engineering, ASCE, Vol. 114,
No. 1,pp 120-136, February 1988.
BIRKHOFF G. and DIAZ J. B. (1955-1956), "Non-linear network

problems", Quarterly of Applied Mathematics, 13, pp. 431-443.
de BOOR C.: (1978), "A practical guide to splines", Springer

Verlag, New York.
BRANDON T. W. (Editor): (1984), "Water distribution systems",

Water practice manuals, Vol. 4, compiled and published by the
Institution of Water Engineers and Scientists, London, U. K.
BROUSSOLLE F.: (1978), "State estimation in power systems:

detecting bad data through the sparse inverse matrix method",
IEEE Transactions on Power Apparatus and Systems, Vol. PAS-97,
No. 3.
BROYDEN C. G.: (1965), "A class of methods for solving non-linear

simultaneous equations", Math. of Computations, pp. 577-593.
CARPENTIER P., COHEN G. AND HAMAM Y.: (1985), "Water network

equilibrium. Variational approach and comparison of numerical
algorithms", 7th Congress on Operational Research, Bologna,
Italy.
CHANDRASHEKAR M. (1980): "Extended set of components in pipe

networks", Journal of the Hydraulics Division, ASCE, Vol. 106,
No. HY1, January 1980, pp. 133-152.
CHANDRASHEKAR M. AND STEWART K. H.: (1975), "Sparsity oriented

analysis of large pipe networks", Journal of the Hydraulics
Division, ASCE, Vol. 101, No. HY4, pp. 341-355.
CHAUDHRY M. H. AND YEVJEVICH V. (Editors), (1981): "Closed-
conduit flow", Water Resources Publications, Colorado, USA.
CHEN D-Q. AND TEWARSON R. P.: (1986), "Use of incomplete

decomposition of the coefficient matrix in solving linear
equations", International Journal for Numerical Methods in
Engineering, Vol. 23, pp. 199-208.
CHIN K. K., GAY R. K. L., CHINA S. H., CHAN C.H. AND YO S.Y.:
(1978), "Solution of water networks by sparse matrix method",
International Journal for Numerical Methods in Engineering,
Vol. 12, pp. 1261-1277.
CLARKE D. A., MCBEAN E. A. AND AL-NASSRI S. A.: (1981),

"Uncertainties in water distribution systems", Journal of
Hydraulics Division. ASCE, Vol. 107, No. HY10, pp. 1263-1267.
376
CLEMENTS K.A., KRUMPHOLZ G.R. AND DAVIS P.W.: (1983), "Power
system state estimation with measurement deficiency: an
observability/measurement placement algorithm", IEEE
Transactions on Power Apparatus and Systems, Vol. PAS-102,No.
7, July, 1983.
COLLINS M. A.: (1980), "Pitfalls in pipe network analysis

techniques", Transportation Engineering Journal, ASCE, Vol.
106, No. TE5, pp.507-521.
COLLINS M. A., COOPER L. AND KENNINGTON J. L.: (1977), "Analysis

of hydraulic networks using minimization principles", Proc.
17th Congress International Association for Hydraulic
Research, Baden-Baden, Vol. 5, pp. 241-248.
COLLINS M. A., COOPER L., HELGASON R., KENNINGTON J. AND LE

BLANC 1. (1978), "Solving the pipe network analysis problem
using optimization techniques", Management Science, 24, 7, pp.
747-760.
COLLINS M. A., COOPER L. AND KENNINGTON J. L. (1979), "Multiple
operating points in complex pump networks", Journal of the
Hydraulics Division, ASCE, Vol. 105, No. HY3, pp. 229-244.
COLLINS A. G. AND JOHNSON R. L.: (1975), "Finite-element method

for water-distribution networks", Journal American Water Works
Association, July 1975, pp. 385-389.
CORNISH R. J.: (1939-1940), "The analysis of flow in networks of

pipes", Journal of the Institution of Civil Engineers, 13, pp.
147-154.
COULBECK B.: (1984), "A computer program for calibration of water
distribution systems", Journal Institution of Water Engineers
and Scientists, Vol. 38,N0.1,February 1984.
COULBECK B. AND ORR C. H.: (1983), " A network analysis and

simulation program for water distribution systems", Research
Report No. 31, Computer Control of Water Supply, Leicester
Polytechnic,
COX M. G. : (1972), "The numerical evaluation of B-splines", J.

Inst. Maths Applies (1972), 10, pp. 134-149.
COX M. G. : (1982a), "Direct versus iterative methods of solution

for multivariate spline-fitting problems", IMA Journal of
Numerical Analysis (1982), 2, pp. 73-81. Also NPL Report DNACS
37/80, October 1980.
COX M. G. : (1982b), "Practical spline approximation", NPL Report

DITC 1/82, February 1982.
COX M. G. : (1986), "Data approximation and surface fitting", in

Johnson M.W. (Editor): "Workshop on neutron scattering data
anal y sis", Rutherford Appleton Laboratory, Chilton, 13-14
377
March, 1986, Institute of Physics Conference Series Number 81,
Bristol, IOP Publishing Ltd., pp. 103-125. Also NFL Report
DITC 81/86.
COX M. G. : (1987a), Data approximation by splines in one and
two independent variables", in Iserles A. and Powell M.J.D.
(Editors): The state of the art in Numerical Analysis",
Oxford University Press, 1987. Also NFL Report DITC 77/86,
October 1986.
COX M. G. : (1987b), "The NPL data approximation subroutine

library: current and planned facilities", NFL Report DITC
104/87, November 1987.
COX M. G., HARRIS P. M. AND JONES H. M. : (1987), "Strategies for

knot placement in least squares data fitting by splines", NFL
Report DITC 101/87, September, 1987.
COX M. G. AND HAYES J. G.: (1973), "Curve fitting: a guide and

suite of algorithms for the non-specialist user", NFL Report
NAC 26, December 1973.
CRABBE R., CREASEY J. D., FIELD D. B., LACEY R. F. AND STIMSON K.

: (1982), "A guide to water network analysis and the WRC
computer program WATNET, Part I", Technical Report TR 177,
Water Research Centre (WRC), WRC Engineering Centre, Swindon,
U K.
CROSS H. (1936), "Analysis of flow in networks of conduits or

conductors", Bulletin No. 286, University of Illinois,
Engineering Experimental Station, Urbana, Illinois.
DELFINER P. AND DELHOMME J.P. : (1973), "Optimum interpolation by

Kriging", in Davis J.C. and McCullagh M.J. (Editors): "Display
and analysis of spatial data", NATO ASI, Nottingham 1973,
published by John Wiley and Sons, 1975.
DELFINER P. : (1976), "Linear estimation of nonstationary spatial
phenomena", in Guarascio M., David M. and Huijbregts C.
(Editors): "Advanced geostatistics in the mining industry",
pp. 49-68, D. Reidel, Hingham, Mass., 1976.
DELHOMME J. P. : (1978), "Kriging in the hydrosciences", Advances

in water resources, Vol. 1, No. 5, 1978
DELHOMME J. P. : (1979), "Spatial variability and uncertainty in

groundwater flow parameters: a geostatistical approach", Water
Resources Research, Vol. 5, No. 2, pp. 269-280, April, 1979,
DEO N.: (1974), "Graph theory with applications to engineering

and computer science", Prentice Hall.
DILLINGHAM J. H.: (1967), Computer analysis of water distribution

systems, part 2", Water and Sewage Works, 114, 2, pp. 43-45
378
DONACHIE R. P.: (1974), "Digital program for water network
analysis", Journal of the Hydraulics Division, ASCE, Vol. 100,
No. HY3.
DUBIN Ch.: (1947), "Le calcul des róseaux maillês. Contribution A

l'application pratique de la Methode Hardy Cross", La Houille
Blanche, Mai-Juin 1947, pp. 213-227.
DUFF I. S.: (1980), "Recent developments in the solution of large

sparse linear equations", Report No. CSS 80, Computer Science
and Systems Division, AERE Harwell, Oxfordshire.
DUFF I. S. (Editor): (1981), "Sparse matrices and their uses",

Academic Press.
DUFF I. S.: (1983), "Direct methods for solving sparse systems of
linear equations", Report No. CSS 131, Computer Science and
Systems Division, AERE Harwell, Oxfordshire.
DUFF I. S.: (1985), "Comments on the solution of sparse linear

equations", Report No. CSS 190, Computer Science and Systems
Division, AERE Harwell, Oxfordshire.
DUFF I. S., GOULD N. I. M., LESCRENIER M. AND REID J. K.

(1987), "The multifrontal method in a parallel environment",
Report No. CSS 211, Computer Science and Systems Division,
AERE Harwell Laboratory, Oxfordshire.
DUFF I. S. AND REID J. K.: (1976), "A comparison of some methods

for the solution of sparse overdetermined systems of linear
equations", Journal IMA, 17, pp. 276-280.
DUFF I. AND REID J. K.: (1978), "Algorithm 529. Permutation to

block triangular form", ACM Transactions Mathematical
Software, 4, pp. 189-192.
DUFF I. S. AND REID J. K.: (1982 a), "The multifrontal solution

of indefinite sparse symmetric linear systems", Report No. CSS
122, Computer Science and Systems Division", AERE Harwell
Laboratory, Oxfordshire.
DUFF I. S. AND REID J. K.: (1982 b), "MA 27 - A set of Fortran

subroutines for solving sparse symmetric sets of linear
equations", Report No. R-10533, AERE Harwell Laboratory,
Oxfordshire.
DUFF I., ERISMAN A. M. AND REID J. K. : (1986), "Direct methods

for sparse matrices", Oxford University Press, London.
DUPUIS P., ROBERT J. L. AND OUELLET Y.: (1987), "A modified

method for pipe network analysis", Journal of Hydraulic
Research, Vol. 25, No.1, pp. 27-40.
EISENSTAT S. C., GURSKY M. C., SCHULTZ M. H. AND SHERMAN A. H.:

(1982), "Yale sparse matrix package. 1: The symmetric codes",
International Journal of Numerical Methods in Enineerina, 18,
379
pp. 1145-1151.
EPP R. AND FOWLER A. G.: (1970),"Efficient code for steady-state

flows in networks", Journal of the Hydraulics Division, ASCE,
Vol. 96, No. HY1, pp. 43-56.
EPP R. AND FOWLER A. G.: (1971),closure of discussions to Epp and

Fowler: (1970),"Efficient code for steady-state flows in
No. HY7, p. 1131.
EVANS D. J. (Editor): (1983), "Preconditioning methods: theory

and applications", Gordon and Breach, New York.
EVANS D. J. (Editor): (1985), "Sparsity and its applications",

Cambridge University Press.
FEATHERSTONE R. E. : (1983), "Computational methods in the

analysis and design of closed conduit hydraulic systems", in
Novak P. (Editor): "Developments in Hydraulic Engineering-1",
Elsevier Applied Science Publisher.
FIETZ T. R.: (1973), discussion to Wood and Charles (1972):

"Hydraulic network analysis using linear theory", Journal of
the Hydraulics Division, ASCE, Vol. 99, No. HY5, pp. 855-857.
FOX J. A. AND KEECH A. E.: (1975), "Pipe network analysis-A novel
steady state technique", Journal of the Institution of Water
Engineers and Scientists, 29, pp. 183-194.
GAMBOLATTI G. : (1980), "Fast solution to finite element flow

equations by Newton iteration and modified conjugate gradient
method", International Journal for Numerical Methods in
GAMBOLATTI G. AND PERDON A.: (1984), "The conjugate gradients in

subsurface flow and land subsidence modelling" pp. 955-984 in
Bear and Corapcioglu (Editors): "Fundamentals of transport
phenomena in p orous media", Martinus Nijhoff Pub., 1984. NATO
ASI Series E, No. 82.
GAY R. K. L.,CHIN K. K., CHUA S. H., CHAN C. H. AND HO S. Y.:

(1978), "Node reordering algorithms for water network
analysis", International Journal of Numerical Methods in
GELB A. (Editor) : (1974), "Applied optimal estimation", MIT

Press.
GEORGE A.: (1981), "Direct solution of sparse positive definite

systems: some basic ideas and open problems", in Duff I. S.
(Editor), "Sparse matrices and their uses", Academic Press.
GEORGE A., HEATH M. T. AND NG E.: (1984), "Solution of a sparse

underdetermined system of linear equations", SIAM J. Sci. &
Stat, Comput., Vol. 5, No. 4, pp. 988-997.
380
GEORGE A. AND LIU J. W-H.: (1981), "Computer solution of large
sparse positive definite systems", Prentice Hall.
GESSLER J. (1981): "Analysis of pipe networks", Chapter 4 in

Chaudhry and Yevjevich (1981).
GILMAN H. D., GOODMAN M. Y. AND de MOYER R. : (1973),

"Replication modeling for water distribution control",
Journal American Water Works Association, April 1973, pp. 255-
260
GLENFIELD & KENNEDY LTD.: "Pressure and flow control valves",

Publication No. 215/R3.
GOLUB G. H. AND VAN LOAN C. F.: (1983), "Matrix computations",

North Oxford Academic Pub. Co. Ltd., U. K., (The Johns Hopkins
University Press).
HALL M. A.: (1976), "Hydraulic network analysis using

(generalized) geometric programming", Networks, Vol. 6, pp.
105-130.
HAMAM Y. M. AND BRAMELLER A.: (1971), "Hybrid method for the

solution of piping networks", Proceedings Institution of
Electrical Engineers, Vol. 118, No. 11.
HAMBERG D. AND SHAMIR U. : (1988 a), "Schematic models for

distribution systems design. I: Combination concept", Journal
of Water Resources Planning and Management, ASCE, Vol. 114,
No. 2, pp. 129-140, March 1988 .
HAMBERG D. AND SHAMIR U. : (1988 b), "Schematic models for

distribution systems design. II: Continuum approach", Journal
of Water Resources Planning and Management, ASCE, Vol. 114,
No. 2, pp. 141-162, March 1988 .
HARARY F.: (1972), "Graph Theory", Addison Wesley, Third Edition.
HARRIS B. (Editor): (1970), "Graph theory and its applications",

Academic Press.
HARRISON C. : (1988), "Water system demand estimates based on

meter book subtotals", International Symposium on Computer
Modeling of Water Distribution Systems, University of
Kentucky, Lexington, May 12-13, 1988.
HAYES J. G. AND HALLIDAY J.: (1974), "The least-squares fitting

of cubic spline surfaces to general data sets", J. Inst. Maths
Applies (1974), 14, pp. 89-103.
HESTENES M. R. AND STIEFEL E.: (1952), "Methods of conjugate

gradients for solving linear systems", National Bureau of
Standards J. Res., 49, (1952), 409-436.
HOEKSEMA R. J. AND KITANIDIS P. K. : (1984), "An application of
381
the geostatistical approach to the inverse problem in two-
dimensional groundwater modeling ", Water Resources Research,
Vol. 20, No. 7, pp. 1003-1020, July 1984.
HOEKSEMA R. J. AND KITANIDIS P. K. : (1985), "Analysis of the
spatial structure of properties of selected aquifers ", Water
Resources Research, Vol. 21, No. 4, pp. 563-572, April 1985.
HUGHES J. P. AND LETTENMAIER D. P. : (1981), "Data requirements

for Kriging: Estimation and network design", Water Resources
Research, Vol. 17, No. 6., pp. 1641-1650, December 1981.
HUIJBREGTS C.J. : (1973), "Regionalized variables and

quantitative analysis of spatial data", in Davis J.C. and
McCullagh M.J. (Editors): "Display and analysis of spatial
data", NATO ASI, Nottingham 1973, published by John Wiley and
Sons, 1975.
IDA N. AND LORD W. : (1984), "Solution of linear equations for
small computer systems", International Journal of Numerical
Methods in Engineering, Vol. 20, pp.625-641.
IRONS B. M.: (1970) , "A frontal solution program for finite

element analysis", International Journal for Numerical Methods
in Engineering, Vol. 2, pp. 5-32.
ISAACS L. T. AND MILLS K. G.: (1980), "Linear theory methods for

pipe networks", Journal of the Hydraulics Division, ASCE, Vol.
106, No. HY1.
JACKSON C. P. AND ROBINSON P. C. : (1985), "A numerical study of
various algorithms related to the preconditioned conjugate
gradient method", International Journal for Numerical Methods
in Engineering, Vol. 21, pp. 1315-1338.
JENNINGS A. : (1977), "Matrix computations for engineers and
scientists", John Wiley and Sons Ltd.
JENNINGS A.: (1971), "Accelerating the convergence of matrix
iterative processes", J. Inst. Math. APPln., 8, 99-110.
JENNINGS A. AND MALIK G. M.: (1978), "The solution of sparse

linear equations by the conjugate gradient method".
International Journal for Numerical Methods in Engineering,
Vol. 12, pp. 141-158.
JEPPSON R. W.: (1976), "Analysis of flow in pipe networks", Ann
Arbor Science Pub.
JEPPSON R. W. AND DAVIS A. L. (1976): "Pressure reducing valves

in pipe network analysis", Journal of the Hydraulics Division,
ASCE, Vol. 102, No. HY7, July 1976, pp. 987-1001.
JEPPSON R. W. AND TAVALLAEE A.: (1975), "Pumps and reservoirs in

networks by linear theory", Technical Note, Journal of the
Hydraulics Division, ASCE, Vol.101, No. HY3, pp. 576-580.
382
KARMELI D., PERI G. AND TODES M.: (1985), "Irrigation systems
design and operation", Oxford University Press, Cape Town.
KERSHAW D.: (1978), "The incomplete Cholesky-Conjugate gradient

method for the iterative solution of systems of linear
equations", Journal of Computational Physics, 26, pp. 43-65.
KESAVAN H. K. AND CHANDRASHEKAR M.: (1972), "Graph-theoretic

models for pipe network analysis", Journal of the Hydraulics
KING I. O.: (1970), discussion to Epp and Fowler:

(1970),"Efficient code for steady-state flows in networks",
Journal of the Hydraulics Division, ASCE, Vol. 96, No. HY11,
pp. 2379-2380.
KITANIDIS P. K. AND VOMVORIS E. G. : (1983), "A geostatistical

approach to the inverse problem in groundwater modeling
(steady state) and one-dimensional simulations", Water
Resources Research, Vol. 19, No. 3, pp. 677-690, June 1983.
KRUMPHOLZ G. R., CLEMENTS K. A. AND DAVIS P. W.: (1980), "Power

system observability: a practical algorithm using network
topology", IEEE transactions on Power Apparatus and Systems,
Vol. PAS-99, No. 4, July/August 1980.
LAINIOTIS D. G. AND DESPHANTE J. G.: (1974), "Parameter

estimation using splines", in Lainiotis D. G. (Editor):
"Estimation theory", American Elsevier, pp. 101-125.
LAM C. F. AND WOLLA M. L.: (1972a), "Computer analysis of water

distribution systems: Part I - Formulation of equations",
Journal of the Hydraulics Division, ASCE, Vol. 98, No. HY2,
pp. 335-344.
LAM C. F. AND WOLLA M. L.: (1972b), "Computer analysis of water

distribution systems: Part II - Numerical solution", Journal
of the Hydraulics Division, ASCE, Vol. 98, No. HY3, pp. 447-
460.
N
LANSEY K. E.: (1988, ) " A procedure for water distribution
network calibration considering multiple loading conditions",
International Symposium on Computer Modelina of Water
Distribution Systems, University of Kentucky, Lexington, May
12-13, 1988.
LEKANE TH. (1979): "Modele de calcul de l'êcoulement en r6gime

permanent dans un rêseau d'eau maill6", Journal of Hydraulic
Research, 17, (1979), No. 2, pp. 149-163.
LEMIEUX P.: (1972), "Efficient algorithm for distribution

No. HY11, p. 1911.
383
LEWIS J. G. : (1977), "Algorithms for sparse matrix eigenvalue
problems", Computer Science Department, Stanford University
Report STAN-CS-77-595.
LIU K. T. H.: (1969). "The numerical analysis of water supply

networks by digital computers", Proc. 13-th Congress of IAHR,
Vol. 1, Subject A, pp. 25-42.
LIVESLEY R. K. : (1983), "Finite elements: an introduction for

engineers", Cambridge University Press, Cambridge.
LO K.L., ONG P.S., McCOLL R.D., MOFFATT A.M. AND SULLEY J.L.:
(1983 a), "Development of a static state estimator. Part I:
Estimation and bad data suppression", IEEE Transactions on
Power Apparatus and Systems, Vol. PAS-102, No. 8,pp. 2486-
2491, August, 1983.
LO K.L., ONG P.S., McCOLL R.D., MOFFATT A.M. AND SULLEY J.L.:
(1983 b), "Development of a static state estimator. Part II:
Bad data replacement and generation of pseudomeasurements",
IEEE Transactions on Power Apparatus and Systems, Vol. PAS-
102, No. 8, pp. 2492-2500, August, 1983.
LONSDALE P. B.: (1985), "Automatic control of pressure in

distribution systems", The Water Officers' Journal, January,
pp. 34-44.
MARKOWITZ H. M. (1957), "The elimination form of the inverse and

its application to linear programming", Management Science, 3,
pp. 255-269.
de MARSILY G. : (1986), "Quantitative hydrogeology. Groundwater

hydrology for engineers", Academic Press, 1986. Enlarged and
translated from the original french edition entitled
"Hydrogeologie quantitative", Masson, Paris, 1981.
MARTIN D. W. AND PETERS . G.: (1963), "The application of Newton's

method to network analysis by digital computer", Journal
Institution of Water Engineers and Scientists, p. 115.
McCORMICK M.: (1969), discussion to Shamir and Howard (1968):

"Water distribution system analysis", Journal of the
Hydraulics Division, ASCE, Vol. 95, No. HY1, pp.481-483.
McCORMICK M. AND BELLAMY B. E.: (1968), "A computer program for

the analysis of networks of pipes and pumps", J. Instn. Engrs.
Aust., 38, 3, pp. 51-58.
MEIJERINK J. A. AND VAN DER VORST H. A. : (1977), "An iterative

solution method for linear systems of which the coefficient
matrix is symmetric M-matrix", Math. COMP., pp. 148-162.
MILLAR W. (1951), "Some general theorems for non-linear systems

possessing resistance", Phil. Mag. (Ser. 7), Vol. 42, pp.1150-
1160.
384
MONTICELLI A. AND WU F.F.: (1985), "Network observability:
identification of observable islands and measurement
placement" IEEE transactions on Power Apparatus and Systems,
Vol. PAS-104, No. 5, May 1985.
MONTICELLI A. AND WU F.F.: (1985), "Network observability

Theory". IEEE transactions on Power Apparatus and Systems,
Vol. PAS-104, No.5, May 1985.
de MOYER R.: (1973), "A statistical approach to modeling and

control of water distribution systems", Ph.D. Dissertation,
Polytechnic Institute of New York, Brooklin,N.Y.,1973.
de MOYER R. AND HORWITZ L. B.: (1975), "A system approach to

water distribution modeling and control", Lexington Books.
MUNKSGAARD N. : (1980), "Solving sparse symmetric sets of linear

equations by preconditioned conjugate gradients", ACM
Transactions on Mathematical Software, VOl. 6, No. 2, pp. 206-
219.
NAHAVANDI A. N. AND CATANZARO G. V.: (1973), "Matrix methods for

the analysis of hydraulic networks", Journal of the Hydraulics
NATIONAL WATER COUNCIL (1980), "Leakage control policy and

practice", Department of Environment Standing Technical
Committee, Report 26, reprinted by the Water Authorities
Association, London, 1985.
de NEUFVILLE R. AND HESTER J.: (1969), Discussion of " Water

distribution system analysis" of Shamir and Howard (1968),
Journal of the Hydraulics Division, ASCE, Vol. 95, No. HY1.
NIESSNER H. AND REICHERT K.: (1983), "On computing the inverse of

a sparse matrix", International Journal for Numerical Methods
in Engineering, Vol. 19, No. 10, p p . 1513-1526.
O'CONNELL P. E. (Editor): (1977), "Real Time Forecasting and

Control", Proceedings of the First International Workshop held
at the Institute of Hydrology, July 4-29, 1977, Wallinford,
England.
O'NEIL P. V. : (1983), "Advanced engineering mathematics",

Woodsworth Pub. Co.
OGBUOBIRI E. C.: (1970), "Dynamic storage and retrieval in

sparsity p rog ramming", IEEE Trans. on Power ApParatus and
Systems, Vol. PAS-89, NO. 1, pp. 150-155
OGBUOBIRI E. C., TINNEY W. F. AND WALKER J. W. : (1970),

"Sparsity directed decomposition for Gaussian elimination on
matrices", IEEE Trans. on Power Apparatus and Systems, Vol.
PAS-89, NO. 1, pp. 141-150.
385
ORMSBEE L.E. AND WOOD D.J.: (1986), "Explicit pipe network
calibration", Journal of Water Resources Planning and
MAnagement, ASCE, Vol.112, No. 2, April 1986. pp 166-182.
ORMSBEE L. E. AND CHASE D. V.: (1988), "Hydraulic network

calibration using nonlinear programming", international
Symposium on Computer Modeling of Water Distribution Systems,
University of Kentucky, Lexington, May 12-13, 1988.
ORTEGA J. M. AND POOLE W. G.: (1981), "An introduction to

numerical methods for differential equations", Pitman.
OSTERBY 0. AND ZLATEV Z. : (1983), "Direct methods for sparse

matrices", Lecture notes in computer science, 157, Springer-
Verlag.
PARLETT B. N.: (1980), "A new look at the Lanczos algorithm for
solving symmetric systems of linear equations", Linear Algebra
and its Applications, 29, pp. 323-346.
PENROSE R.: (1955), "A generalized inverse for matrices", Proc.

Cambridge Philosophical Society, 51, 406-413 (1955).
PILATI S. and TODINI E . (1984), "La verifica delle reti

idrauliche in pressione", Istituto di Costruzioni Idrauliche,
Facolta D'Ingegneria dell'Universita di Bologna. (in italian).
PISSANETZKY S. : (1984), "Sparse matrix technology", Academic

Press.
PITCHAI R.: (1966), "A model for designing water distribution
pipe networks", Ph. D. Thesis Harvard University, Cambridge,
Mass.
PRESS W. H., FLANNERY B. P., TEUKOLSKY S. A. AND VETTERLING W.
T.: (1986), "Numerical Recipes. The art of scientific
computing", Cambridge University Press.
RAHAL C.M., STERLING M.J.H. : (1981), "Dynamic parameter tunning

for water distribution network models in an extended period
simulation". Proc. Instn. Civ. Engrs.,71,2,March, 1981,pp
151-164.
RAHAL C.M., STERLING M.J.H. AND COULBECK B.: (1980), "Parameter
tunning for simulation models of water distribution networks",
Proc, Instn. Civ. Engrs., September 1980, pp 751-762.
RAO C. R. : (1962), "A note on a generalized inverse of a matrix

with applications to problems in mathematical statistics",
Journal Royal Statistical Society, Series B, Vol. 24, No. 1,
pp. 152-158.
RAO H. S. AND BREE D. W.: (1977), "Extended period simulation of

water systems - Part A", Journal of the Hydraulics Division,
ASCE, Vol. 103, No. HY2, pp. 97-108.
386
RATCLIFFE B.: (1986), The performance and selection of pressure
reducing valves", Technical Report TR 238, WRC Engineering
Centre.
REID J. K.(Editor) : (1971), "Large sparse sets of linear

equations", Academic Press
REID J. K. : (1971), "On the method of conjugate gradients for
the solution of large sparse systems of linear equations", in
Reid J. K. (Ed.): "Large sparse sets of linear equations",
Academic Press
REID J. K. : (1981), "Frontal methods for solving finite-element

systems of linear equations", in Duff I. S. (Editor): "Sparse
matrices and their uses" , Academic Press.
REID J. K.: (1976), "Sparse matrices",in Jacobs D. (Editor), "The

state of the art in Numerical Analysis", Proc. of the
Conference on the State of the Art in Numerical Analysis,
held at the Univ. of York, 12- 15 April 1976, Academic
Press, 1977.
REID J. K.: (1986), "Sparse matrices", Report No. CSS 201,

Computer Science and Systems Division, Harwell Laboratory,
Oxfordshire.
RUHE A. : (1977), "Computation of eigenvalues and eigenvectors",
in Barker V. A. (Editor): "Sparse matrix techniques".
RUUS E. (1981): "Head losses", Chapter 2 in Chaudhry and

Yevjevich (1981).
SAMUELSSON A., WIBERG N. E. AND BERNSPANG L.: (1986), "A study of
the efficiency of iterative methods for linear problems in
structural mechanics", International Journal for Numerical
Methods in Enaineerina, Vol. 22, pp. 209-218.
SARTORETTO F.: (1984), "A modified conjugate gradient method for
the solution of sparse linear systems on microcomputers",
Engineering Software for microcomputers. Proc. of the first
International Conference, Venice, Italy, 2-5 April 1984, pp.
325-333. Riveridge Press, Swansea.
SCOTT D. S. : (1981), "The Lanczos algorithm", in Duff I. S.

(Editor): "Sparse matrices and their applications".
SHAMIR U.: (1973), "Water distribution system analysis", IBM

Research Report RC 4389, Yorktown Heights, N. Y., June 1973.
SHAMIR U.: (1974), "Optimal design and operation of water

distribution systems", Water Resources Research, 10:1:27-36
SHAMIR U. AND HOWARD C. D. D.: (1968), "Water distribution
systems analysis", Journal of the Hydraulics Division, ASCE,
Vol. 94, No. HY1, pp. 219-234.
387
SHERMAN A. H.: (1975), On the efficient solution of sparse
linear and non-linear equations", Report 46, Ph. D. Thesis,
Department of Computer Science, Yale University.
SHIMAUCHI S., TABATA A. AND SHINOMIYA F.: (1985), "Distribution

network control for water supply systems", Instrumentation and
control of water and wastewater treatment and transport
systems, Proc. 4th IAWPRC Workshop, Houston, Denver USA, 1985.
SMITH D. K.: (1982), "Network optimisation practice: a

computational guide", Ellis Horwood, Chichester, England.
SONG C.C.S. AND YANG C.T. (1980), "Minimum stream power: theory",
Journal of the Hydraulics DiViSi011, ASCE, Vol. 106, No. HY9,
pp. 1477-1487.
SONG C.C.S. AND YANG C.T. (1982), "Minimum energy and energy
dissipation rate", Journal of the Hydraulics Division, ASCE,
Vol. 108, No. HY5, pp. 690-706.
"Continuous simulation of flow in pipe

STEI=O?ks"1,)*ActiT8NO: 5, pp. 258-262.
STEWART G. W.: (1973), "Introduction to matrix computations",

Academic Press.
STIMSON K. : (1982a), "A guide to water network analysis and the

WRC computer program WATNET. Part 2 -The interactive version",
Technical Report TR 177 Part 2, WRC Engineering Centre.
STIMSON K. : (1982b), "A guide to water network analysis and the

WRC computer program WATNET. Part 3 - The interactive graphics
version", Technical Report TR 177 Part 3, WRC Engineering
Centre.
STIMSON K. AND BRAMELLER A.: (1981), An integrated mesh-modal

method for steady state water distribution network analysis",
Journal of the Institution of Water Engineers and Scientists,
35, 2, p. 186. (full text paper).
STOER J. AND BULIRSH R. : (1980), "Introduction to Numerical

Analysis", Springer Verlag, New York.
TEWARSON R. P. : (1973), "Sparse matrix", Academic Press.
THOMAS P. D. AND BROWN R. A. : (1987), "LU decomposition of

matrices with augmented dense constraints", International
Journal for Numerical Methods in Engineerina, Vol. 24, PP.
1451-1459.
TINNEY W. F. AND WALKER J. W. : (1967), "Direct solutions of

sparse network equations by optimall y ordered triangular
factorization", Proc. IEEE, Vol. 55, pp. 1801-1809, November
1967.
TODINI E. (1979), "Un metodo del gradiente per la verifica delle
388
reti idrauliche", Bolletino degli Inaegneri della Toscana, No.
11 , pp. 11-14 (in italian).
TODINI E. AND PILATI S. : (1987), "A gradient algorithm for the

analysis of pipe networks", Proceedings International
Conference on Computer Applications for Water SUPPlY and
Distribution, Leicester Polytechnic, 8-10 September.
VAN DER VORST H. A. : (1981), "Iterative solution methods for

certain sparse linear system with a non-symmetric matrix
arising from PDE-problems", J. COMP. Phys., 44, pp. 1-19.
VARGA R. S. : (1962), "Matrix iterative analysis", Prentice Hall.
VOYLES C. F. AND WILKE H. R.: (1962), "Selection of circuit

arrangements for distribution network analysis by the Hardy
Cross method", Journal of the American Water Works
Association, March 1962, pp.285-290.
WALSKI T.: (1983), "Technique for calibrating network models",

Journal of Water Resources Planning and Management, ASCE, Vol.
109, No.4, October, 1983. pp 360-372.
WALSKI T.: (1984), "Analysis of water distribution systems", Van

Nostrand Reinhold.
WALSKI T.: (1986), "Case study: pipe network model calibration

issues", Journal Water Resources Planning and Management,
ASCE, Vol.112, No. 2, April, 1986, pp. 238-249.
WALSKI T. : (1987), discussion to "Explicit pipe network

calibration" by Ormsbee and Wood (1986), Journal Water
Resources and Management, ASCE, Vol. 113, No. 4, July, 1987,
pp. 591-593.
WALSKI T. : (1988a), "Model calibration in Austin, Texas or

Sherlock Olmes meets Hardy Cross", International Symposium on
Computer Modeling of Water Distribution Systems, University of
WALSKI T.: (1988b), "Equipment needs for field data collection in

support of modeling", International Symposium on Computer
Modeling of Water Distribution Systems, University of
WARGA J. (1954), "Determination of steady-state flows and

currents in a network", Instrument Society of America, Vol.
9,Pt. 5, Paper 54-43-4.
WILLIAMS G. N.: (1973), "Enhancements of convergence of pipe

network solutions", Journal of the Hydraulics Division, ASCE,
Vol. 99, No. HY7, pp.1057-1067.
WILSON R. J. : (1985), "Introduction to Graph Theory", Longman.

WOOD D. J.: (1981a), "Algorithms for pipe network analysis and
389
their reliability", Research Re p ort No. 127, Water Resources
Research Institute, University of Kentucky.
WOOD D. J.: (1981b), discussion to Isaacs and Mills (1980):
"Linear theory methods for pipe network analysis", Journal of
iy drauLia r
th el r l'aiom ASCE, March 1981, pp. 384-385.
WOOD D. J. AND CHARLES C. O. A.: (1972), "Hydraulic network
analysis using linear theory", Journal of the Hydraulics
WOOD D. J. AND CHARLES C. 0. A.: (1973), closure to discussions

to Wood and Charles (1972): "Hydraulic network analysis using
linear theory", Journal of the Hydraulics Division, ASCE,
Vol. 99, No. HY11, p. 2129.
WOOD D.J. AND RAYES A. G.: (1981), "Reliability of algorithms

for pipe network analysis", Journal Hydraulics Division, ASCE,
Vol. 107, No. HY10, pp. 1145-1161.
YANG C.T. AND SONG C.C.S. (1979), "Theory of minimum rate of

energy dissipation", Journal of the Hydraulics Division, ASCE,
Vol. 105, No. HY7. .1s2
ZARGHAMEE M. S. : (1971), "Mathematical model for water

distribution systems", Journal of the Hydraulics Division,
ASCE, Vol. 97, No. HY1.
ZISSERMAN A.: (1984), "Program F Matinv for the solution of the

sparse linear system A x = b ", Adv. Eng. Software (G.B.),
Vol.6, No. 1, pp. 2-8.
ZOLLENKOPF K.: (1970), "Bi-factorization: Basic Computational

Algorithm and programming techniques", in Reid J. K. (Editor):
"Lar g e sparse sets of linear equations", Conference on
larg e sets of linear equations, Institute of Mathematics
and its Applications, Oxford, April, 1970.

Salgado R. (1988) PH D Thesis Vol I

Uploaded by

Copyright:

Available Formats

Salgado R. (1988) PH D Thesis Vol I

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Salgado R. (1988) PH D Thesis Vol I

Uploaded by

Copyright:

Available Formats

UNIVERSITY OF NEWCASTLE-UPON-TYNE

CIVIL ENGINEERING DEPARTMENT

COMPUTER MODELLING OF WATER SUPPLY DISTRIBUTION NETWORKS

USING THE GRADIENT METHOD.

NEWCASTLE UNIVERSITY LIBRARY

Thesis submitted in fulfilment of the requirements for the

degree of Doctor of Philosophy in Civil Engineering.

I would like to acknowledge the constant and enthusiastic

support given to this work by Professor P. E. O'Connell, Head of

the Water Resources Division of the Civil Engineering Department,

University of Newcastle-upon-Tyne (United Kingdom). He supervised

permanent encouragement has been invaluable through all this

I am indebted to Professor E. Todini, Head of the Institute of

Hydraulic Constructions of the University of Bologna (Italy). He

its further improvement.

It has been a privilege and a pleasure to work with Professors

O'Connell and Todini.

Some institutions made possible my stay in Newcastle

University and I am most grateful to them: The Overseas

Development Administration provided the funds for a three years

scholarship; The British Council managed this scholarship and

also provided extra support for attending Conferences and

courses. I am particularly grateful for the support provided by

the chilean office of The British Council and of that received

from the staff of Newcastle Regional Office. The University of La

Serena (Chile), granted me with a leave of absence which I fully

appreciate and my colleagues at the Civil Engineering Department

have shared the extra burden due to my absence.

Newcastle-upon-Tyne provided a stimulating environment and

to use his Kriging suite of programs.

My famil y and friends from Chile constantly supported me and I

truly appreciate their encouragement.

A water distribution network analysis method known as the

gradient method, due to Todini (1979), has been generalised and

subjected to an extensive program of testing and evaluation. The

method has been extended to include pumps and some pressure

regulating valves, and an original physically-based method has

been proposed for modelling the latter devices. Also, a

generalised version of the algorithm which considers the nodal

demands as a linear function of the pressures has been

introduced. The gradient method has been tested with numerous

examples, showing remarkable robustness and convergence speed

when compared with the most efficient traditional methods. The

gradient algorithm does not break down with disconnected

The performance of the gradient algorithm when using seven

different linear solvers, including direct and iterative methods,

has been investigated. A multifrontal linear solver has been

identified as the most efficient method when enough computer

memory is available (routine MA27 of the Harwell Library); if

storage is limited, a preconditioned (modified) conjugate

gradient method is the recommended linear solver. A good

compromise between memory and speed is represented by the one-way

An automatic calibration algorithm has been proposed which

estimates the true pipe resistance parameters, based on the

estimation of the unmeasured piezometric heads and unmeasured

flows. For the piezometric head estimation, three different

another based on bi-cubic splines and a third based on an

original deterministic one-dimensional interpolation procedure.

The latter producing the closest estimates with respect to the