Relative Gain Array Analysis for Uncertain

Process Models
Dan Chen and Dale E. Seborg
Dept. of Chemical Engineering, University of California, Santa Barbara, CA 93106

Relati®e gain array (RGA) analysis has been widely used in process control to identify
promising control structures and to characterize the degree of process interactions be-
tween controlled and manipulated ®ariables. Howe®er, the influence of process model
uncertainty on RGA analysis has recei®ed little attention. Analytical expressions for
RGA uncertainty bounds are deri®ed for 2 = 2 control problems and for general, n= n
control problems. Both worst-case bounds and statistical uncertainty bounds are de-
ri®ed. Se®eral simulation examples illustrate the new results. The RGA uncertainty
bounds pro®ide useful information concerning model accuracy requirements and the
robustness of decentralized control systems.

Introduction
RGA was introduced by Bristol Ž1966. as a measure of tribute to model uncertainty such as plantrmodel mismatch,
process interactions in multi-input, multi-output control changes in operating conditions, drift of physical parameters,
problems. Due to its simplicity and utility, the RGA analysis and so on. Thus, process models are never perfect. For un-
has been widely used to identify promising decentralized certain plant models, an RGA analysis based on a nominal
Žmulti-loop. control systems based on limited information, process model can lead to incorrect conclusions. Conse-
steady-state gains. The books by McAvoy Ž1983a. and quently, it is important to know how sensitive the RGA anal-
Shinskey Ž1988. describe many practical applications. Further ysis is to model uncertainty.
research has shown that the RGA is related to many funda- In this article, analytical expressions for RGA uncertainty
mental closed-loop system properties such as stability, ro- bounds are derived for two important classes of model uncer-
bustness, decentralized integral controllability, and failure tainty descriptions: worst-case bounds Žhard limits. and sta-
tolerance ŽGrosdidier et al., 1985; Yu and Luyben, 1987; Sko- tistically-based bounds Žsuch as three sigma limits.. New re-
gestad and Morari, 1987; Chiu and Arkun, 1991; Hovd and sults are developed for both 2=2 control problems and gen-
Skogestad, 1992; Chen et al., 1994; Zhu and Jutan, 1996; Lee eral n= n control problems.
and Edgar, 2000.. The related concept of relative disturbance
gain ŽRDG. has been proposed ŽStanley et al., 1985. and ex- Definition and Properties of RGA
tended ŽChang and Yu, 1992, 1994. in order to evaluate the
For an n= n system, Bristol Ž1966. defined each element
effect of disturbances. Because the original RGA analysis was
in the RGA as the ratio of the open-loop gain with all other
based on only steady-state information, it does not consider
loops open to the closed-loop gain with all other loops closed:
the dynamic behavior of the process in response to distur-
bances and set point changes. Consequently, dynamic ver-
Ž ⭸ yir⭸ u j . u k/ j
open-loop gain
sions of the RGA have been proposed ŽWitcher and McAvoy, ␭i j s s Ž1.
1997; Bristol, 1978; Tung and Edgar, 1981; Gagnepain and Ž ⭸ yir⭸ u j . y k/ j
closed-loop gain
Seborg, 1982; Jensen et al., 1986; McAvoy et al., 2001..
Although the process control literature is replete with ␭ i j is the relative gain between yi , the ith controlled variable,
analyses of RGA properties based on process models, the ef- and u j , the jth manipulated variable. The RGA matrix ⌳
fect of model uncertainty on RGA analysis has received little can be written in terms of the steady-state gain matrix K s
attention. There are many factors in practice that can con-  K i j 4, and the transpose of its inverse w Ky1 xT s  K Xi j 4

T
⌳ s  ␭ i j 4 s  K i j K Xi j 4 s Km w Ky1 x Ž2.
Correspondence concerning this article should be addressed to D. E. Seborg.

302 February 2002 Vol. 48, No. 2 AIChE Journal

where m denotes element-by-element multiplication. A dy- Analysis for 2=2 Control Problems
namic version of the RGA can be obtained by replacing the Consider a control problem with two controlled variables
steady-state gain matrix K in Eq. 2 with the corresponding and two manipulated variables. The steady-state gain matrix
transfer function matrix G Ž s ., and then substituting ss j ␻ for this 2=2 control problem is
ŽMcAvoy, 1983b; Grosdidier et al., 1985.. Thus the steady-
state RGA is the limiting case of the dynamic RGA as s™ 0. K 11 K 12
The RGA matrix ⌳ has the following properties ŽSeborg et Ks
K 21 K 22
Ž6.
al., 1989; Skogestad and Postlethwaite, 1996.:
Ž1. The sum of the elements in each row or column is one.
and the corresponding RGA matrix is
Ž2. The RGA elements are dimensionless and thus not af-
fected by choice of units or scaling of variables. ␭11 1y ␭11
Ž3. Any permutation of rows and columns in the K matrix ⌳s Ž7.
1y ␭11 ␭11
results in the same permutation in the RGA.
Both the steady-state and the dynamic RGA have these where
properties. In this article, for simplicity of notation, only the
steady-state RGA is considered. However, the results can 1
␭11 s Ž8.
easily be generalized to the dynamic RGA. 1y ␬
The effect of model uncertainty on RGA analysis has re- K 12 K 21
⌬
ceived surprisingly little attention. Grosdidier et al. Ž1985. ␬s Ž9.
derived an expression for the local sensitivity of ␭ i j to changes K 11 K 22
in the corresponding process gain K i j
and ␬ is referred to as the interaction quotient or interaction
measure ŽRijnsdorp, 1965.. Clearly, ␬ s1 is a singular point
d ␭i j dK i j
s Ž 1y ␭ i j . Ž3. for the RGA matrix.
␭i j Kij Because the RGA matrix depends on the steady-state gains,
model uncertainty in K produces uncertainties in the calcu-
In a derivation based on this expression, Yu and Luyben lated RGA. In this article, two types of model uncertainty
Ž1987. showed that the K will become singular if a single, descriptions are considered.
arbitrary element K i j is perturbed by the following amount
Worst-case bounds for steady-state gains
Kij Suppose that symmetric upper and lower bounds are avail-
⌬ K i j sy Ž4. able for each steady-state gain, that is, the unknown steady-
␭i j
state gain K i j varies within a range of " ⌬ K i j around a
nominal value, Kˆi j . The corresponding range for the interac-
This perturbation provides an upper limit on the maximum tion quotient ␬ will be denoted as
change in a single process gain that will maintain integral
controllability ŽYu and Luyben, 1987.. However, Eqs. 3 and 4 ␬ lF␬ F␬ h Ž 10.
are based on the restrictive assumption that only one process where
gain changes. Thus, these results are not applicable to the
more realistic situation where uncertainties are associated K 12 K 21
␬ 1 s min Ž 11.
with more than one process gain. Furthermore, the element- K 11 K 22
by-element uncertainties are usually correlated ŽSkogestad,
K 12 K 21
1992.. ␬ h s max Ž 12.
Grosdidier et al. Ž1985. have derived a more general result K 11 K 22
for integral controllability. If Kˆ is a nominal gain matrix and
I ⌳ I 1 and I ⌳ I ⬁ are matrix norms for the ⌳ Ž Kˆ., then in- If the range of ␬ in Eq. 10 does not include ␬ s1, then
tegral controllability will be preserved if the perturbed gain the range for ␭11 is
matrix K satisfies the following bound
1 1
F ␭11 F Ž 13.
1y ␬ l 1y ␬ h
I K y KˆI 1
F Ž5. If the range of ␬ does include ␬ s1, then the sign of ␭11
I KˆI 2 max Ž I ⌳ I 1 , I ⌳ I ⬁ .
changes at the singular point where ␬ s1 and the uncer-
tainty range for ␭11 is
While Eq. 5 is an important result, it does address the princi-
pal concern of this article: to what extent does a prescribed 1 1
degree of uncertainty in Kˆ result in uncertainty in the calcu- y⬁F ␭11 F and F ␭11 F⬁ Ž 14.
1y ␬ h
1y ␬ l
lated RGA and analysis based on it?
In this article, uncertainty bounds for the RGA matrix are Next, a transfer function model which has been used in
derived based on general classes of model uncertainty that many previous studies will be used to illustrate these results.
include correlated uncertainties for the elements of the n= n Example 1. Wood and Berry Ž1973. developed the follow-
gain matrix. ing empirical model of a pilot-scale distillation column that is

AIChE Journal February 2002 Vol. 48, No. 2 303

used to separate a methanol-water mixture Table 1. Comparison of Uncertainty Analysis Methods
␣ Uncertainty in All Elements Uncertainty in K 11 Only
XD Ž s . 12.8 eys y18.9 ey3 s RŽ s. 0.01 1.92 F ␭11 F 2.08 1.99F ␭11 F 2.03
16.7sq1 21 sq1 0.1 1.52 F ␭11 F 4 1.81F ␭11 F 2.21
s Ž 15. 0.25 y⬁F ␭11 Fy2.53 and 1.22 F ␭11 F⬁ 1.50 F ␭11 F 2.52
6.6 ey7s y19.4 ey3 s
XB Ž s . SŽ s.
10.9 sq1 14.4 sq1
makes ␬ h s1. Therefore, the upper bound for ␣ is given by
where X D and X B are the overhead and bottoms composi-
tions of methanol, respectively; R is the reflux flow rate and 0 F ␣ F 0.17 Ž 20.
S is the steam flow rate to the reboiler.
The nominal steady-state gain matrix is If the model uncertainty satisfies this constraint, the corre-
sponding ␭11 range is always greater than 0.5 and the RGA
12.8 y18.9 analysis suggests the same controller pairing.
K̂ s Ž 16.
6.6 y19.4 Table 1 compares the results of two uncertainty analyses
for Example 1. The results for Cases 1᎐3 are shown as ‘‘un-
The corresponding RGA matrix and interaction quotient ␬ certainty in all elements.’’ The results for ‘‘uncertainty in K 11
are only’’ were calculated using Eq. 3. Table 1 indicates that con-
sidering only uncertainties in the corresponding gain element
ˆs 2.01 y1.01
⌳ Ž 17. can produce overly optimistic estimates of the uncertainty in
y1.01 2.01 ␭ i j.
␬ˆ s 0.502 Ž 18. This example demonstrates that model uncertainty can
have a serious effect on control structure selection if the
which suggest the following controller pairing: X D y RrX B y standard RGA analysis is employed. Model uncertainty
S. should be incorporated into the RGA analysis in order to
Assume that the uncertainty bounds for the steady-state ensure the robustness of the decentralized control system.
gains are given by
Statistical description of uncertainty bounds
< ⌬ K i j < F ␣ < Kˆi j < Ž 19. Because worst-case situations usually have a low probabil-
ity of occurring, the RGA analysis of the previous section
Three different ␣ values will be considered. Because ␬ and l
may be very conservative for more typical operating condi-
␬ h can be calculated using Eqs. 11 and 12, the range of ␭11 tions that are much more probable. Thus, statistical descrip-
for each value of ␣ can be obtained. tions of model uncertainty are more reasonable. They can be
Case 1: ␣ s 0.01. According to Eqs. 11 and 12, the upper developed from a wide variety of sources, including statistical
and lower bounds for ␬ are ␬ l s 0.48 and ␬ h s 0.52, so that information for phenomenological model parameters, empiri-
the bounds for ␭11 are given by Eq. 13 cal model parameters, or the frequency response ŽCloud and
Kouvaritakis, 1987; Correa, 1989; Goodwin and Salgado,
1.92 F ␭11 F 2.08
1989; Stengel and Ryan, 1989; Schaper et al., 1992.. Addi-
tional process knowledge is often available in the form of
Thus, the recommended controller pairing from the RGA
engineering heuristics and information about the range of
analysis is X D y RrX B y S.
operating conditions. These sources of information should
Case 2: ␣ s 0.1. Now ␬ l s 0.34 and ␬ h s 0.75 and
also be considered when developing the uncertainty descrip-
1.52 F ␭11 F 4 tion.
Consider 2=2 systems where the nominal values of the
The suggested pairing is still X D y RrX B y S. steady-state gains and the covariance matrix are denoted by
Case 3: ␣ s 0.25. Now ␬ l s 0.18, ␬ h s1.39, and thus the
range of ␬ includes 1. From Eq. 14 it follows that Kˆ11 Kˆ12
K̂ s Ž 21.
Kˆ21 Kˆ22
y⬁F ␭11 Fy2.53 and 1.22 F ␭11 F⬁
⌬
Ý Ž K . s cov Ž K i j , K mn . Ž 22.
Thus, within the uncertainty range for ␣ s 0.25, the RGA
matrix becomes singular and ␭ i j changes sign. Consequently, ˆi j can be calculated from Eqs. 8 and 9
The nominal value ␭
no single controller pairing can be recommended for the en- ˆ
based on the nominal values of steady-state gain matrix K.
tire uncertainty range because both positive and negative val-
Because ␭ i j is a function of steady-state gains K kl , it can be
ues of ␭11 are possible.
approximated by introducing a Taylor’s series expansion and
In this example, if ␭11 ) 0.5 for the entire uncertainty range
truncating after the first-order terms
in Eq. 13, then the same controller pairing is recommended
by the RGA analysis. According to Eq. 8, it is easy to con- 2 2 ⭸␭ i j
clude that ␭11 is greater than 0.5 ify1- ␬ -1. For this ex- ␭ i j f ␭ i j < Kˆ q K kl y Kˆkl .
ample, ␬ l will not be equal to y1 for ␣ F1, while ␣ s 0.17
Ý Ý
k s1 ls1
ž /Ž
⭸ K kl K̂
Ž 23.

304 February 2002 Vol. 48, No. 2 AIChE Journal

 ˆi j s ␭ i j < K̂ . Then, it follows Eq. 23 that
Let ␭ Case 3: ␣ s 0.25. The variance and uncertainty of ␭11 are
given by
2
2
2 2 ⭸␭ i j
Ž ␭i j y ␭ˆi j . f Ž K kl y Kˆkl . ␴␭211 f 0.114
Ý Ý
k s1 ls1
ž / ⭸ K kl K̂
Ž 24.
or
Ž 34.

2 2 2 2 ⭸␭ i j
s Ý Ý Ý Ý
k s 1 l s 1 m s 1 ns 1
ž / ⭸ K kl K̂
0.995F ␭11 F 3.02 Ž 35.

⭸␭ i j For all three cases, the recommended controller pairing

=
ž / ⭸ K mn Kˆ
Ž K kl y Kˆkl .Ž K mn y Kˆmn . Ž 25 . Ž y 1 y u1ry 2 y u 2 . is valid for the entire model uncertainty
range. Comparison of Examples 1 and 2 indicates that the
results for the hard uncertainty bounds are much more con-
Thus, the variance of ␭ i j can be approximated as
servative, because the worst-case situation is considered, while
2 in practice, it may have a very low probability of actually oc-
␴␭2i j s E Ž␭ ˆ
i j y ␭i j . Ž 26. curring.
2 2 2 2 ⭸␭ i j ⭸␭ i j
Model uncertainty descriptions from process identification
f Ý Ý Ý Ý
k s 1 l s 1 m s 1 ns 1
ž /ž /
⭸ K kl Kˆ ⭸ K mn Kˆ
cov Ž K kl , K mn .
Statistically-based model uncertainty descriptions are often
Ž 27. generated when process identification techniques are em-
ployed to develop empirical process models from input-out-
where E denotes the expectation operator. Thus, a statistical put data. For example, commercial software packages pro-
description of the uncertainty bounds for ␭ i j has been de- duce confidence intervals for estimated model parameters
rived based on the nominal value and covariance of K. andror the frequency response of the model. However, a de-
The Wood-Berry distillation column model is used again to tailed description of these identification methods and the re-
demonstrate the utility of statistically-based uncertainty sponse uncertainty description is beyond the scope of this ar-
bounds. ticle. The objective of this article is to derive the uncertainty
Example 2. Assume that the steady-state gains are re- description of the RGA based on model uncertainty descrip-
ported as Kˆi j "3 ␴K i j where the standard deviation for each tions which have already been obtained.
steady-state gain is given by The following analysis is valid for a wide class of linear,
dynamic stochastic models that includes state-space models
1 and auto-regressive, integrated moving average ŽARIMA.
␴K i j s cov Ž K i j , K i j . s ␣ < Kˆi j <
' Ž 28. models. However, for purposes of illustration, we will con-
3
sider a particular class of models, finite impulse response
and each gain is uncorrelated, that is, the covariance between ŽFIR. models with stochastic disturbances
different gains is zero. A comparison of Eqs. 19 and 28 indi-
cates that 3 ␴K i j corresponds to the range of uncertainty for y Ž t . sG Ž q,⌰ . u Ž t . q H Ž q,⌰ . e Ž t . Ž 36.
the worst-case analysis of the previous section. This choice
was made so that the results obtained using worst-case bounds where
and statistical bounds can be compared. Next, three values of Nu
y Ž t . , e Ž t . g R n y , uŽ t . g R and
␣ will be considered.
Case 1: ␣ s 0.01. By using Eq. 27, the variance of ␭11 is ⌰ s w ␪ 1 , . . . , ␪n x
T
Ž 37.
obtained ⬁

␴␭211 f 0.000183 Ž 29.

G Ž q,⌰ . s Ý g Ž k ,⌰ . qyk Ž 38.
k s1
⬁
Therefore, the 3 ␴ limits for ␭11 can be represented as H Ž q,⌰ . s1q Ý h Ž k ,⌰ . qyk Ž 39.
k s1
ˆ11 "3 ␴␭11
␭11 s ␭ Ž 30.
Here, eŽ t . is a sequence of independent random variables
or with zero means and the same variance ␴ 2. Assume that ⌰
s w ␪ 1, . . . , ␪n xT is the vector of model parameters to be iden-
1.97 F ␭11 F 2.05 Ž 31. tified and that ⌰ ˆ denotes an asymptotically normal and unbi-
ased estimate of ⌰. The estimated covariance matrix ÝŽ ⌰ .
Case 2: ␣ s 0.1. The variance and uncertainty of ␭11 are of the parameter estimation error is assumed to be obtained
given by during process identification. Therefore
␴␭211 f 0.0183 Ž 32. T
ˆ y ⌰ . Ýy1 Ž ⌰ . Ž ⌰
Ž⌰ ˆ y ⌰ . ™ ␹ q2 Ž 40.
or
where q is the number of model parameters and ␹ q2 is the
1.60 F ␭11 F 2.42 Ž 33. chi-square distribution with q degrees of freedom. This result

AIChE Journal February 2002 Vol. 48, No. 2 305

enables the construction of confidence bounds The Ž1y ␣ .=100% confidence intervals can be derived
from Lemma 1 with Eq. 41, 47 and 48 as
T
ˆ y ⌰ . Ýy1 Ž ⌰ . Ž ⌰
P Ž⌰ ˆ y ⌰ . F ␹ a,2 q s1y ␣ Ž 41. y1
⌬ K T Ž ␻ ,⌰ . Ž K ⌰ Ý Ž ⌰ . K ⌰
T
. ⌬ K Ž ␻ ,⌰ . F ␹␣2, q Ž 51.
where 1y ␣ is the probability that the true parameters lie
´ ⌬ K T Ž ␻ ,⌰ . Ýy1 Ž K Ž ␻ ,⌰ . . ⌬ K Ž ␻ ,⌰ . F ␹␣2, q Ž 52.
within the specified ellipsoidal confidence region and ␹ a,2 q is
the upper ␣ percentage point of ␹ q2. Let K i j Ž ␻ ,⌰ . be the where K Ž ␻ s 0,⌰ . is the steady-state gain matrix. For sim-
magnitude of the transfer function from input i to output j plicity of notation, assume ␻ s 0 and omit the ⌰ depend-
of a 2=2 system and K Ž ␻ ,⌰ . s w K i j Ž ␻ ,⌰ .x 2=2 . The sensitiv- ence. Thus K and Kˆ denote the steady-state gain matrix and
ity of the function K i j Ž ␻ ,⌰ . to the parameters ⌰ is given by its estimate.
the gradient Equation 23 for the steady-state RGA element ␭ i j can be
⭸ expressed as
K i j,⌰ Ž ␻ ,⌰ . s K i j Ž ␻ ,⌰ . Ž 42.
⭸ ⌰T ˆi j f ␭ i j, K ⌬ K
␭i j y ␭ Ž 53.
which is considered to be a 1= n row vector. Define a 4= n ⭸␭ i j ⭸␭ i j ⭸␭ i j ⭸␭ i j
matrix K ⌰ as where ␭ i j, K s . Therefore, the
⭸ K 11 ⭸ K 21 ⭸ K 12 ⭸ K 22
K 11,⌰ Ž ␻ ,⌰ . Ž1y ␣ .=100% confidence intervals for ␭ i j can be derived by
K 21,⌰ Ž ␻ ,⌰ . using Lemma 1 with Eqs. 27 and 52
K⌰s Ž 43.
K 12,⌰ Ž ␻ ,⌰ . T y1
Ž␭ ˆ
. Ž ␴␭ . Ž ␭ y ␭ˆ . F ␹␣
i j y ␭i j
2
ij ij
2
,q Ž 54.
K 22,⌰ Ž ␻ ,⌰ . ij

2
´Ž␭ y␭ ˆ . F ␹␣ ␴␭
ij ij
2
,q
2
Ž 55.
ˆ . can be approximated by
The error in estimation of K i j Ž ␻ ,⌰ ij

a truncated Taylor series expansion The extension of these confidence intervals to the dynamic
RGA is straightforward because frequency-dependent
ˆ . y K i j Ž ␻ ,⌰ . f K i j,⌰ Ž ␻ ,⌰ . Ž ⌰
K i j Ž ␻ ,⌰ ˆ y⌰. Ž 44. K Ž ␻ ,⌰ . and Ýw K Ž ␻ ,⌰ .x are known.
A simulation example is considered to demonstrate how
The covariance matrix of K Ž ␻ ,⌰ . can be expressed as
RGA confidence intervals can be established based on pro-
⌬
cess identification.
Ý Ž K . s cov Ž vec Ž K . . s E ž vecŽ K .yvecŽ Kˆ. vec Ž K . Example 3. A simplified model for the Shell Heavy Oil
T Fractionator problem has been reported by Adusumilli et al.
yvec Ž Kˆ. / Ž 45. Ž1998.

where E denotes the expectation operator and vecŽ ⴢ . is an y1 4.05ey27s 1.77ey28 s u1 5.88 ey27s
operator that stacks the columns of a matrix in a column vec- 50 sq1 60 sq1 50 sq1
tor. For example, if a matrix As w a1 a 2 x n=2 , then s q d
5.39 ey18 s 5.72 ey14 s 6.90 ey15 s
y2 u2
a1 50 sq1 60 sq1 40 sq1
vec Ž A . s Ž 46.
a2 Ž 56.
⌬
ˆ .y K Ž ␻ ,⌰ .x4=1. Then Eq. 42
Let ⌬ K Ž ␻ ,⌰ . s vecw K ␻ ,⌰ where y 1 and y 2 are the top and side end points, respec-
can be written as tively; u1 and u 2 are top and side draws, respectively.
The steady-state gain matrix is
ˆ y⌰.
⌬ K Ž ␻ ,⌰ . f K ⌰ Ž ⌰ Ž 47.
4.05 1.77
Ks Ž 57.
by a Taylor’s series expansion. Therefore, the covariance ma- 5.39 5.72
trix ÝŽ K Ž ␻ ,⌰ .. can be expressed by ŽLjung, 1985.
The corresponding ␭11 value is ␭11 s1.70 and the recom-
Ý Ž K Ž ␻ ,⌰ . . f K ⌰ Ý Ž ⌰ . K ⌰
T
Ž 48. mended control variable pairing is y 1 y u1ry 2 y u 2 .
A commercial software package ADAPTX ŽLarimore, 1996.
Lemma 1 (Wahlberg and Ljung, 1992). Let x g R n, Ýg was used for the process identification. It employs a subspace
n=n
R be positive definite, and assume that identification technique, Canonical Variate Analysis ŽCVA.,
to identify a linear, stochastic state space model and calcu-
x T Ýy1 x F1 Ž 49. late the frequency domain response of the model. An uncer-
tainty description in the form of 95% confidence limit is also
Consider y s Ax, where y g R p, pF n, Ag R p=n and A has calculated.
full row rank. Then A MIMO PRBS input signal was designed as per
y1 Adusumilli et al. Ž1998. and used as the excitation signal for
y T Ž AÝ AT . y F1 Ž 50. process identification. The sampling period was chosen as ⌬ t

306 February 2002 Vol. 48, No. 2 AIChE Journal

Figure 1. SNR =3: frequency response with 95% confi- Figure 2. SNR =10: frequency response with 95% confi-
dence bands. dence bands.

s 4 min. By using the ADAPTX software for CVA identifi- <cov Ž K 12 , K 21 . < F 0.116, <cov Ž K 12 , K 22 . < F 0.123,
cation with a 2,500-point data set, a CVA based model was
identified. Thus, the frequency-dependent gain matrix and its <cov Ž K 21 , K 22 . < F 0.136 Ž 63.
statistical uncertainty descriptions could be obtained. Then,
the estimated RGA values and its uncertainty bounds were Then, the upper bound for ␴␭211 can be obtained from Eq. 27
calculated.
Case 1. The signal-to-noise ratio for identification is SNR
s 3. ␴␭211 F 0.0495. Ž 64.
Figure 1 shows the magnitude and 95% confidence inter-
vals of the frequency response for the identified model. The For ␣ s 0.05, ␹ 0.05,64 s9.13. Then using Eq. 55, the 95% con-
solid line is the frequency response of the identified models fidence bands for ␭11 are
and the dashed lines represent the 95% confidence bands.
The dash-dot line is the frequency response of the true sys-
tem. The estimated steady-state gains K i j and the variances y0.27 F ␭11 F 3.79 Ž 65.
from the CVA identification are
For this range of possible ␭11 values, the recommended con-
4.45 1.94 trol variable pairing is ambiguous. The y0.27 F ␭11 - 0.5
K̂ s Ž 58.
5.86 5.93 range suggests the y 1 y u 2ry 2 y u1 pairing, while the y 1 y
u1ry 2 y u 2 pairing is suggested for 0.5- ␭11 F 3.79. In view of
␴K211 s 0.0933, ␴K212 s 0.105, ␴K221 s 0.129, ␴K222 s 0.145 Ž 59 . ˆ11 s1.76, it seems more appropriate to
the nominal value of ␭
select the y 1 y u1 pairing. However, a more accurate identi-
The nominal value of ␭11 can be calculated using Eq. 8 fied model would be desirable in order to resolve this ambi-
guity.
ˆ11 s1.76
␭ Ž 60. Case 2. The signal-to-noise ratio is SNRs10.
The identified frequency responses together with 95% con-
The Cauchy-Schwartz inequality ŽTaylor, 1982. fidence bands are shown in Figure 2. Also, the solid line is
the frequency response of the identified model and the
<cov Ž x, y . < F ␴x ␴y Ž 61. dashed lines are the 95% confidence bands. The dash-dot
line is the frequency response of the true system.
gives the upper bounds for the covariances cov Ž K i j , K mn . The steady-state gain matrix and the variance for each ele-
ment are
<cov Ž K 11 , K 12 . < F 0.099, < co® Ž K 11 , K 21 . < F 0.110,
4.03 1.64
<cov Ž K 11 , K 22 . < F 0.116 Ž 62. K̂ s
5.37 5.57
Ž 66.

AIChE Journal February 2002 Vol. 48, No. 2 307

 ␴K211 s 0.0253, ␴K212 s 0.0272, ␴K221 s 0.0348, ␴K222 s 0.0375 Here, K i j, k l denotes the submatrix of K with rows i and k
and columns j and l removed. For the special case of is k
Ž 67. and js l, Eq. 73 is identical to Eq. 3 in Grosdidier et al.
Ž1985..
Therefore, the nominal value of ␭11 and its variance are Next, several different uncertainty descriptions for n= n
problems are considered.
ˆ11 s1.64
␭ Ž 68.
Worst-case bounds
␴␭211 F 0.0121 Ž 69.
Suppose that the uncertainty region for the steady state is
given by Kˆi j " ⌬ K i j . Then, the corresponding uncertainty
The corresponding 95% confidence interval for ␭11 is bound for ␭ i j is

0.67 F ␭11 F 2.62 Ž 70. n n ⭸␭ i j

<␭
ˆi j y ␭ i j < F Ž ⌬ K kl .
and the recommended RGA pairing is y 1 y u1ry 2 y u 2 for
Ý Ý
k s1 ls1
ž / ⭸ K kl
Ž 74.

the entire range.

Analysis for n= n Systems Statisitical uncertainty description

Assume that K, ˆ the nominal steady-state gain matrix, and
The results obtained for 2=2 systems will now be extended
ÝŽ K ., the covariance matrix for each steady-state gain are
to n= n systems. For a n= n system with the steady-state
known where
gain matrix

Ks Kij Ž 71. ⌬
n= n Ý Ž K . s cov Ž vec Ž K . . s E ž vecŽ K .yvecŽ Kˆ. vec Ž K .

the i, jth element of the RGA matrix ⌳ is ŽGrosdidier et al., T

1985. yvec Ž Kˆ. / Ž 75.

K i j det Ž K i j .
␭ i , j s Ž y1 .
iq j
Ž 72. The nominal value of each ␭ i j can be obtained from Eq. 72
det Ž K . with Kˆ replacing K. The variance is given by

Here, K i j denotes the submatrix that remains after the ith T

row and jth column of K are deleted. The partial derivative ␴␭2i j s Ž ␭ i j, K . KˆÝ Ž K . Ž ␭ i j, K . Kˆ Ž 76.
of ␭ i j with respect to K kl is given by

where
⭸␭ i j
⭸ K kl
⭸␭ i j ⭸␭ i j ⭸␭ i j
␭ i j, K s ⭈⭈⭈ Ž 77.
°␭ Ž 1y ␭ .
ij ij
is k and js l
⭸ K 11 ⭸ K 21 ⭸ K nn
Kij
␭ i j ␭ kl denotes the row vector of the partial derivatives of ␭ i j with
s y~ K kl
is k or js l
respect to each element of the steady-state gain matrix K.
Equation 76 is the extension to n= n systems of the covari-
Ž y1. iq jq k q l K i j det Ž K i j, k l . ␭ i j ␭ kl ance formula Eq. 27 for 2=2 systems.
¢ det Ž K .
y
K kl
i/ k and j/ l
Example 4. Ogunnaike et al. Ž1983. considered a transfer
function model for a pilot scale, binary distillation column
Ž 73. used to separate ethanol and water

0.66 ey2.6 s y0.61ey3 .5 s y0.0049 eys u1

y1
6.7sq1 8.64 sq1 9.06 sq1
y2 1.11ey6.5 s y2.36 ey3 s y0.012 ey1 .2 s
s u2 Ž 78.
3.25sq1 5sq1 7.09 sq1
y33.68 ey9.2 s 46.2 ey9 .4 s 0.87 Ž 11.61 sq1 . eys
y3
8.15sq1 10.9 sq1 Ž 3.89 sq1.Ž 18.8 sq1. u3

308 February 2002 Vol. 48, No. 2 AIChE Journal

where y 1 is overhead mol fraction of ethanol, y 2 is the mol Based on the uncertainty ranges of ␭ i j , the only appropri-
fraction of ethanol in the side stream, and y 3 is temperature ate pairing for the entire uncertainty ranges is: y 1 y u1ry 2 y
on Tray 19; u1 is overhead reflux flow rate, u 2 is side stream u 2ry 3 y u 3.
draw-off rate, and u 3 is reboiler steam pressure. By calculating the uncertainty bounds of the RGA matrix
The nominal steady-state gain matrix is given by for the entire uncertainty range of the plant model, we can
check whether a single control structure selection is appro-
0.66 y0.61 y0.0049 priate for all possible plants. Therefore, the robustness of the
K̂ s 1.11 y2.36 y0.012 Ž 79. control structure is achieved.
y33.68 46.2 0.87

Therefore, the nominal RGA matrix is Conclusions

Analytical expressions for RGA uncertainty bounds have
1.945 y0.673 y0.272 been derived for 2=2 and n= n control problems. Two char-
ˆ s y0.664
⌳ 1.899 y0.235 Ž 80. acterizations of process model uncertainty have been consid-
y0.281 y0.225 1.506 ered: worst case bounds and the types of statistical descrip-
tions that are readily obtained during process identification.
Several simulation examples illustrate the new results.
This RGA matrix suggests the diagonal controller pairing. The new RGA uncertainty bounds provide useful informa-
Assume that the uncertainty for each steady-state gain can tion concerning the recommended control structure and the
be expressed as uncertainty associated with the recommendation. Further-
more, RGA uncertainty analysis can be used to estimate the
< ⌬ K i j < F ␣ < Kˆi j < Ž 81. maximum degree of uncertainty in the plant model that will
not affect the recommended controller pairing. This useful

Case 1: ␣ s 0.01. The uncertainty range for ⌳ is

1.92 F ␭11 F1.97 y0.698F ␭12 Fy0.650 y0.291F ␭13 Fy0.253

y0.689F ␭21 Fy0.641 1.87 F ␭22 F1.93 y0.246 F ␭23 Fy0.223 Ž 82.
y0.300 F ␭31 Fy0.262 y0.237 F ␭32 Fy0.214 1.50 F ␭33 F1.52

The recommended RGA controller pairing is obvious and

unambiguous: y 1 y u1ry 2 y u 2ry 3 y u 3.
Case 2: ␣ s 0.1. The uncertainty range for ⌳ is

1.65F ␭11 F 2.24 y0.914 F ␭12 Fy0.434 y0.459F ␭13 Fy0.085

y0.902 F ␭ 21 Fy0.641 1.57 F ␭22 F 2.23 y0.351F ␭23 Fy0.118 Ž 83.
y0.470 F ␭31 Fy0.092 y0.342 F ␭32 Fy0.109 1.39F ␭33 F1.62

Again, the preferred controller pairing is: y 1 y u1ry 2 y u 2ry 3

y u 3.
Case 3: ␣ s 0.25. Now the uncertainty range for ⌳ is

1.21F ␭11 F 2.97 y1.56 F ␭12 F 0.167 y0.739F ␭13 F 0.382

y1.54 F ␭ 21 F 0.168 1.07 F ␭22 F 3.06 y0.668F ␭23 F 0.173 Ž 84.
y0.754 F ␭31 F 0.380 y0.654 F ␭32 F 0.182 1.22 F ␭33 F1.91

The same controller pairing is suggested for this case.

Case 4: ␣ s 0.5. The uncertainty range for ⌳ is

0.477 F ␭11 F 3.41 y1.87 F ␭12 F 0.527 y1.21F ␭13 F 0.662

y1.85F ␭ 21 F 0.524 0.244 F ␭22 F 3.55 y0.818F ␭23 F 0.348 Ž 85.
y1.23F ␭ 31 F 0.664 y0.807 F ␭32 F 0.356 0.931F ␭33 F 2.08

AIChE Journal February 2002 Vol. 48, No. 2 309

information provides insight concerning the required accu- Ljung, L., ‘‘Asymptotic Variance Expressions for Identified Black-box
racy of the plant model and how much the operating condi- Transfer Function Models,’’ IEEE Trans. Auto. Control, AC-30, 834
Ž1985..
tion can change before the controller pairing decision be- McAvoy, T. J., Interaction Analysis, Instrument Society of America,
comes ambiguous. Research Triangle Park, NC Ž1983a..
Although only the steady-state RGA has been considered McAvoy, T. J., ‘‘Some Results on Dynamic Interaction Analysis of
in this article, all of the results can be directly extended to Complex Control Systems,’’ Ind. Eng. Chem. Process Des. De®., 22,
42 Ž1983b..
the frequency-dependent RGA matrix in order to gain infor-
McAvoy, T. J., Y. Arkun, R. Chen, D. Robinson, and P. D. Schnelle,
mation about dynamic behavior. ‘‘A New Approach to Defining a Dynamic Relative Gain,’’ Proc.
IFAC DYCOPS-6 Symp., 481, Chejudo Island, Korea Ž2001..
Ogunnaike, B. A., J. Lemaire, M. Morari, and W. H. Ray, ‘‘Ad-
Literature Cited vanced Multivariable Control of a Pilot Plant Distillation Column,’’
Adusumilli, S., D. E. Rivera, S. Dash, and K. Tsakalis, ‘‘Integrated AIChE J., 29, 632 Ž1983..
MIMO Identification and Robust PID Controller Design Through Rijnsdorp, J. E., ‘‘Interaction in Two-variable Control Systems for
Loop Shaping,’’ Proc. 1998 Amer. Control Conf., 1230, Philadel- Distillation ColumnsᎏI and II,’’ Automatica, 1, 15 Ž1965..
phia, PA Ž1998.. Schaper, C. D., D. E. Seborg, and D. A. Mellichamp, ‘‘Probabilistic
Bristol, E. H., ‘‘On a New Measure of Interactions for Multivariable Approach to Robust Process Control,’’ Ind. Eng. Chem. Res., 31,
Process Control,’’ IEEE Trans. Auto. Control, AC-11, 133 Ž1966.. 1694 Ž1992..
Bristol, E. H., ‘‘Recent Results on Interactions for Multivariable Seborg, D. E., T. F. Edgar, and D. A. Mellichamp, Process Dynamics
Process Control,’’ AIChE Meeting, Miami, FL Ž1978.. and Control, Wiley, New York Ž1989..
Chang, J. W., and C. C. Yu, ‘‘Relative Disturbance Gain Array,’’ Shinskey, F. G., Process Control Systems, McGraw-Hill, New York
AIChE J., 38, 521 Ž1992.. Ž1988..
Chang, J. W., and C. C. Yu, ‘‘Synthesis of Controller Structures for Skogestad, S., ‘‘Robust Control,’’ Practical Distillation Control, W. L.
Robust Load Performance,’’ Int. J. Control, 60, 1353 Ž1994.. Luyben, ed., Van Nostrand Reinhold, New York, 291 Ž1992..
Chen, J., J. S. Freudenberg, and C. N. Nett, ‘‘The Role of the Condi- Skogestad, S., and M. Morari, ‘‘Implications of Large RGA Ele-
tion Number and the Relative Gain Array in Robustness Analysis,’’ ments on Control Performance,’’ Ind. Eng. Chem. Res., 26, 2323
Automatica, 30, 1029 Ž1994.. Ž1987..
Chiu, M.-S., and Y. Arkun, ‘‘A New Result on Relative Gain Array, Skogestad, S., and I. Postlethwaite, Multi®ariable Feedback Control:
Niederlinski Index and Decentralized Stability Condition: 2=2 Analysis and Design, Wiley, New York Ž1996..
Plant Cases,’’ Automatica, 27, 419 Ž1991.. Stanley, G., M. Marino-Galarraga, and T. J. McAvoy, ‘‘Shortcut Op-
Cloud, D. J., and B. Kouvaritakis, ‘‘Weighting Sequences, Optimal erability Analysis. 1. The Relative Disturbance Gain,’’ Ind. Eng.
Truncation, and Optimal Frequency-Response Uncertainty Chem. Process Des. De®., 24, 1181 Ž1985..
Bounds,’’ IEE Proc. Part D, 134, 153 Ž1987.. Stengel, R. F., and L. E. Ryan, ‘‘Multivariable Histograms for Analy-
Correa, G. O., ‘‘A System Identification Problem Motivated by Ro- sis of Linear Control System Robustness,’’ Proc. 1989 Amer. Con-
bust Control,’’ Int. J. Control, 50, 575 Ž1989.. trol Conf., 937, Pittsburgh, PA, Ž1989..
Gagnepain, J. P., and D. E. Seborg, ‘‘Analysis of Process Interactions Taylor, J. R., An Introduction to Error Analysis, University Science
with Applications to Multiloop Control Systems Design,’’ Ind. Eng. Books, Mill Valley, CA Ž1982..
Chem. Process Des. De®., 21, 5 Ž1982.. Tung, L. S., and T. F. Edgar, ‘‘Analysis of Control-Output Interac-
Goodwin, G. C., and M. E. Salgado, ‘‘A Stochastic Embedding Ap- tions in Dynamic Systems,’’ AIChE J., 27, 690 Ž1981..
proach for Quantifying Uncertainty in the Estimation of Re- Wahlberg, B., and L. Ljung, ‘‘Hard Frequency-Domain Model Error
stricted Complexity Models,’’ Adapt. Control Signal Process, 3, 333 Bounds from Least-squares Like Identification Techniques,’’ IEEE
Ž1989.. Trans. Auto. Control, 37, 910 Ž1992..
Grosdidier, P., M. Morari, and B. R. Holt, ‘‘Closed-loop Properties Witcher, M. F., and T. J. McAvoy, ‘‘Interacting Control Systems:
from Steady-state Gain Information,’’ Ind. Eng. Chem. Fund., 24, Steady State and Dynamic Measurement of Interaction,’’ ISA
221 Ž1985.. Trans., 16, 35 Ž1977..
Hovd, M., and S. Skogestad, ‘‘Simple Frequency-dependent Tools Wood, R. K., and M. W. Berry, ‘‘Terminal Composition Control of a
for Control System Analysis, Structure Selection and Design,’’ Au- Binary Distillation Column,’’ Chem. Eng. Sci., 28, 1707 Ž1973..
tomatica, 28, 989 Ž1992.. Yu, C. C., and W. L. Luyben, ‘‘Robustness With Respect to Integral
Jensen, N., D. G. Fisher, and S. L. Shah, ‘‘Interaction Analysis in Controllability,’’ Ind. Eng. Chem. Res., 26, 1043 Ž1987..
Multivariable Control Systems,’’ AIChE J., 32, 959 Ž1986.. Zhu, Z.-X., and A. Jutan, ‘‘RGA as a Measure of Integrity for De-
Larimore, W. E., ADAPTX Users Manual, Adaptics Inc., Mclean, VA centralized Control Systems,’’ Trans. IChemE, 74 Part A, 35 Ž1996..
Ž1996..
Lee, J., and T. F. Edgar, ‘‘Computational Method for Decentralized
Integral Controllability of Low Dimensional Processes,’’ Comput.
Chem. Eng., 24, 847 Ž2000.. Manuscript recei®ed May 4, 2001, and re®ision recei®ed Aug. 30, 2001.

310 February 2002 Vol. 48, No. 2 AIChE Journal