Reconciliation of Process Flow Rates

by Matrix Projection
Part II: The Nonlinear Case

Flow rate and concentration measurements in a steady state process are recon- C. M. CROWE
ciled by weighted least squares so that the conservation laws and other constraints
are obeyed. Two projection matrices are constructed in turn, in order to decom- Department of Chemical Engineering
pose the problem into three subproblems to be solved in sequence. The first McMaster University
matrix eliminates all unmeasured component flow rates and concentrations from Hamilton, Ontario, Canada L8S 4L7
the equations; the second then removes the unmeasured total flow rates. The
adjustments to component flow rates are iteratively determined, starting with
guessed values of unmeasured total flow rates.
Chi-squareand normal test statistics are derived by linearizing the equations, to
allow detection of gross errors in imbalances and adjustments of measurements.

SCOPE
The cornerstone for monitoringplant performance is a set of structive method of finding the RBS by a projection matrix that
steady state balances for component and total flow rates. Such effectively blanked out all unmeasured flows in the linear case.
flow rates are normally obtained from measurements of total The approach here is to extend the technique of Crowe et al.
flow rates and of concentrations, which are subject to random to the nonlinear case by constructing two successive projection
and sometimes gross errors, and thus in general violate conser- matrices. The first eliminates all unmeasured component flow
vation laws. The measurements should be reconciled, in some rates and concentrations; the second then eliminates the un-
“best” sense, to obey those laws and any other constraints that measured total flow rates from the balance equations. The
are required to be enforced. original problem is thus divided into three sequentially solved
The linear case, where it is assumed that the total flow rate is subproblems. Guesses of the unmeasured total flow rates are
measured in any stream in which a concentration is measured, used to solve for the adjustments to component flow rates.
was discussed by Crowe et al., (1983). In this paper, that as- These guesses are iteratively updated in the second subprob-
sumption is omitted so that the balance equations contain lem until convergence is achieved. Then those unmeasured
products of unknowns and thus are nonlinear (actually bilin- component flow rates that are determinable are found in the
ear). third subproblem.
Conflict with conservation laws can only arise in sections of Gross errors in the measurements can arise from instrument
the process where measurements have been made in all malfunction or miscalibration, sampling errors, and unsus-
streams in a balance equation. Thus previous work by Vacla- pected leaks or departures from steady state. Measurements
vek (1969), Mah et al. (1976), and Romagnoli and Stephano- that are in gross error should be detected before the data are
poulos (198 1) was directed to devising path-tracing or combin- reconciled so that they can be corrected or eliminated. Since
atorial algorithms for efficiently finding a maximal set of the equations are bilinear, the statistical analysis requires local
sections of the process with redundant measurements -the linearization about the measurement values, in order to find
reduced balance scheme (RBS). Unfortunately, such searches test criteria for the imbalances in component flow rates and for
must be done separately for each component, or each element the adjustments to particular measured values.
in the case of reactions. Crowe et al. (1983) proposed a con-

CONCLUSIONS AND SIGNIFICANCE

A pair of projection matrices has been applied to the equa- any one stream. The construction of these projection matrices
tions defining the conservation laws and other constraints. is direct and each requires the inversion of a matrix that is also
This approach finds a maximal set of balances on which recon- needed for subsequent steps in the solution.
ciliation can be done. There are no restrictions imposed upon Statistical tests for the set of imbalances or residuals in the
the placement of measurements, nor upon what is measured in constraints (chi-square)and for individual imbalances and ad-

616 April 1986 Vol. 32, No. 4 AIChE Journal

justments of measurements (unit normal variate) are devel- then be used jointly to find which imbalances or adjustments
oped by linearizing the constraint equations. These tests can are grossly in error.

INTRODUCTION Stanley and Mah (1981a) developed the concepts of global

and local observability of state variables, given a set of measure-
The problem to be addressed is that of reconciling measure- ments and constraints, for the nonlinear problem. They also
ments of total flow rates and concentrations such that the con- (1981b) applied graph theory to mass-energy flow networks to
servation laws and possibly other constraints are obeyed, by a classify unmeasured variables as globally (or locally) observ-
“best” choice of adjustments to the measurements. In an ear- able or unobservable. The global (or local) redundancy of a
lier paper, Part l, Crowe et al. (1983) presented a solution to measurement was determined by the effect of its deletion on
the linear case where it was assumed that the total flow rate in the global (or local) observability.
any stream was measured if any concentrations were measured Knepper and Gorrnan (1980) presented the theory of nonlin-
in that stream. A summary of the published work done on the ear estimation of constrained data sets and provided statistical
linear case was given in Part I, and more general reviews have tests for the estimate’s being free from gross error. Romagnoli
been presented by Hlavacek (1977) and Mah (1981). (1983) recently discussed a recursive technique for sequential
In this paper, no restrictions are imposedupon the location of addition or deletion of either measurements or constraints.
streams in which total flow rates or concentrations are mea- Beckmann (1982) used nonlinear programming to adjust
sured nor upon which concentrations are measured in any measurement data in batch and steady state processes. Data
stream. As a result, the balance equations will, in general, con- reconciliation of an operating chemical process was studied by
tain products of pairs of unknowns and thus will b e bilinear. Ham et al. (1979), and of mineral and metallurgical processes
Vaclavek et al. (1976a,b) developed the theoretical basis of the by Cutting (1976), Hodouin and Everell (1980), Hodouin et
bilinear problem and discussed conditions under which the al. (1982),Mularetal. (1976),RagotandAubrun(1980),Smith
original balance equations could be reduced in number. They and Ichiyen (1973), White et al. (1977) and Wiegel (1972). A
assumed that either all o r none of the concentrations in astream computer program for nonlinear least squares estimation was
were measured and thus could partition the streams into four reported by Laguitton (1980).
categories, namely:

Category Total Flow Concentrations PROBLEM STATEMENT

1 M M If vector r represents the n true component and total flow
2 U M rates in all streams of a process, the m material balances and
3 M U other constraints among them can b e written as
4 U U
Bz+STC=O (1)
where M = measured, U = unmeasured. where matrix S represents the (p X m) stoichiometric matrix
and ( is the composite vector of p extents of reaction for all
However, with an arbitrary distribution of measurements, the reactions in the process, as described in Part I.
classification must refer to components in streams, in which The columns of the (m X n) matrix B are then permuted and
case categories 3 and 4 can be combined since the component partitioned so that
flow rate is unmeasured in each case. Thus, there will be three
categories of variables: B -+ [~oI~,IB2lB,I
1. Total flow rate and concentration measured and adjust- where columns of B, correspond to exactly known flow rates k
able. and those of B, (i = 1 , 2 , 3 ) to n, components in categories 1 , 2 ,
2. Concentration measured and adjustable. and 3 respectively.
3. Total flow rate unknown or measured; component flow The flow of component c in stream j , in category 1 is defined
rate unmeasured.
The classification by Vaclavek e t al. (1976a) of unknown vari- by
ables into determinable and indeterminate can be extended Xcj = q c c , (2)
here in that each of categories 2 and 3 separately may be so where Mj is total flow rate and ccj is concentration.
classified. The indeterminate variables in each case will corre- The measured values will be represented by - over the sym-
spond to dependent columns of the respective matrices in the bol and the estimated values by Thus, ^.
balance equations. Vaclavek e t al. also divided the measure-
ments according to whether they should b e corrected or not. id= ** + a* (3)
This division will depend not only on the flow sheet but on the where a* is the adjustment in stream j for the measured flow
variance-covariance (dispersion) matrix of the measurements. rate of component c.
Only if the dispersion matrix is diagonal can this division be For a stream with unmeasured total flow rate, we use d, Gfor
done from the flow sheet alone. the measured concentration and its adjustment. Then,
In order to reduce the number of balance equations to a
minimum number, without sacrificing information, Vaclavek et d,, = ci,, + s, (4)
al. proposed a two-step reduction. The formulation presented for the estimated concentration of component c in stream 4 .
here will lead to a different two-step reduction scheme which The unknown total flow rate will be N,.
first eliminates variables of category 3 and other unmeasured As in Part I, we define
variables, such as extents of reaction, and then eliminates the
total flow rates corresponding to variables of category 2. P = [B3(ST] (5)

AIChE Journal April 1986 Vol. 32, No. 4 617

and represent all unknown component flow rates and extents of There are several advantages to defining Y and Z separately.
reaction by vector 1;. Then the constraints must b e obeyed by First, the inverses of P, and D , are more efficient to compute
the estimated values, so that separately than that of a larger combined matrix. Secondly,
separate conditions are obtained for indeterminacy of variables
+ +
~~k B,(z+ a) B , N ( ~ +s) ~ t=.o + (6) in category 3 and of total flow rates in category 2. Thirdly, Y is a
where N is the diagonal matrix of unknown flow rates, with as constant matrix_so that Z contains all of the statistical variability
many entries for each stream as there are measured concentra- due to that of d .
tions in that stream. The reconciliation problem can then b e The third step in the simplification involves rewriting the
defined. second term in the objective function, Eq. 7 , as
Problem P i S T Z ' S = (NS)T(NZdN)-l(NB) (18)
If estimates are made of N , say N o , then define
0 A'oZdNo (19)
subject to Eq. 6. Here, s1and s d are estimates of the respective
variance-covariance matrices. so that problem P1 can be restated as P2.
There are several steps that can b e taken to simplify this
Problem P2
problem. The first is to define, as in Part I, the matrix Y such that
PTY = 0 (8) (a) Solve:
with the columns of Y forming a basis for the null space of PT. aSNb)
+
Min G(a,NS)A a ' 2 i 1 a (NS)TXF1(NS) (20)
Then, Eq. 6 can be reduced to
subject to Eq. 15.
+
Y T [ B , ~ B,(Z + a) + B +
~ N ( 41
~ =o (9) (b) From the solution to part (a), solve Eq. 9, rewritten with
This is the first stage of the reduction in the number of equa- Eq. 1 4 as
tions by eliminating all category 3 variables. The construction
of Y can be done, as in Part I, by partitioning
D n = -YTIBok + B1(Z + a) + B 2 ( N 6 ) ] (21)
for n.
(c) From the solutions to parts (a) and (b), solve

where P, is a maximal square nonsingular matrix contained in P.

Pt- = -[Bok + B1(Z+ a) + BzN(d + S)] (22)
W e can then choose for c'.
The necessary condition for a solution to exist for each ofEqs.
Y = [-Pz P i 1111 (11) 21 and 22 is satisfied by the definitions of Y and Z. If the inde-
By definition, the righthand column partition of P depends on terminate variables are set to zero, the remaining unknowns
the lefthand columns, so that Eq. 8 is satisfied. The dependent can b e solved for by
columns of P correspond to indeterminate variables in cate- n d = D;'[ZIO]R(21) (23)
gory 3.
It should be noted that when P1is not in the upper left corner, and
the columns of each partition of Y T in Eq. 11 should be per- c d = Pi'[ZIO]R(22) (24)
muted to correspond respectively to the rows of P, and P 2 .
The second step of simplication involves defining a matrix 2 whereR(21), R(22) are therighthandsides ofEqs. 2 1 and22, of
that is a basis for the null space of which we retain only the rows corresponding to D , , P1,respec-
tively.
D T P [Rz,d,Rz,dz ... BzLdC. . . ITY (12) Since we already have the inverses of D , and P,, the equa-
where Bpt is the set of columns of B, corresponding to stream t , tions are efficiently solved. The space of all solutions to Eqs. 2 1
and d, is the vector of measured concentrations in that stream. and 2 2 can b e found from the fact that
Then,
(25)
DTZ P 0 (13)
and Thus
D n = Y TB2Nd (14)
where n is the vector of distinct unknown total flow rates in
category 2, so that Eq. 9 can be simplified t o for arbitrary H. Similarly, ageneral solution for ucan b e written.
ZITIBok + B1(Z + a) + BzNS] = 0 (15) There is no guarantee that the solutions for nd or l)d will be
nonnegative so that the general solutions could be used to find
This then further reduces the size of the problem b y eliminat- vectors n and t;, which conform to additional specifications such
ing total flow rates in category 2. W e can represent Z analo- as nonnegativity.
gously to Y, namely if
Problem Solutions
It remains for us to solve problem P2(a).The Lagrangian Lzis
defined by
with D, a maximal square nonsingular submatrix in D.
ZT= [- D,Di'(Z] (17)
1
2
+
L2 6 - [aTF1a (NS)lZi'(NS)]-

The total flow rates that correspond to dependent columns of D

+
-ITZTYqBok B1(Z a) B,(NS)] + + (27)
are indeterminate. with Lagrange multiplier vector, 1.

618 April 1986 Vol. 32, No. 4 AIChE Journal

 4(1) stream. Thus
f 1

LTE = p (35)
where 1 is a vector of ones. Then
M = (1=*)/p (36)
and c are found from Eq. 33.
(d) In the absence of any of the above conditions, we can
minimize the weighted sum of squares of the adjustments as
follows:
Min
M
J = [(E - E)%;'(E - i') + (&f- ~l?)~/a2,] (37)

Figure 1. Flow diagram of Vaclavek's example s( j ) : Stream Note that J is a function only of M because c can be eliminated
by Eq. 33.
s(Category j ) .
If we set dJ/dM = 0, we obtain
Then derivatives are taken with respect to a and (NS) and set - M M 3 + ah(n;rZ.- i)%;li= o
M 4 (38)
equal to zero, giving respectively, which can be readily solved using Newton's method, with an
a = P,BpZA (28) initial guess of M = M.
and
NS=P,BaZA (29) APPLICATION TO VACLAVEK'S EXAMPLE
Note that if D has full row rank, Z = 0, a = 0 , and N 6 = 0.
FromEq. 15, The flow scheme of the example of Vaclavek et al. (1976b)is
shown in Figure 1 . In this case, all concentrations were assumed
ZrYTIBla + B&"] =-ZVT[B,k + Blx] (30) to b e measured or none were, so that the streams themselves
From Eqs. 28 and 29, with can be classified according to categories 1-3. The category of
each stream is shown in parentheses. The matrices B1, B , and
H A Y T ( B I Z I B T +B p Z p B ; ) Y (31) B, are shown in Table 1, with Z as the identity matrix of order
we obtain equal to the number of components and the remaining entries
I = -(ZTHZ)-'ZTyTIBok BIZ] + zeroes. The matrix Y can then be defined as:
(32)
This allows computation of I, a, and N S from Eqs. 32, 28, and y'=[. . . .
Z I Z . . l
.] (39)
29. Problem P2(b) can then be solved for n. Now, because 2,
was computed using a guess for N (or n), we can iterate by We note that the remaining balances involve node 4 alone and
updating Zz and solving P2(a) again. This shows a fourth advan- nodes 1, 2 , 3 , and 6 combined.
tage of splitting matrices Z and Y, namely that this iteration Now, the matrix D can b e written as
does not involve variables of category 3.
If we wish to apportion the adjustments aj of component flow
rates x, in streamj among the total flow rate M , and the concen-
trations c,, we will require In Vaclavek's case, there were four components so that D had
full column rank. In that case,
Ej = i j / M j (33)
This then leaves only one further condition to b e specified.
Having suppressed the subscriptj for simplicity, we can list the Thus, adjustments would be zero for the streams incident to
following possible conditions: node 4 and adjustments to flows in streams 1.8, and 9 could be
(a) A particular concentration c, is exactly known. Then found as problem P2(a). Then estimates of the total flows of
streams of category 2 incident to node 4 can be found.
M = x,/c, (34) Now, if there were fewer than four components, one or more
and Eq. 33 gives the other concentrations. total flows in streams 11, 13, 14, and 15 would be indetermin-
(b) Total flow rate M is exactly known. Then c are all calcu- ate but the remainder could b e estimated. If there were more
lated from Eq. 33. than four components, the null matrix in Eq. 41 would be re-
(c) The concentrations must add up to a specified quantity p; placed by -D&' as discussed above. Then, some balance
for example, the mass or mole fractions of all species in the envelope($ around node 4 would b e retained in the reduced
TABLE & ,Be, B,
1, MATRICES FOR VACLAVEK'S EXAMPLE

Stream Number

1 4
I

-I
8

-z
.

. .
. -z
9
.

z
1

:I[: z1r !I
2

Z
5

-I
I
10

-I
11 13 14

-1
I
Z 1
15 2 3
-Z
I
6

-I

I
7

-Z

I
1 6

-I

AIChE Journal April 1986 Vol. 32,No. 4 619

problem P2(a), adjustments to the streams incident to it would Thus E has a degenerate normal distribution,
not all be zero, and the total flow rates could again be esti-
mated.
-
E N(O,HZT-’ZTH), (53)
By comparison, Vaclavek e t al. excluded node 4 from the whose elements can be tested by
reduced balance scheme at the first stage, where node 1 was
retained alone, and nodes 2 , 3 and 6 were combined as balance Zi A €k/(HzT-’zTH)f/z (54)
envelopes. On the second stage of reduction, these four nodes Mah and Tamhane (1982) have discussed the power of such
were combined. Vaclavek et al. imposed the condition that the tests. They pointed out that to require each e k ,a j , and dt to lie
sum of concentrations in any stream equal unity. This would not within a given level of confidence would lead to unwarranted
affect the results if the measured concentrations were normal- rejection of a value which is consistent with N(O,1), an error of
ized in advance. However, without such a priori normalization, type I. They suggested, following Sidak (1967). setting a confi-
an additional constraint for each stream in category 3 would b e dence level j? for each measurement of a set of s independent
added, which in turn, would add a row to Z T . Then the adjust- measurements such that jointly, the confidence level was a,
ments to concentrations in streams 1 1 , 1 3 , 1 4 , and 15would b e with
nonzero, as expected. p P 1 - (1 - a)’P (55)
Sidak showed that this would provide a test for the a confidence
STATISTICAL TESTS FOR GROSS ERRORS level of the distribution of the maximum of such a set, which
gives an upper bound on a type I error.
The statistical analysis is complicated by the dependence of Z Then in testing elements of e using z z , and in the light of
on the random vector d. In the Appendix, aTaylor series analy- Eq. 42, one would find the value o f p using s = r,, the rank of Z
sis is used to derive approximate normal distributions of (and T ) .
-
e P ZTyTB13i. N(0,T) (42) However, for tests of z;, z i , and zf, not only are the variance
matrices singular with rank r,, but a a n d 6 arejointly dependent
a - NO,Qi) (43) on E and on e, through Eqs. A16 and A17, which give
-ZTYT(Bla + B z N 6 ) = ZTe = e (56)
(44) It seems reasonable to suggest again using s = r, to find j3 for
with Q1, Qz usually singular. testing z;, z j , and z f , although Iordache et al. (1984) have
The data can then be tested as a whole since argued for a value intermediate between r, and the total num-
ber of elements in a and 6 or in E , respectively. In any event, the
xz 4 eTT-’e (45) practical aim is to flag suspect values of e, a, 6, and€ so that they
is a chi-square statistic with the number of degrees of freedom can be reexamined and any faults in the measurements can be
equal to the rank of Z . identified. When the tests with one set of measurements do not
As in Part I, linear combinations of e can b e tested against the clearly identify suspect values, tests using averages of several
unit normal distribution since sets of successive measurements should help to reveal any per-
sistent gross errors since the variances are all reduced in pro-
Z P
w Te
(w TTw)‘Iz
- N(0J) portion to the number of such sets.

In particular, for tc = u k ,the kth unit vector DISCUSSION

(47) The main advantage of the proposed method is that it reduces

the computing load required for matrix inversion. The determi-
Individual elements of a, 6 can be tested analogously by nation of Y and Z need only be done once and only the inversion
of ( Z T H Z ) is required at each iteration to solve for total flow
rates n.
The formulation as problem P2 has the advantage over P1 of
and separating the calculations of a and ( N 6 )from that of Nitself. In
problem. P1, they must be solved together and Z , if used, will
change at each iteration.
(49) The savings in matrix inversion are substantial compared to
the general method of Britt and Luecke (1973), which would
By comparing elements of e, a, and S whose absolute values
require the inversion of
are too large to be consistent with Eqs. 46,48, and 49, one has
information with which to diagnose which measurements are
the most likely cause of the apparent gross errors.
at each iteration, instead of ( Z T H Z ) .
One disadvantage of using zias a test statistic for the residuals
A computer program to implement the method was written
in the constraints is that Z T is not simply related to balances
and tested on a VAX 11/780 computer. Among the literature
around nodes of individual species. An alternative is to define
cases studied were those of Cutting (1976), Ham e t al. (1979).
(with k = 0 )
Mular et al. (1976), Smith and Ichiyen (1973), and Wiegel
E A YT[B,Z + B , N 4 (50) (1972). The results corresponded reasonably closely to those
reported by these authors, when direct comparison was pos-
By using Eqs. 9, 28, and 29, we can obtain
sible.
E= -HZ1 (51) The flotation circuit of Smith and Ichiyen (1973), shown here
in Figure 2, is a useful example to illustrate the application of
and with Eqs. 32 and 4 2
theory. Concentrations of C u and Zn were measured in all
E = HZT-le (52) streams except number 8 but no total flow rates are given, so

620 April 1986 Vol. 32, No. 4 AIChE Journal

 1
1 - 2 p 2 3 * 3 - 4
is the fact that the Zn concentration in stream 8 must be nega-
tive. This was also true in Smith and Ichiyen's results although
they do not remark on it.
8
h In case 2, the concentration of Zn in stream 1 is assumed to be
unknown. Although this results in acceptable values of the test
statistics, the negative concentration of Zn remains.
In case 3, we imposed a further restriction that nzn.8=
O.OO1nzn,, in order to prevent a negative concentration. The
Figure 2. Smith and Ichiyen's flotation circuit Unit 1, unit cells;
result is that concentrations in streams 5 and 6 and dcu,lnow
Unit 2, Cu flotation; Unit 3, Zn flotation.
have values of za that are too high. It is clear that there is
something wrong with the data, and indeed Smith and Ichiyen

['I
that only relative flows can be estimated. Thus we define the
feed flow to b e unit mass. The B matrices are then noted that the X-ray analyses for Zn were in error. There is in

[-' :]
fact a second set of data taken six minutes earlier, given in
Stream: 1 2 3 4 5 6 7 Figure 2 of Smith and Lewis (1969). If one repeats the adjust-
I - z. . -I . . ments with the average of these two sets, the faultiness of the
data is even more glaring, since the variances are all halved.
B= ; B, =
'
. I - I . . - I '
. . . I - I .
8 CONCLUSION

r .i The theoretical approach to solving the problem of reconcil-

ing flow measurement data for the linear case, as presented in
Part I, has been extended to the nonlinear case. The unknown
component flow rates and extents of reaction are deleted, as in
where I is the (3 X 3) identity matrix so that three balance the linear case, by the constant projection matrix Y. Then, the
equations are imposed for each of the four nodes (i.e., Cu, Zn unknown total flow rates are deleted by a second projection
and total mass). matrix, Z, which is stochastic because of the definition, Eq. 13,
The matrix Y can then b e seen to b e in terms of measured concentrations, d. The original reconcilia-
tion problem is then reformulated as a sequence of three sub-
I . . . problems, P2(a), (b), and (c). The first two are iteratively solved
YT=[: z 1 z ] in alternation, and after convergence the third is solved sepa-
rately.
Matrix D is given by

D= [ -d,
d, -dg
.
.
d, -d,
. -d,
. d, -d,
.
.
. -d,
ACKNOWLEDGMENT

The author is grateful to the Natural Sciences and Engineer-

where d, = Id,, d,, 11: for stream i, in mass fractions. The ing Research Council, Ottawa, Canada, for financial support.
relative standard deviation of the measurements was the same He wishes to thank P. Rosza, Technical University of Budapest,
as in Smith and Ichiyen (1973), namely 6.56%. Hungary, for his interest in this work and for many stimulating
The results of the reconciliations are shown in Table 2. In discussions. He also wishes to acknowledge the assistance of C.
case 1, all the data were used and it is clear that several values of Feierabend, McMaster University, in preparing the computer
z" are too large, as is the chi-square value. Even more disturbing program.

TABLE
2. SMITH AND ICHIYEN (1973) FLOTATION
CIRCUIT
Stream
1 2 3 4 5 6 7 8
&%
, 1.93 0.45 0.13 0.09 19.86 21.44 0.51 -
&"% 3.81 4.72 5.36 0.41 7.09 4.95 52.10 -
Case 1 dC"% 1.928 0.45 0.128 0.09 19.88 21.43 0.513 35.36"
(all data)
x* = 13.1 0.38 WC"I 0.77 0.38 0.38 0.77 2.01 0.38 -
Ix%.o.95= 7.8) 4.95 dzn% 4.78 4.95 0.41 7.07 4.92 51.88 -14.4'
lLl 3.55 0.77 1.91 0.38 0.77 2.01 0.38 -
Total Flow 1 .o 0.924 0.915 0.835 0.0761 0.0845 0.0808 0.00844
Case 2 dC"% 1.933 0.45 0.13 0.09 19.84 21.44 0.511 36.36'
(dZn.1 unknown) WC"I 0.72 0.72 0.11 0.11 0.72 0.72 0.11 -
xe = 0.52 &"% 5.20" 5.04 5.22 0.41 7.11 4.94 52.03 -15.4'
lG"l - 0.72 0.68 0.11 0.72 0.72 0.11 -
Ix%,o.85= 6.0) Total Flow 1.0 0.923 0.915 0.830 0.0765 0.0846 0.0854 0.00817
Case 3 1.86 0.46 0.129 0.09 21.2 20.13 0.51 16.3'
2.57 2.67 0.19 2.67
0.092 - 2.71
- 0.09 -
5.18" 5.08 5.18 0.41 6.64 5.23 52.04 0.25
nzn.2
xe = 7.83 IGnI - 0.93 0.92 0.09 2.57 2.57 0.09 -
&,85 = 7.8) Total Flow 1 .o 0.932 0.913 0.829 0.0677 0.0867 0.0843 0.019
* Value reauired to satisfv balance.

AIChE Journal April 1986 Vol. 32, No. 4 62 1

NOTATION e L ZTYTB,x (Al)
and of the adjustments a and 6,based on the assumption that
a = vector of adjustments to measured component flow x,d are independently normally distributed:
rates (n,X 1)
B = balance matrix of process (m X n) 6x - N(0,SJ (A21
6d - N(O,&)
B, = columns of B corresponding to fixed component flow
rates (m X no) 643)
B, = columns of B corresponding to measured component with
flow rates (m X n,) 6xLx-i
B, = columns of B corresponding to components with mea-
sured concentrations in streams with unmeasured total SdPd-2 (A4)
flow rates (m X n,) Let us define
B3 = columns of B corresponding to components with un- T B ZTHZ (A51
measured concentrations (m X n3)
WenotefromthedefinitionsinEq. 12ofDandEq. 13 ofZ, that
c = concentration vector corresponding to Bl(nl X 1)
d = measured concentrations corresponding to Bz(nzX 1) azT
-D=-ZTYTIO . . . OB,,,, 0 . . . 01
D = matrix, Eq. 12 (r, X s,) (A6)
ad,
E = expectation operator
e = imbalance vector, Eq. A1 where B,., is the column of B, corresponding to d, and i corre-
H = matrix, Eq. 31 sponds to any category 2 concentration. Furthermore,
I = identity matrix 2P Z(2) (A7)
J = objective function, Eq. 37
k = vector of fixed component flow rates, corresponding to and by definition
Bo(% x 1) ZTD(d) = 0 (AS)
L = Lagrangian, Eq. 27
m = number of constraints Then, from Eqs. 9 and 14
M = total flow rates of streams with measured component ZTYTB,4 = 0 649)
flow rates (s, X 1)
n,= number of variables in category i = 0, 1, 2, 3 Expanding the residual e in a Taylor series to first order, and
n = vector of unmeasured flow rates of streams with mea- taking account of Eq. A8
sured concentrations, d(sz X l) air
N = diagonal matrix, with each component of n repeated to e = ZTYTB,Sx + -Y T B , f 6d, (-410)
correspond to the number of measured concentrations 1 ad,
in that stream (nzX nz) Now, with Eqs. A6, 9, and 14
p = number of independent reactions
P = [B, : S 1'
+
e = ZTYTB16x ZTYTB2NSd (All)
Qi = matrices, Eqs. A20 and A22 for i = 1, 2(n, X n,) Thus,
r p= rank of matrix P
s,= number of streams in category i = 1, 2
E(e) = 0 ('412)
S = stoichiometric matrix for process ( p X m) E(eeT)= ZTYT(B,ZIBT+BzNxdNBl)YZ (A13)
T = ZTHZ(r,X r z ) In the limit, 8, + N Z d N if the solution is convergent, so that
i: = vector of unmeasured component flow rates corre-
from Eqs. A5 and 31 we obtain
sponding to B,, combined with unknown extents of re-
+
action [(n, p) X 11 E(eeT)= ZTHZ = T 6414)
x = vector of component flow rates corresponding to Thus, the residual e is approximately normally distributed as
B,(n, x 1)
Y = matrix, Eq. 8 [m X (m - rp)] e -N(0,f) (A15)
Z = matrix, Eq. 13 [r, X (ry- rD)] The distributions of a and 6 can also be approximated from
za= vector of unit normal variates for test of adjustment a Eqs. 28, 29, and 32, that is
Greek Letters
a = -S,BPZT-'e (-416)
6 = vector of adjustments to concentrations d(nz X 1) 6 = - N - 5,B p Z T - l e 6417)
1 = vector of Lagrange multipliers (rr X 1)
t. = vector of extents of reaction ( p X 1) A Taylor series expansion with 6 = 0 gives
:a = variance of M a = -PI B P Z f - l e ('41 8 )
S = variance-covariance matrix:
2, = component flow rates x(nl X n,) 6= - N - ' L , B ~ Z ? ' - l e (A191
Lc= concentrations c(nl X n,) Thus,
& = concentrations d(nz X np)
L, = matrix, Eq. 19 (nzX n2) E(Q)= O,E(6)= 0
E(aaT)= L I B ~ Z f - ' Z T Y T B , Z 0
l Q, (A20)
APPENDIX: STATISTICAL ANALYSIS
E ( B T )= N - 1 Z 2 B ~ Z T - ' Z T Y T B , Z z N - 1 (A21)
The estimates i ,2 can be regarded as maximum likelihood If 2, = NZdN
estimates, subject to Eq. 15, ofx andd, respectively. We wish to
deduce the mean and variance of the residual (with k = 0 for E ( B T )= &NB;YZT-'ZTYT&md P 9 2 (A22)
simplicity) Note that Q, , Q, will be singular if a, 6respectively have more

622 April 1986 Vol. 32, No. 4 AIChE Journal

elements than the order of T .Thus,a, Ghave degenerate normal surement Errors for Complex Chemical Plants,” Chem. Eng. Sci., 35,
distributions. 1,067 (1981).
Smith, H. W., and N. Ichiyen, “Computer Adjustment of Metallurgical
Balances,” Can. Znst. Mining Metall. (C.Z.M.) Bull., 66, 9 7 (Sept.,
LITERATURE CITED 1973).
Smith, H. W., andC. L. Lewis, “Computer ControlExperimentsat Lake
Beckman, J. R., “Data Adjustment for Nonreactive Batch or Steady Dufault,” Can. Znst. Mining Metall. (C.1.M.) Bull., 62, 109 (Feb.,
State Processes,” Chem. Eng. Commun., 15, 357 (1982). 1969).
Britt, H. I., and R. H. Luecke, “TheEstimation ofparameters in Nonlin- Stanley, G.M., and R. S. H. Mah, “Observability and Redundancy in
ear Implicit Models,” Technometrics, 15, 233 (1973). Process Data Estimation,” Chem. Eng. Sci., 36,259 (1981a).
Crowe, C. M., Y.A. Garcia Campos, and A. Hrymak, “Reconciliation of -, “Observability and Redundancy Classification in Process Net-
Process Flow Rates by Matrix Projection. I: The Linear Case,” AZChE works. Theorems and Algorithms,” Chem. Eng. Sci., 36, 1,941
I. , (1983).
29,881 (1981b).
Cutting, G . W., “Estimation of Interlocking Mass Balances on Complex Vaclavek, V., “Studies on System Engineering 111. Optimal Choice of
Mineral Beneficiation Plants,” Znt. ]. Mineral Proc., 3, 207 (1976). the Balance Measurements in Complicated Chemical Engineering
Ham, P. G., G .W. Cleaves, and J. K. Lawlor, “Operation DataReconcil- Systems,” Chem. Eng. Sci., 24,947 (1969).
iation: An Aid to Improved Plant Performance,” Proc. loth World Vaclavek, V., M. Kubicek, and M. Louchka, “Calculation of Material
Petroleum Congr., Bucarest, 4, 281 (1979). Balances for Chemical Engineering Systems with Due Allowance for
Hlavacek, V., “Analysis of a Complex Plant-Steady State and Tran- Measurement Errors,” Theor. Found. Chem. Eng., 9, 242 (1976a).
sient Behavior,” Comput. Chem. Eng., 1 , 7 5 (1977). (Translation of Teor. Osnovy Khim. Tekhnol. 9, 270 [ 19751.)
Hodouin, D., and M. D. Everell, “A Hierarchical Procedure for Adjust- -, “Calculation of Material Balances for Chemical Engineering Sys-
ment and Material Balancing of Mineral Process Data,” Znt. J.Mineral tems with Due Allowance for Errors in Measurement Classification of
Proc., 7,91 (1980). Stream Parameters,” Theor. Found. Chem. Eng., 10, 256 (1976b).
Hodouin, D., T. Gelpe, andM. D. Everell, “Sensitivity AnalysisofMate- (Translation of Teor. Osnovy Khim. Tekhnol. 10,281 119761.)
rial Balance Calculations- An Application to a Cement Clinker White, J. W., R. L. Winslow, andG. J. Rossiter, “AUsefulTechnique for
Grinding Process,” Powder Technol., 32, 139 (1982). Metallurgical Mass Balances- Applications in Grinding,” Znt.J. Min-
lordache, C., R. S. H. Mah, and A. C. Tamhane, “Detection of Gross eral Proc., 4, 39 (1977).
Errors in Process Data Using Measurement Test,” Paper 43c, AIChE Wiegel, R. L., “Advances in Mineral Processing Material Balances,”
Spring Meet. Anaheim, CA (1984). Can. Metall. Quart., 1, 413 (1972).
Knepper, J. C., and J. W. Gorman, “Statistical Analysis of Constrained
Data Sets,” AZChEJ., 26, 260 (1980). Manuscript received Aug. 16,1984. and revision received May 14,1985.
Laguitton, D., “Material Balance of Mineral Processing Flowsheets.
FORTRAN Program MATBALB Users’ Manual,” Division Report
MRP/MSL 80-33 (IR),CANMET, Ottawa (1980). ERRATA
Mah, R. S. H., “Design and Analysis of Process Performance Monitoring
Systems,” Chem. Proc. Control 2, Proc. Eng. Found. Conf., SeaIsland, In the paper titled “Reconciliation of Process Flow Rates by Matrix
GA, AZChE, 525 (1981). Projection” by C. M. Crowe, Y.A. Garcia Campos, and A. Hrymak (29,
Mah, R. S. H., G. M. Stanley, and D. M. Downing, “Reconciliation and p. 881, December, 1983), the following corrections are made:
Rectification of Process Flow and Inventory Data,” Znd. Eng. Chem. p. 881, right column, line 10 in Scope: “malfunction of miscalibration”
Proc. Des. Deo., 15, 175 (1976). should read “malfunction or miscalibration”
Mah, R. S. H., and A. C. Tamhane, “Detection ofGrossErrors in Process p. 882, right column, line 6 above Eq.1,Eq. 2, and lines 1.2 following
Data,” AZChE]., 28, 828 (1982). Eq. 2 as well as Eq. 8 on p. 883: “E” should read ‘‘5”
Mular, A. L., et al., “Mass Balance ofa Grinding Circuit,” CZM Bull., 69, p. 884, left column, line 3 under “Detection of Errors”: “unknown
124 (Dec., 1976). variance” should read “known variance”
Ragot, J., and M. Aubrun, “A Useful Technique for Metallurgical Mass p. 885, add at end of Eq. 35: = a,
Balances- Applications in Flotation,” Proc. 3rd ZFAC Symp. Auto- p. 887, Table 3, rightmost column heading: “X” should read “x” twice;
mation in Mining, Mineral and Metal Processing, Montreal, 229 and Table 3, line 5, leftmost column: “Straints” should read “Con-
(1980). straints”
Romagnoli, J. A., “On Data Reconciliation: Constraints Processing and p. 888, reference to Kuehn and Davidson: “1967” should read “1961”;
Treatment of Bias,” Chern. Eng. Sci., 38, 1,107 (1983). and reference to Swenker: “Messergenbnissen” should read “Mes-
Romagnoli, J. A., and G. Stephanopoulos, “On the Rectification of Mea- sergebnissen”

AIChE Journal April 1986 Vol. 32, No. 4 623