A Regression Technique With

Dynamic Parameter Selection

for Phase-Behavior Matching
Rajeev K. Agarwal, SPE, Vau·Kun U, SPE, and Long Nghlem, SPE, Computer Modelling Group

Summary. The major problem in phase-behavior matching with a cubic equation of state (EOS) is the selection of regression param-
eters. Many parameters can be selected as the best set of parameters; therefore, a dynamic parameter-solution scheme is desired to
avoid tedious and time-consuming trial-and-error regression runs. This paper proposes a regression technique in which the most signif-
icant parameters are selected from a large set of parameters during the regression process. This technique reduces the regression effort
considerably and alleviates the problem associated with a priori selection of regression parameters. The technique's success is demon-
strated by matching experimental data for a light oil and a gas condensate.

Introduction
Cubic EOS's generally do not predict laboratory data of oil/gas Several strategies are available that exploit that structure. Coats and
mixtures accurately without tuning of the EOS parameters. I The Smart I used a modified linear programming least-squares al-
practice often has been to adjust the properties of the components gorithm, while Watson and Lees used a modification of the
(usually the heavy fractions)-e.g., Pc. Tc, and ",-to fit the ex- Levenberg-Marquardt algorithm9 to solve the nonlinear least-
perimental data. squares problem. In this paper, a modification of the adaptive least-
The objective function in the regression involves the solution of squares algorithm of Dennis et al. 2 is used. The present method
complex nonlinear equations, such as flash and saturation-pressure departs from Dennis et al.'s approach in the use of different non-
calculations. A robust minimization method is therefore required linear optimization concepts for selecting the search direction and
for rapid convergence to the minimum. In this paper, a modification the step size. Details of the regression method are described in the
of Dennis et al. 's2 adaptive least-squares algorithm is used. The Appendix.
modification involves the use of some other nonlinear optimization
concepts on direction and step-size selection. 3 Calculation of Jacobian
Dynamic selection of the most meaningful regression parameters
It was found early in the investigation that the key to an efficient
from a larger set of variables is described. This feature is extremely
algorithm would be a fast and accurate estimation of the Jacobian
useful in EOS fitting because it alleviates the problem of deciding
matrix, J =0 rlox. In the Appendix, it is shown that matrix J is
a priori the best regression variables, which is extremely difficult.
also used to determine the Hessian matrix, V2f
It should be stressed that the regression procedure will not correct
The Jacobian J is obtained through numerical differentiation.
the deficiencies of the EOS used and that the EOS predictive capa-
Although the EOS can be differentiated analytically with respect
bility depends entirely on the type and accuracy of the data used
to the regression parameters Xj, the residuals, ri, are extremely
in the regression. For predictive purposes, attempts should be made
complex functions of these parameters (e.g., the GOR of the last
to ensure that the tuned parameters are within reasonable physical
step in a differential liberation) and are more suitable to numerical
limits.
differentiation.
All calculations in this paper are performed with the Peng-
The calculations of ri also involve the iterative solution of non-
Robinson EOS, 4 although the scheme is general and can be ap-
linear systems of equations (e.g., flash and saturation-pressure cal-
plied to any EOS.
culations) where the results are available only to some accuracy
fj. Thus, in the calculation of J through numerical differentiation,
Regre••lon Method
the perturbations of the independent variables' Xj must not be
The implementation of the dynamic-parameter-selection strategy masked by the convergence tolerances Ei or the round-off errors
for tuning the EOS is a nonlinear optimization problem. The goal associated with the calculations of rio It has been found that a per-
is to minimize a weighted sum of squares turbation of 1% of the independent variables Xj is adequate to
nm compute J through numerical differentiation.
f(x)= E rl, ................................... (1)
i=1 Selection of Regre••lon Parameter.
x
where =(Xb x2' .. 'Xn )T = vector of nr regression parameters Given a global set of np regression parameters, the method selects
an active subset of nr parameters with which regression will be
and nm = number of measurements to be fitted. Usually, nm > nr.
The elements of r(x) denoted by ri(x) are nonlinear functions performed. The global set of regression parameters is supplied by
of x; i.e., the user and includes all or some of the following: Pci. Tci , Vci •
VI, dij , and 8.
ri=wi£ei(x)-e;"1/ei, i=l. .. n m , . . . . . . . . . . . . . . . . . . . . (2) The interaction coefficients between hydrocarbons are estimated
from the following equation 10:
where ei(x)=values calculated with the EOS, ei=corresponding
experimental measurements, and Wi = weight associated with the d ij =I-[2VJVcf/(Vc;'+VcJ)]8 . ..................... (3)
ith experimental data point. Note that the differences between the The volume translation techniques of Peneloux et al. II are used
experimental and calculated values are normalized by dividing by to correct the molar volume as follows:
ei in Eq. 2. This brings the magnitudes of the ri to comparable
values. nc
The minimization off(x) may be solved by various methods for V'=V- E y;Vf, ................................ (4)
nonlinear parameter estimation5 and for nonlinear optimization. 6,7 i=1
The general-purpose optimization methods, however, do not take
advantage of the structure of the nonlinear least-squares problem. where V=molar volume from the EOS, V' = corrected molar
volume, and Yi=mole fraction of Component i. Eq. 4 has been
Copyright 1990 Society of Petroleum Engineers proved to improve density predictions of the EOS.

SPE Reservoir Engineering, February 1990 llS

 TABLE 1-LIGHT-OIL COMPOSITION (mol%)
Enter Initial values of the np xI with the respective bounds Xhmin and ~......
CO 2 3.03
CaiculateXj. (~-~.min)/(x.mox -~.min) N2 0.33
C1 41.33
C2 8.89
C3 5.95
iC 4 1.34
nC 4 2.78
iC s 1.22
nC s 1.44
C8+ 33.69
API gravity of C 8 + at 15°C, °API 35.3
Specific gravity of C 8 + 0.8478
Molecular weight of C 8 + 202

5sn r snp • Fig. 1 shows a flow chart of the dynamic selection of

regression variables.

No Exampl••
The Peng-Robinson EOS4 is used in alt calculations.

Light-Oil System. Table I gives the composition of a light reservoir

fluid. This fluid is modeled by the seven pseudocomponents shown
No
in Table 2. The experimental datum to be matched is the
pressure/composition diagram of light-oil/C0 2 mixtures. The 25
potential regression parameters for this system are acentric factors,
"'i; critical pressures, Pci; critical temperatures, Tci ; CO 2 /hydro-
carbon interaction coefficients, deo 2 ,i; and power for hydro-
carbonlhydrocarbon interaction coefficients, 8 (Eq. 3), where the
Fig. 1-Flow chart for selecting the active regre88lon pa- subscript i refers to all the hydrocarbon components of the fluid.
rameters. Table 3 gives properties of the original fluid and the regressed
fluid. Fig. 2 shows the experimental pressure/composition data,
as well as the calculated values obtained from the tuned and un-
The regression parameters Xj are scaled by the upper bound tuned EOS. It can be seen that the fit obtained by regression is ex-
ij,max and lower bound ij,min of the corresponding parameter i j cellent.
from the above list so that they always lie between zero and unity; Table 3 also shows the five variables chosen by the regression
i.e., procedure: the Tc of the two lightest hydrocarbons, the Tc of the
Xj=(ij-ij,min)/(ij,max-ij,min)' ..................... (5) heaviest hydrocarbon, the interaction coefficient between CO 2 and
C l3 through C 20 (Component 6), and the hydrocarbon interaction
Given initial estimates x!> ,j = 1 .. . np ' the regression algorithm coefficient power in Eq. 3. This example indicates that the five
first sorts them in the de~ending order of loj/oxj I. From these regression parameters selected among the 25 candidates by the
np parameters, the first nr parameters that correspond to the nr method are very adequate to match the experimental data.
largest loj/oxj I are chosen for regression. nr is supplied by the
user. Gas-Condensate System. The composition of the reservoir fluid
The regression starts with these nr parameters, and if at any time used in the Third SPE Comparative Solution project 12 is given in
during the regression loj/oxkl becomes less than Iril, i= l. . .n m , Table 4. The fluid contains a large percentage of methane and is
the variable Xl is dropped from the regression set and is replaced represented by the seven pseudocomponents shown in Table 5. The
by the next Xj from the initial. sO£!ed list. Indeed, txx:ause all Xj experimental data to be matched consist of saturation-pressure/com-
are scaled between zero and uruty, If loj/oxkl < Ir;l for 1= I .. . nm, position data and swelling-factor/composition data.
Xk will likely go out of bounds at the subsequent iteration and
therefore should be dropped from the active parameter list. Another
condition where Xl is dropped is when it tries to go out of bounds
TABLE 3-PROPERTIES OF ORIGINAL AND
after consecutive iterations.
REGRESSED LIGHT OIL
The above technique allows the regression on nr variables among
np variables at a time. It also selects the most sensitive parameters Pc (atm) Tc(K)
Component
to operate on and discards these as they fail to reduce the objective
Number Component Initial Final Initial Final
functionj(x). Our experience indicates that for most regression
problems involving reservoir fluids, it is desirable to have 1 CO 2 72.80 72.80 304.2 304.2
2 C 1 ,N 2 45.31 45.31 190.1 176.0'
3 C2 , C3 45.64 45.64 332.9 383.0'
TABLE 2-MODEL FLUID FOR LIGHT OIL 4 C4 , Cs 35.54 35.54 438.2 438.2
5 C 8 through C 12 26.69 26.69 621.0 621.0
Component Composition 6 C 13 through C 20 19.84 19.84 747.3 747.3
Number Pseudocomponent (mol%) 7 C 21 + 14.65 14.65 888.2 941.0'
1 CO 2 3.03 Initiald co .1=0.1 ;=2,3 ... 7
2 C 1 , N2 41.66 Final d co :.1=0.1 ;=2,3 ... 7, ;¢6
3 C2 , C3 14.84 d C02 ,6 = 0.051'
4 C4 , Cs 6.78 Initial () = 1.0
5 C 8 through C 12 16.64 Final () = 0.0'
6 C 13 through C 20 11.65
7 5.40 • Parameters used during regression.
C 21 +

116 SPE Reservoir Engineering, February 1990

 TABLE 4-GAS·CONDENSATE COMPOSITION (mol%)

1.21
1.94
65.99
1 INITIAL MODEL 8.69
'5,000
5.91
-------- FINAL MODEL
0
2.39
I 2.78
o EXPERIMENTAL DATA 1
1.57

A
'0,000
8. 1.12
J6
1.81
W 6.59
~
::>
II')
."
35,000 API gravity of C 7 + at 60 o F. °API 51.4
W
~
1 I Specific gravity of C 7 + 0.7737

//
0.. Molecular weight of C 7 + 140
z
Q 30,000
< ","
~ /" "
::>
< /' 1
~, ... " TABLE 5-MODEL FLUID FOR GAS CONDENSATE
." 25,000 ,..,..1 _ "
~ ......0'"
......
Component Composition
~...
~ ... Number Pseudocomponent (moJO/o)
1 C 1 • N2 67.93
20,000 2 C 2• CO 2 9.90
0 20 .0 60 80 100 3 C3 5.91
4 C4 • Cs 7.86
MOLE , CO2 5 C6 1.61
6 C 7 through C 12 5.18
Fig. 2-Llght·oIl/CO 2 mixtures. 7 C 13 + 1.41

For this example, several cases were run to illustrate some pitfalls For all these cases, the interaction coefficients between the first
in EOS tuning, to stress that caution must be exerted when the component and the second through seventh components were also
potential regression parameters are selected, and to show the im- part of the potential regression parameter set.
portance of choosing the upper and lower bounds of the regression Table 6 shows the properties of the original fluid model, along
parameters, The cases described below were run. with those determined in Cases 1 through 3. The experimental data
Case J- The critical pressures and temperatures of all and the regression results for the saturation-pressure/composition
hydrocarbon components except C 6 were treated as potential data are plotted in Fig. 3, and those for swelling-factor/composition
regression parameters, with the lower and upper bounds being 0.7 data in Fig. 4.
and 1.3 times the initial values, respectively. Case 1 gives the best fit, followed by Case 2 and then Case 3.
Case 2-The critical pressures and temperatures of all compo- Note, however, that for Case 1, the Tc have lost the property of
nents except those of C6 were treated as regression parameters. monotonic increase with increasing carbon number. A possible cause
Also, the upper and lower bounds were much narrower than in Case of such behavior is that a match of the experimental data can be
1 to ensure that after regression Pc and Tc would foilow the trend achieved only by drastically changing the current set of regression
of decreasing Pc with increasing carbon number and increasing Tc parameters so that the trend of Tc cannot be preserved. To preserve
with increasing carbon number. the trend of component properties, bounds need to be placed on
Case 3-The critical properties of only the two heaviest fractions these variables. This is not a limitation of the regressi9n technique,
were chosen as regression parameters, again with the restrictions but the consequence of some inadequacies of the fluid representation.
on the upper and lower bounds to maintain the monotonic increase Case 2, on the other hand, maintains the desirable trend of
and decrease in Tc and Pc with carbon number. decreasing Pc and increasing Tc. However, the fit is not as good

TABLE 6-PROPERTIES OF ORIGINAL FLUID AND

REGRESSED FLUIDS FOR GAS·CONDENSATE SYSTEM

Component Pc (atm) Tc(K)

Number Component Initial Case 1 Case 2 Case 3 Initial Case 1 Case 2 Case 3
--- --- --- --- ---
1 C 1 • N2 45.08 48.09* 54.10* 45.08 188.7 229.0* 226.4* 188.7
2 C 2 • CO 2 50.36 50.36 60.43* 50.36 305.3 169.2* 336.0* 305.3
3 C3 41.90 41.90 39.92* 41.90 369.8 369.8 369.8 369.8
4 C4 • Cs 35.61 35.61 35.61 35.61 433.9 433.9 433.9 433.9
5 C6 32.46 32.46 32.46 32.46 507.5 507.5 507.5 507.5
6 C 7 through C 12 26.96 16.68* 23.33* 23.33* 586.7 664.1 * 654.6* 554.4*
7 C 13 + 19.30 9.65* 22.44* 13.51 * 729.3 944.2* 765.8* 733.6*

Initial d II = 0 for all i and j.

Case 1-d7 ,1 =0.111*; all other dll=O.
Case 2-d 2•1 =0.2*; d 3,l =0.13*; d 4 ,1 =0.001*; d 7,l =0.195*; all other dll=O.
Case 3-d 7 ,1 =0.2*; all other dl}=O.
Case 4-initial volume translation (kmol/m 3 ): V/ =0. i= 1 ... 7. Final volume translation
(kmol/m 3 ): v1t .. -0.002*; Vi =0.025*; all other V/ =0.
'Parameters used during regression.

SPE Reservoir Engineering, February 1990 117

 .- 0

--
EXPERIMENTAL DATA
INITIAL MOOEL
~
.9
..
/~~-
............
.700
..._.. _- CASE I
CASE 2
•.....•. ..0
.,00 ---- CASE 3
~~ .....
0- ~./
l
.300 /'<. . '
' ,...
~~~
'
...-
o EXPERIMENTAL DATA

!onon ,, j<.. ....."

' - - INITIAL MCJOEL
'

~
3.0

/~/
...II;! .'00
~
,, ......
....6
~ C>
~ 3900
/,//...... z
'"::I /
/,/.,. ...
:::;
w
~ 3700 " . .<... ~
on
2.0

/~

3500
" ../
... -;.~:ii

1.0 "'-!!!!!::;::=-~----"""'----~--T"""---r---~
3300+-----~--~----~----~--~----~--__, 10 20 30 50 60 70
10 20 30 60

MOLE "OF INJECTED FLUID MOLE "OF INJECTED FLUID

Fig. 3-Gas-condensatellnjected-fluld pressure/composition Fig. 4-Gas-condensstellnjected-fluld swelling-factor dia-

diagram. Injected fluid composition (mol%): 94.68% C 1 , gram. Injected fluid composition (mol%): 94.68% C 1 , 5.27%
5.27% C 2 , and 0.05% C 3 • C 2 , and 0.05% C 3 •

as Case 1. Also note that, in Case I, the EOS is tuned by changing cluding only a small number of component properties in the
a few parameters by large amounts, whereas in Case 2, the tuning regression parameter set. Note that the results from Case 3 are still
is done by adjusting more parameters by smaller amounts. Because adequate for most engineering calculations.
the bounds on the variables were narrowed in this particular case, Case 4-The fluid model obtained from Case 3 was used to fit
the critical pressures and temperatures of Components I and 2 hit the liquid dropout for the constant-volume-depletion study by ad-
the upper bounds, as did the critical temperature of Component 6. justing the volume translation of the lightest and heaviest com-
These variables were subsequently discarded from the active ponents.
regression set and replaced by other variables. The results are shown in Fig. 5. The values of the volume trans-
As expected, the match in Case 3 is not as good as in Cases 1 lations before and after the regression are also given in Table 6.
and 2 because the available regression parameters are further re- This case has been included to show that more than one type of
stricted. This case is included to show the possible effects of in- data from different experiments can be matched reasonably by
changing very few regression parameters.

26 o EXPERIMENTAL POINTS
Discussion
- - - - INITIAL MODEL In the examples, the upper and lower bounds of the variables were
- - - - - fiNAL MODEL selected arbitrarily. In practice, one would have narrow bounds for
those variables that are known with a fairly large degree of cer-
tainty, such as the critical properties of the light components. The
narrow bounds render the objective function relatively insensitive
;e- to the corresponding variables. Since the variables are scaled be-
, .... ---- ", tween zero and unity in Eq. 5, the derivatives iJjliJXj are very small
,,
22
w
\ for variables with narrow bounds, and these will likely be dropped
~ ,,
\
\ during the regression. Conversely, larger (but realistic) bounds
g \
should be plaoed on variables that are known with a small degree
I- 20
,," \\
\ of certainty. Thus, variable scaling makes variables more sensitive
~
,,"
as their bounds become broader, which is a very desirable feature.
\\
~
Note that there are two ways by which a variable can be discarded
/
, \
\
\
from the active regression set: when it loses its "sensitivity" and
0 when it hits its bounds for two successive iterations. In the first
5 18 // \
I \ case, the sensitivity is determined by comparing the variable
"
:::;
/
/
I \
\
\
derivatives with the deviations between calculated and experimental
I \ results. A variable that has lost its sensitivity at a particular iter-
16 \ ation may regain it later. Because the method presented does not
\
allow for the reintroduction of a discarded variable, the method
will converge to a solution that may not be a "global" minimum.
However, because the purpose of the algorithm is not to find the
,.~------~------r-----~------~------~
500 1000 1500 2000 2500 3000
"global" or "true" minimum of the regression problem, but rather
to find a minimum that gives an adequate fit of the experimental
PRESSURE (psig) data, the method works extremely well. The method has been found
Fig. 5-Gas condensate: constant-yolume-depletlon study. to be quite robust and does not require initial guesses close to the
final solution.

118 SPE Reservoir Engineering, February 1990

Conclualon. 4. Peng, D. Y. and Robinson, D.B.: "A New Two-Constant Equation of
State," Ind. de Eng. Chem. Fund. (1976) 15, 59-64.
This paper proposes the use of a regression scheme for EOS that 5. Bard, Y.: Nonlinear Parameter Estimation, Academic Press Inc.,
does not require a priori selection of regression parameters. The Washington, DC (1974) 83-123.
scheme also discards those active regression parameters that are 6. Himmelblau, D.M.: Applied Nonlinear Programming, McGraw-Hill
found to have outlived their usefulness in the regression process. Book Co. Inc., New York City (1972) 221-86.
This regression scheme performs very well and chooses regression 7. Schittkowski, K.: "The Nonlinear Programming'Method of Wilson,
parameters in a very sensible and expected manner, eliminating Han, and Powell With an Augmented Lagrangian Type Line Search
tedious trial-and-error runs. Function-Parts I and II," Numer. Math. (1981) 38, 83-127.
A modification of the adaptive technique proposed by Dennis et 8. Watson, S.T. and Lee, W.J.: "A New Algorithm for Automatic History
Matching Production Data," paper SPE 15228 presented at the 1986
al. 2 has been used in this work. It involves an automatic switching SPE Unconventional Gas Technology Symposium, Louisville, KY, May
between a Gauss-Newton and a secant method to approximate the 18-21.
Hessian matrix. The search directions are taken by using a region- 9. More, J.J.: "The Levenberg-Marquardt Algorithm: Implementation and
of-confidence approach. Theory," Lecture Notes in Mathematics, No. 630, NumericalAnalysis,
It should be stressed that caution should be exercised when an G. Watson (ed.), Springer-Verlag, New York City (1978) 105-16.
EOS is used that has been tuned to some experimental data. The 10. Li, Y.-K., Nghiem, L.X., andSiu, A.: "Phase Behavior Computations
predictive range of a tuned EOS should not be very different from for Reservoir Fluids: Effect of Pseudo-Components on Phase Diagrams
the experimental conditions that have been matched; otherwise, un- and Simulation Results," J. Cdn. Pet. Tech. (Nov.-Dec. 1985) 24,
reliable calculations may result. For example, an EOS tuned with No.6, 29-36.
11. Peneloux, A., Rauzy, E., and Freze, R.: "A Consistent Correction
hydrocarbon-gas-injection data that contain a small amount of C02 for Redlich-Kwong-Soave Volumes," Fluid Phase Equil. (1982) 8,
should not be used for predicting C02 flooding processes. 7-23.
12. Kenyon, D.E. and Behie, G.A.: "Third SPE Comparative Solution
Nomenclature Project: Gas Cycling of Retrograde Condensate Reservoirs," JPT(Aug.
dij = interaction coefficient between Components i and j 1987) 981-97; Trans., AIME, 287.
ej = value calculated from EOS
Appendix-Adaptive Nonlinear Regre ••lon
ef = measured value
/ = sum of squares of residuals (objective function) Adaptive Quadratic Model. The least-squares problem is restated
H = Hessian of objective function here as follows.
I = identity matrix mi,!l/(x)=r(x)Tr(x), ......................... (A-I)
J = Jacobian matrix x
nm = number of experimental data points where/(x)=r(x)Tr(X). To facilitate the regression effort, a
np = number of potential regression parameters simplified quadratic model, q, of the objective function / is first
nr = number of active regression parameters established. Expanding/in Taylor series around the kth iterate point,
Pc = critical pressure Xk, and ignoring the third- and higher-order terms results in
q = quadratic model of /
qk(x)~(Xk)+V/(Xk)TAXk+ 'hAxfV2/(Xk)AXk, ..... (A-2)
rj = weighted residual
S = approximation to second-order terms of Hessian of / where A Xk = x - Xk. The first and second derivatives of / are
Tc = critical temperature V/(Xk)=2J(Xk)Tr (Xk) ........................... (A-3)
V = molar volume from EOS
V' = corrected molar volume and V2/(Xk) =2[J(Xk)TJ(Xk) +S(x)], ................. (A-4)
VI = volume translation for Component i where J is the Jacobian matrix,
Wj = weight associated with ith measurement
x = vector of regression parameters J(Xk )jj =orj(Xk )/o(Xk)j'
Xj = physical properties (regression parameters) nm
Xj,max = upper bound for Xj and S(Xk)= E rj(Xk)·V2 ri (Xk)·
Xj,min = lower bound for Xj i=l
Yj = mole fraction of Component i
Eq. A-4 shows that the first term of the second expansion of/
o = radius of confidence is available through the Jacobian J. However, the second term, S,
(J = power for calculating interaction coefficients
is not readily available and is quite difficult to evaluate. Usually,
between hydrocarbons S is neglected, which is valid when the residuals rj(x) are small.
A = parameter used in Eq. A-9 For large residuals, it is necessary to include S, and the latter can
w = acentric factor be approximated by a secant method. 2
When Eqs. A-2 through A-4 are used, the quadratic model
Superscripts becomes
k = iteration level
T = transpose qi (x) = r?'i:+ 2A xitJ{i'k +Axf (J[Jk+Sk)AXk •.... (A-5)
where the subscript k denotes values evaluated at Xk. The super-
Acknowledgment script s in qS indicates that the S term is included. The minimum
This research was supported. by the members of the Computer of the quadratic model can be located easily with Newton's method.
Modelling Group. This results in the following iterative scheme:

Reference. (J[Jk+Sk)(Xk+l - Xk)= -J['i: .................... (A-6)

1. Coats, K.H. and Smart, G.T.: "Application ofa Regression-Based EOS Initially, So is set to 0, and subsequent Sk, k?:.I, are calculated
PVT Program to Laboratory Data," SPERE(May 1986) 277-99; Trans.,
from a recursive correlation resulting from a secant approximation.
AIME,282.
2. Dennis, J.E. Jr., Gay, D.M., and Welsch, R.E.: "An Adaptive Non- For convenience, the quadratic model in Eq. A-5 will be referred
linear Least-Squares Algorithm," ACM Trans. Math. Software (Sept. to as the secant model in subsequent discussions.
1981) 7, No.3, 348-68. Note that initially, So =0, Eq. A-5 reduces to the Gauss-Newton
3. Chen, H-S. and Stadtherr, M.A.: "A Modification of Powell's Dogleg model, q8, with the following iterative scheme.
Method for Solving Systems of Nonlinear Equations," Compo and Chern.
Eng. (1981) 5, No.3, 143-50. J[Jk(Xk+l-Xk)=-J['i: ........................ (A-7)
SPE Reservoir Engineering, February 1990 119
It is well known that for small k, the Gauss-Newton model, q8, However, it has been found that a good step AX'k can be computed
gives a better prediction to the objective function than the secant much more simply by use of Powell's dogleg method. 3
model, if, until the second derivative infonnation, S, is sufficiently
accurate through secant approximations. An ideal scheme thus would Quadratic-Model Switching. As discussed earlier, the secant
use q8 initially and then switch to if once an accurate S is model, if, and the Gauss-Newton model, q8, are accurate at
available. To retain generality, the proposed scheme allows the different times during the regression process. This requires a
switching from q8 to if and vice versa. The choices of the quad- switching criterion between the two quadratic models. The tradi-
ratic models and the radius of confidence of the models are inti- tional approach is to start the regression with So =O-i.e., with q8.
mately connected. Further description of the switching criterion is Because it takes a certain number of iterations to build sufficient
given later. information on Sk for it to be useful, the Gauss-Newton model is
used for a prescribed number of iterations and is replaced subse-
Quadratic-Model Constrained Minimization. Because quadratic quently by the secant model. To retain flexibility and generality,
approximations to f(x) are used, there exists a region 0 outside the dynamic switching method proposed by Dennis et al. is used
which the model fails. This imposes a validity limit to the size of in this work. Because both the radius of confidence and the choice
Ax in the iterative scheme resulting from these quadratic models, of the quadratic model affect the efficiency of the regression scheme,
which is expressed as the strategy is to adjust both during the regression to maintain a
reasonable convergence rate. This approach consists of trying the
min qk(X), ..................................... (A-8)
alternative quadratic model with the same 0 when the present model
subject to II Axil so, where q could be either q8 or if and 11·11 fails to reduce the objective function. If the alternative model does
denotes the Euclidean norm. The solution to the minimization (Eq. not lead to a more successful step, then the present model is re-
A-8) does not consist of simply truncating the step size as calcu- tained and the radius of confidence is reduced. Alternatively, an
lated from Eq. A-6 or A-7, but may involve the calculation of a attempt will be made to increase 0 for valid steps.
different step direction because of the constraint.
The solution to the minimization problem (Eq. A-8) is 8. M.trlc Conv.....on F.cton
AX'k=-(Hk+M)-l(J[rk)' ....................... (A-9) o API 141.5/(131.5+ 0 API) g/cm 3
atm x 1.013 250* E+05 kPa
where IIAX'k1l <0, ),=0 ..................... " ... (A-lOa) OF (OF-459.67)/1.8 K
Ibm mol/gal x 1.198264 E+02 kmoI/m 3
and II AX'k II =0, ),>0 ............................ (A-lOb)
·Converslon factor Is exact. SPERE
In nonlinear programming, ). is usually referred to as the
Original SPE manuacript r_lved for review April 8, 1887. Paper accepted for publication
Levenberg-Marquardt weighting factor. Because ). is not known Sept. 5, 1989. RevIeed manuscript received Jan. 10,1989. Paper(SPE 16343) first pr_ted
a priori, an iterative method for updating). is usually required. 8,9 at the 1887 SPE California Regional Meeting held In Ventura, April 8-10.

120 SPE Reservoir Engineering, February 1990