Assessment of Control Loop Performance

THOMAS J . HARRIS

Department of Chemical Engineering, Queen 's University, Kingston, Ontario, K7L 3N6
For processes described by linear transfer functions with additive disturbances, the best possible control in the mean
square sense is realized when a minimum variance controller is implemented. It is shown that an estimate of the best
possible control can be obtained by fitting a univariate time series to process data collected under routine control. No
'identifiability' constraints need be imposed. The use of this technique is demonstrated with pilot plant and production data.

Pour les procCdCs dCcrits par des fonctions de transfert IinCaires avec des perturbations additives, le meilleur contr6le
possible au sens des moindres carrCs s'obtient avec un rCgulateur il variance minimum. On montre qu'on peut obtenir
une estimation du meilleur contrble possible en adaptant une sCrie chronologique univariCe pour traiter des donnCes
recueillies lors d'un rCtrocontr6le de routine. II n'est pas nCcessaire d'imposer des contraintes d'.identificabilitC>>.L'utili-
sation de cette technique est dCcrite par des donnCes d'une usine pilote et des donnCes de production.

Keywords: minimum variance control, autocorrelation, time series analysis, control loop performance.

Y, = w ( z - ' ) / h ( ~ - ' ) U , - +
T here are many techniques available for the design of
feedback control strategies. These methods may be very
simple, such as those used in statistical process control, or
~ D, . . . . . . . . . . . . . . . . (1)

w ( z - ' ) and h ( z - ' ) are polynomials of order ( T , s )

they may be more elaborate strategies, such as pole place-
ment designs and linear quadratic controllers. Regardless of
respectively in the backward shift operator z - which ',
is defined such that z-' Y, = Y,- I . b is the number of whole
the control strategy, it is important to have some benchmark periods of delay in the process and is calculated as
against which its performance can be evaluated. The theo-
retical 'best achievable' control, as measured by the mean
square error, is such a benchmark. If the theoretical best b= 1 +f= 1 + integer ( T ~ / T.). . . . . . . . . . . . . ( 2 )
achievable control represents a significant improvement over
the current performance, alternate controller tuning or feed- T is the control interval and T~ is the process delay arising
back control strategies can be considered if this improved from true process deadtime or analysis delay. f is the
performance is warranted. However, the best achievable per- number of integer periods of delay. Only processes which
formance itself may not be adequate. In these cases alter- are open loop stable are considered.
nate approaches, such as feedforward control, reduction of The process disturbance D,, is represented by an
deadtime and different loop pairings must be used to achieve Autoregressive-Integrated-Moving-Average (ARIMA) time
a reduction in variability. series model of order (p,d,q), i.e., a linear filter driven by
The purpose of this paper is to describe a very simple tech- white noise:
nique for ascertaining the best theoretically achievable feed-
back control performance as measured by the output mean
square error. A univariate time series model is tit to process
data. From this time series model, the best control perfor-
mance can be estimated if the number of whole periods of (a,) is a sequence of independently and identically
delay is known. An important feature of this method is that distributed normal random variables with mean zero and
it is not necessary to 'perturb' the process with extraneous constant variance. v is an abbreviation for 1 - z - ' . The
test signals. Data may be used from any linear process which monic polynomials 4 ( z - ' ) and O(z-') are of order p
is operating under normal closed loop control with a linear, and q respectively, and are stable. The presence of d > 0
time-invariant feedback control strategy. roots on the unit circle allows for nonstationary behavior
The outline of this paper is as follows. In the first section in the mean value of the disturbance. The observations are
the process description is given. This is followed by a brief collected at equispaced control intervals. [ Y,, U,) repre-
derivation of a minimum variance controller, and of the sent deviations of the process and manipulated variables from
resulting properties of a process operating under minimum var- some reference values. The reference value for { Y,]
iance control. It is then shown that the theoretical best achiev- is taken as the target value or setpoint for closed loop iden-
able performance in the mean square sense can be estimated tification, and is taken as the mean value in open loop
from process data collected under 'normal' closed loop condi- situations.
tions. A method for estimating this mean square performance Time series models provide a very flexible means of
is discussed. This is followed by a Monte-Carlo simulation to modeling process dynamics and disturbances. These distur-
illustrate the statistical properties of the estimator. Finally, this bances may be of a meandering, drifting type nature, or they
technique is applied to pilot plant and production data. may be more deterministic in nature, such as randomly occur-
ring steps, ramps, or exponential rises to a new level (Mac-
Gregor et al., 1984; Astrom and Wittenmark, 1984).
Minimum Variance Control When the model Equations (1) and (2) are known, a con-
troller can be designed to minimize the variance of Y,.
The process is described by the superposition of a discrete (Recall that Y, is the deviation of the process variable from
dynamic model relating the controlled variable Y,, to the its setpoint). The resulting linear controller is (Astrom, 1970;
manipulated variable U,, and a disturbance D,. Box and Jenkins, 1970):

856 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 67, OCTOBER, 1989
 The relationship between minimum variance controllers
(4) and Proportional-Integral-Derivative (PID) controllers and
other strategies such as Dahlin's algorithm are discussed by
where D, f b l f is the minimum variance forecast for D, + h Palmor and Shinnar (1979) and by Harris et al. (1981). The
given information only available to time r. This forecast is relationship between minimum variance controllers and
obtained by writing D,+h as the sum of a minimum variance methods used in statistical process control is discussed by
forecast D,+bi,and a prediction error e f + h l rwhich
, is Box et al. (1974). In these papers, it is shown that many feed-
uncorrelated with the forecast, i.e., back control algorithms have the same structure as minimum
variance controllers. Although a particular controller, such
as a PID algorithm, may not have been designed to explicitly
Dr+b = er+b/[ + Dt+blr . . . . . . . . . . . . . . . . . . . . . . (5)
minimize the variance of Y, it may still give performance
The minimum variance forecast and forecast error are uni- which is very close to that of the minimum variance con-
quely determined by solving the following Diophantine equa- troller. This is readily verified by examining the autocorre-
tion for the polynomials $ ( z - ' ) and T ( 2 - I ) : lation function of the process variable. The autocorrelation
function of the data, commonly known as the correlogram,
e(z-') = $ ( z - ' ) + ( z - ' ) V d + z - ' f + ' ) T ( z - ' ) . . ' (6) and the Fourier transform of the autocorrelation function,
which is the spectrum of the data, also provide valuable infor-
$(z-') is a monic polynomial of order f, and T ( z - ' ) is mation for diagnosing the sources of process disturbances
a polynomial of order max [q-f-1, p + d - 11. With this nota- and diagnosing control loop performance (Pryor, 1982;
tion, the minimum variance forecast and forecast error are: Ohtsu and Kitagawa, 1984; DeVries and Wu, 1978).
It may not be possible to implement a minimum variance
controller due to the fact that (i) excessive control action is
(7) required or (ii) the numerator of the process dynamics
w ( z - ' ) has a zero, whose magnitude, in the z-plane exceeds
= D,+b - Dr+blr= $ ( Z - ' ) U r + b . . . . . . . . . (8) one. The former typically arises when: i) the sampling
interval is short relative to the dominant process time cons-
$ ( z - ' ) can also be obtained from the first f + 1 terms of the tants or ii) the fractional period of delay is significant. In
long division of + ( z - ' ) v into O ( z - ' ) . T ( z - ' ) is the such cases, constraining the control effort is required, which
remainder term. The linear control law, given by Equation in turn inflates the output variance. However, the variance
(4) provides the best control in the mean square sense when estimate given by Equation (10) provides a theoretical lower
the ( a t ' s ) are normally distributed. When the ( a , ' s ] are not bound on the output variance under linear feedback control.
distributed normally, Equation (4) still gives the best mean
square performance, amongst all linear controllers, (Astrom, Estimation of Minimum Variance Performance
1970).
When it is concluded that the existing control strategy is
Testing for Minimum Variance Control not giving minimum variance performance, it is of interest
then to have a technique for estimating the best possible per-
Under minimum variance control, the process output is formance. Ideally, this technique should be simple. In addi-
the error in forecasting the disturbance, i.e., D,+b - tion, it is desirable to use data collected under conditions
D,+b/,. This error is a moving average time series model which represent normal or routine process operation, i.e.,
of order f, where there are no external inputs used to perturb the process.
The following analysis shows when it is possible to do this.
r, = $ ( z - I ) u , = (1 + $'z-' + . . . + $,fz-f)u, . . . (9) Let the process be controlled with a linear, time-invariant
controller, having the transfer function G,(z - I ) , i.e.,
The variance of Y, is given by
..........................
v a r ( & ) = (1 + $7 + ... + $;)a: . . . . . . . . . . . (10) U, = -G,(z-')Y, (11)

A moving average process of orderfhas the property that Substituting this equation into the process description, Equa-
its autocorrelation function is zero beyond lagf. This property tion (l), and using the identity (6) gives
enables one to check whether any control strategy is giving
minimum variance control (Astrom, 1970; Astrom and Wit- (12)
tenmark, 1973). The sample autocorrelations can be com-
puted, and either: i) compared to their approximate 95% where
confidence intervals (Box and Jenkins, 1970, p. 177) or ii)
used to construct a portmanteau test to determine whether
the correlations beyond lagfare zero (Ljung and Box, 1978).
If the hypothesis is accepted that the observations are from
a moving average process of order j-, further reductions in
the variance of the process output cannot be achieved with
and
the existing control strategy. Further reductions can only be
achieved through process modifications, such as reducing
the inherent variability ui, reducing the deadtime, incor-
D ( z - ' ) = O(z-')6(z-') .................... (14)
porating feedforward control, eliminating the source of dis-
turbances or possibly finding another manipulated variable Alternatively, the transfer function between Y, and a, can be
to control the process variable. expressed as:

THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 67, OCTOBER, 1989 857
 proportional to the number of observations used to fit
the time series.
DeVries and Wu (1978) proposed that a time series b_efit
The polynomials F(z-') and G ( z - ' ) follow from the Equa- a,
to the Y's and be used as a performance measure. a, is
tions ( 1 3)-( 15). ?n estimate of the best achievable control only whenf = 0.
The first term on the right hand side of Equation (12) a, does however provide a lower bound on the achievable
is independent of the second term in this expression, performance of the control strategy if it were possible to
since the latter only involves values of or equiva- eliminate the process time delays.
lently a,-j, for j > f. The key feature to note from Applications
Equation (12) is that the first f moving average terms of
the closed loop transfer function are not affected by any The sampling properties of the f + 1 step ahead estimate
form of feedback. The invariance of these parameters to for the prediction error are first investigated via simulation.
feedback is simply a recognition that a feedback control This is followed by applications of this method to pilot plant
strategy, linear or nonlinear, cannot return the process and production data.
output to its target value until the process deadtime or
time delay has elapsed. The significance of this result is SAMPLING PROPERTIES OF THE ESTIMATORS
that it is not necessary to jointly estimate the parameters
of the process dynamics and disturbances. It is only In this section, the sampling properties of the f + 1 step
necessary to determine the transfer function of the closed ahead prediction error variance, given by Equation (16),
loop system, i.e., Equation (15). The order and parameters are investigated via simulation. The process descrip-
of this transfer function can be uniquely determined from tion is:
a time series fit only to the Y's. Consequently, i) it is
not necessary to perturb the manipulated variable to obtain 14.647
information about the process dynamics and ii) it is not Y,=
( 1 - 0.352-')
ur - 4
necessary to insist on any 'identifiability ' restrictions (Box
and MacGregor, 1976), as is usually the case for closed loop
identification. If a nonlinear controller is used, as is the + (1 - 1 . 3 5 1~ +
~ '0 . 3 5 ~ a,~ ~. . ). . . . . . . . . . . (18)
case in some statistical control methods, then it is necessary
to simultaneously estimate the parameters of the process
dynamics and disturbances. Y, represents the deviation from target value of the dry
To estimate the theoretically achievable minimum variance basis weight on a paper machine. U, is the deviation thick
performance, it is necessary to estimate the f + l step stock flow rate. Model identification, development of the
ahead prediction error variance of the process. This requires minimum variance controller and industrial implementation
determining the order and estimating the parameters of of this controller are discussed in Astrom (1970) and Priestley
the polynomials F ( z - ' ) and G(z-') in Equation (15). f (1981).
must also be known. Methods for accomplishing this are The simulated process was controlled with a discrete
well known (Box and Jenkins, 1970; Newton, 1988). proportional integral controller of the form:
Typically, the parameters of the polynomials are estimated
by maximizing the conditional likelihood function of v U ,= -0.13Y, + O.OllY,-I . . . . . . . . . . . . . . . . (19)
the observations. The prediction error variance is estimated
by Four hundred observations were simulated. The [ a,'s]
-2
Uf+' = a:(l + $! +... + $); . . . . . . . . . . . . . . (16)
were normally distributed, with mean zero and standard
deviation 0.257. The closed loop was modeled as an
autoregressive process of order n , i.e. F(z-')= 1 . The
In Equation (16) af
is the maximum liklihood el_stimate order n was determined using the final prediction error
of the one step ahead prediction error and the $'s are concept of Akaike (1969). The f + l step ahead predic-
the estimates of the moving average coefficients obtained tion error was calculated from Equations (16) and (17).
by dividing the denominator of the identified transfer This procedure was repeated five hundred times with a
function into the numerator. This is equivalect to solving different set of random numbers used in every simulation.
the following Diophantine equation for $ ( z - ' ) and A histogram of c ~ f + ~ / a is
f+~shown in Figure 1 for
T ( Z - I): these simulations. From the analysis of this and other
simulations, one finds that in order to obtain a good
&-I) = i(Z-')e(z-') + Z - ~ + ' ) T ( Z - .' ). . . . . (17) estimate of the prediction error variance, the number
of observations needs to be one hundred to one hundred
The circumflexes in Equation (17) emphasize that the and fifty times (f+l). Comparable results were obtained
quantities are estimated. The estimate of the prediction by modelling the closed loop as an ARIMA (n, 0,n-1)
error variance given by Equation (16) does not take into process.
account the uncertainty associated with the fact that the
parameters are not exact, but rather are estimates of the POLYMER PRODUCTION DATA
parameters. Properties of predictors for time series which
employ estimated parameters have been investigated by Series L in Box and Jenkins (1970) is a collection of
Box and Jenkins (1970), Baillie (1979) and Fuller and 312 hourly observations of a measure of melt viscosity
Hasza (1981). Their results indicate that the prediction of a polymer from its target value. This data is from an
error variance is inflated by the use of estimated para- industrial process and was collected under closed loop
meters. However, this variance inflation factor is inversely conditions with no extraneous perturbations of the input
858 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 67. OCTOBER. 1989
 0.50

0.25

0.00

-0.25

-0.50
0.80 0.90 1.00 1.10 1.20 0 10 20 0
/.
LAG
Figure 1 - Simulation of Prediction Error Standard Deviation. Figure 3 - Correlogram of Prodtiction Data
0.50 I

- 2 5 1 -
0 100 200 300

OBSERVATION

Figure 2 - Production Data

POLYMER PRODUCTION DATA

signal. This data is plotted in Figure 2. The standard devia-

tion of the process variable is 4.74. The autocorrelation func-
tion of the data, along with its approximate ninety-five
percent confidence intervals are plotted in Figure 3. For this
process, it is known that f = 0. Thus, we reject the
hypothesis that the process is operating under minimum
0 10 20 30
variance control.
This data was fit and adequately modeled by an ARIMA
LAG
(6,0,0) and and ARIMA (2,0,1) model. Figure 4 shows Figure 4 - Correlogram of ARIMA (2,0,1)Model Fit to Produc-
the theoretical autocorrelation function for the parameters tion Data
of the ARIMA (2,0,1) model used to represent this time (1 - 1.32 z - I + 0.51 z-*)Y, = ( 1 - 0.9 z-')~,
series. This matches very closely the obseryed auto-
correlation function, Figure 3. The estimate of u1 = 4.08 n-butane is discussed in Harris et al. (1980). Figure 5A
is an estimate of the lower bound on the achievable per- shows the process variable - the hot spot temperature
formance if a minimum variance controller were to be and the manipulated variable - the butane flowrate to
implemented. the reactor. A proportional integral controller was used
From Figure 2, we note that observations 162 and 265 to regulate the hot spot temperature. The sample auto-
might be outliers. These values were replaced by the average correlation function of the hot spot temperature is shown
of their two nearest neighbours. The process standard devi- in Figure 6A. For this process it is known that f = 0. It
ation was recalculated as 4.44. The orders and parameters is clear that this process is not under minimum variance
of the final time series were essentially unchanged. u1 was control. The hot spot temperature was adequately repre-
estimated as 3.86. sented by an AR(2) model with complex rogts close to
the stability boundaries. Using this model u1 was esti-
PILOT PLANT REACTOR DATA mated as 1.95. Figure 5B shows the process variable and
the manipulated variables when an adaptive controller
The application of self-tuning regulators for the control was used to regulate the hot splot temperature. The sample
of a pilot plant reactor carrying out the hydrogenolysis of autocorrelation function of the hot splot temperature,
THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 67, OCTOBER, 1989 859
 RANGE 20-110 125-210
Nomenclature
VAR [ Y ] 9.7 3.3
VAR[ V U ] 4.5 130.0 a, = white noise driving force
A B b = number of whole eriods of delay, Equation (2)
300, 1 C(z-’) = polynominal in z-‘, Equation (13)
d = degree of difference in disturbance model
0

d
in Qr = process disturbance at time t
= b step ahead minimum variance forecast
Dr+b/,
I-
W 280
D(z- ) = polynominal in z - I , Equation (14)
?! = b step ahead forecast error
= number of integer periods of delay, Equation (2)
F(z-’; = polynomial in z-’, Equation (15)
G(z - ) = polynomial in z - I , Equation (15)
Gc(z-‘) = transfer function of controller, Equation (1 I)
P = order of +(z-:)
4 = order of e ( z - I )
r = order of 6 ( z - )
S = order of w ( z - )
T = control interval
T(z-’) = polynomial in z - ’ , Equations (6) and (17)
u, = manipulated variable at time t, (deviation
5 O J : : : : : ;
100
: : : : ’ !
200 variable)
OBSERVATION y, = controlled variable at time f, (deviation from
Figure 5 - Pilot Plant Reactor Data setpoint)
A: PI Controller Z-I = backward shift operator
B: Adaptive Controller
Greek Symbols
1.007 1 1.00-

V = difference operator
6 ( z-I) = denominator polynomial in process dynamics,
Equation (1)
0, = standard deviation of a,
q + l = standard deviation off+ 1 step ahead prediction
error
?d = process deadline
+(z-’) = autogressive component in disturbance model,
I Equation (3)
0 5 10 5 10
= moving average component in disturbance
O(z-’)
LAG LAG model, Equation (3)
w(z-l) = numerator polynomial in process dynamics,
Figure 6 - Correlogram of Pilot Plant Reactor Data
A: PI Controller Equation (1)
B: Adaptive Controller
Superscripts
Figure 6B, indicates that this is essentially minimum vari-
ance control. The standard deviation of the hot spot temper- = maximum likelihood estimate
ature is 1.82 which does not differ significantly from the
estimated value of 1.95. We note in this figure that the References
manipulated variable is exhibiting bang-bang control action.
This arises because the transfer function of the process has Akaike, H., “Fitting Autoregressive Models for Prediction”, AM.
Inst. Statist. Math. 21, 407-419 (1969).
a zero outside the unit circle. To eliminate this behavior, it Astrom, K. J., “Introduction to Stochastic Control Theory”, Aca-
is necessary to constrain the control action. This is discussed demic Press (1970).
in Harris et al. (1980). Astrom, K. J. and B. Wittenmark, “On Self-Tuning Regulators”,
Automatica 9, 457-476 (1973).
Summary Astrom, K. J. and B. Wittenmark, “Computer Controlled Systems:
Theory and Practice”, Prentice-Hall (1984).
It has been shown that the best achievable performance, Baillie, R. T., “The Asymptotic Mean Squared Error of Multistep
when measured by the mean square error, can be estimated Prediction from the Regression Model with Autoregressive
from a time series fit to closed loop process output data alone. Errors”, Journal of the American Statistical Association 74,
It is only necessary to know the integer portion of the process 175-184 (1979).
time delay divided by the control interval. No ‘identifiability’ Box, G. E. P. and G. M. Jenkins, “Time Series Analysis: Fore-
casting and Control”, Holden-Day (1970).
constraints need be applied. Application of the methodology Box, G. E. P., G. M. Jenkins and J. F. MacGregor, “Some Recent
to pilot plant and production data indicates that this approach Advances in Forecasting and Control”, Part 11, Applied Statistics
gives very credible estimates of the variance that can be real- 23, 158-179 (1974).
ized under minimum variance control. Box, G. E. P. and J. F. MacGregor, “Parameter Estimation with
Closed Loop Operating Data”, Technometrics 18, 37 1-380
Acknowledgements ( 1976).
DeVries, W. R. and S. M. Wu, “Evaluation of Process Control
The author gratcliilly :ic.knowlcdgcs support from the Natural Effectiveness in Paper Basis Weight via Multivariate Time-Series
Sciences and Engineering Kcsciircli Council o f Canada. Analysis”, IEEE AC 23, 702-709 (1978).

860 ‘I‘Hti CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 67, OCTOBER, 1989
Fuller, W . A. and D. P. Hasza, “Properties of Predictors for Newton, H. J . , “TIMESLAB: A Time Series Analysis Labora-
Autoregressive Time Series”, Journal of the America1 Statistical tory”, Wadsworth & BrooksKole (1988).
Association 75, 155-161 (1981). Ohtsu, K . and G. Kitagawa, “Statistical Analysis of the AR Type
Harris, T. J., J. F. MacGregor and J. D. Wright, “Self-Tuning Ship’s Autopilot System”, J. Dynamic Systems, Measurement
and Adaptive Controllers: An Application to Catalytic Reactor and Control, 106, 193-202 (1984).
Control”, Technometrics 22, 153-164 (1980). Palmor, Z. J. and R. Shinnar, “Design of Sampled Data Con-
Harris, T. J., J. F. MacGregor and J. D. Wright, “An Overview trollers”, Ind. Eng. Chem. Proc. Des. Dev. 18, 8-30 (1979).
of Discrete Stochastic Controllers: Generalized PID Algorithms Priestley, M. G., “Spectral Analysis and Time Series”, Academic
with Dead-Time Compensation”, Can. J. Chem. Eng. 60, Press (1981).
425-432 (1981). Pryor, C., “Autocovariance: Power Spectrum Analysis - Derive
Ljung, G. M. and G. E. P. Box, “On a Measure of Lack of Fit New Information from Process Data”, Control Eng. 2, 103-106
in Time Series Models”, Biometrika 65, 297-303 (1978). (October 1982).
MacGregor, J. F., T. J. Harris and J. D. Wright, “Duality Between
the Control of Processes Subject to Randomly Occurring Deter-
ministic Disturbances and ARIMA Stochastic Disturbances”, Manuscript received August 5, 1988; revised manuscript received
Technometrics 26, 389-397 (1984). February 17, 1989; accepted for publication March 20, 1989.

THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING, VOLUME 67, OCTOBER, 1989 86 1