Professional Documents
Culture Documents
Abstract pH control is recognized as an industrially impor- encountered in industrial processes such as distillation, pH
tant, yet notoriously difficult control problem. Wiener models, neutralization and heat exchange. Alduwaish et al. [1] showed
consisting of a linear dynamic element followed in series by that multilayer feedforward neural networks can be used in
a static nonlinear element, are considered to be ideal for rep-
resenting this and several other nonlinear processes. Wiener the identification and control of Wiener models. In the case
models require little more effort in development than a standard of pH neutralization, as is considered here, the suitability
linear step-response model, yet offer superior characterization of the Wiener model for representing the process has been
of systems with highly nonlinear gains. These models may be recognized by several researchers. Pajunen [15] investigated
incorporated into model predictive control (MPC) schemes in several different techniques for fitting a Wiener model to the
a unique way which effectively removes the nonlinearity from
the control problem, preserving many of the favorable properties pH process. Kalafatis et al. [12] noted that a pH process
of linear MPC. In this paper, Wiener model predictive control can be accurately modeled by a Wiener model if the process
(WMPC) is evaluated experimentally, and also compared with stream is significantly larger than the titrating flowrate yet
benchmark proportional integral derivative (PID) and linear also recognized that good results may be obtained even when
MPC strategies, considering the effects of output constraints and the titrating flowrate is a significant fraction of the process
modeling error.
flowrate.
For the control of pH, many researchers suggested a variety
I. INTRODUCTION of control strategies and algorithms. Gupta and Coughanowr
[10] developed a controller based on a first-order linear model,
M ODEL predictive control (MPC) allows us to use the
detailed knowledge of a process, in the form of a
dynamic model, as an aid to controlling that process within
but which adapted the parameters such that the a constant
closed-loop gain was maintained, while Shinskey [17] com-
the required constraints [9]. Typically, linear models are used pensated for the nonlinear process gain using a piecewise
for this, despite the fact that essentially all industrial processes linear gain. Wright and Kravaris [18] developed an adaptive
exhibit some degree of nonlinear behavior. This is due to the controller which dealt with the nonlinearity of the titration
significant increase in complexity of the predictive control curve by means of strong acid equivalents. More recently,
problem resulting from the use of a nonlinear model [2]. Proll and Karim [16] demonstrated the use of nonlinear
Linear MPC employs models which are linearized about ARX models for real-time model-predictive control of pH
the operating point as an aid to predicting the response of the neutralization. Fruzzetti et al. [7] presented an MPC of pH
controlled process. This strategy proves to be quite successful neutralization using Hammerstein models, thus motivating this
even in controlling mildly nonlinear processes. The higher study.
the degree of nonlinearity, however, the greater the level of The work presented here includes the development of a
mismatch between actual process and the representative model, Wiener model, consisting of a cubic spline to represent the
hence resulting in a deterioration of controller performance. nonlinearity and a step-response model describing the dynam-
For extending linear MPC to the control of nonlinear ics, with the subsequent incorporation of this model into a
processes, a model is required that can represent the salient MPC scheme. Experimental results dealing with both uncon-
nonlinearities but possibly without the complications associ- strained and constrained control as well as the performance
ated with general nonlinear models. It is in fulfilling this need in the face of significant model mismatch are presented. This
that Wiener models are particularly useful. Wiener models paper significantly extends the work previously presented on
consist of a linear dynamic element followed in series by a the unconstrained control of pH neutralization [13] using
static nonlinear element [3], which translates to a process with WMPC by considering constrained control, the issue of model
linear dynamics but a nonlinear gain. This is a structure that mismatch and the advantages of filtering the feedback signal.
can adequately represent many of the nonlinearities commonly
II. WIENER MODEL IDENTIFICATION AND VALIDATION
Manuscript received April 3, 1997. Recommended by Associate Editor, T. In its most general form, a Wiener model consists of a
Ogunnaike.
S. J. Norquay is with Orica Ltd., Matraville, NSW 2036, Australia. linear dynamic element followed in series by a static nonlinear
A. Palazoglu is with the Department of Chemical Engineering and Materials element as shown in Fig. 1. Provided it is memoryless, the
Science, University of California, Davis, CA 95616 USA. nonlinear element can be chosen from a virtually unlimited
J. A. Romagnoli is with the Department of Chemical Engineering, The
University of Sydney, Sydney NSW 2006, Australia. number of structures, ranging from simple algebraic functions
Publisher Item Identifier S 1063-6536(99)05855-8. to complex neural networks. An important factor governing
10636536/99$10.00 1999 IEEE
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
438 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 7, NO. 4, JULY 1999
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
NORQUAY et al.: APPLICATION OF WIENER MODEL PREDICTIVE CONTROL 439
time intervals, called the prediction horizon. A performance Using the equations above, we may now represent the vector
function consisting of only a least-squares error, however, can of (1) by
result in excessively large changes in the input variable so this
is rectified by the addition of a term which penalizes large (7)
changes in input. A penalty on the absolute value can also be
where
added which results in a performance function of
(1)
where .. ..
. .
where
(3)
(13)
where are the impulse-response coefficients and are related (14)
to the step-response coefficients as and
It is also assumed that and for with
When the step-response and impulse-response models are
used to predict the output, a correction is also added to ..
.
account for unmeasured disturbances and model mismatch. To .. .. ..
. . .
determine this term, we calculate the present model output
value from past inputs then compare it with the actual process
output, thus defining For the case of constrained control, the optimization prob-
(4) lem is expressed as a standard quadratic programming prob-
lem, as in the unconstrained case
and assume that this value remains constant over the entire pre-
diction horizon. Now, the first-order filter output is represented (15)
by
but now subject to
(5)
(16)
Here, the corrected predicted value of the intermediate variable
is given by where
(6) (17)
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
440 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 7, NO. 4, JULY 1999
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
NORQUAY et al.: APPLICATION OF WIENER MODEL PREDICTIVE CONTROL 441
following equations:
HCO H
(25)
H CO
CO H
(25a)
HCO
Bu H
(26)
HBu
Bu H
(27)
HBu
H OH (28)
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
442 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 7, NO. 4, JULY 1999
! !
Fig. 7. Plots (a), (c), and (e) show the system response to a step change in
! !
Fig. 9. (a), (c), and (e) show the unconstrained system response to a
the setpoint from pH7.0 pH8.5 pH7.0 for Wiener MPC, linear MPC and step change in the acid stream flowrate from 1.0 0.5 1.01/min for
PID, respectively. Plots (b), (d), and (f) show the corresponding manipulated Wiener MPC, linear MPC and PID, respectively. (b), (d), and (f) show the
variable moves in l/min. Time is in seconds. corresponding manipulated variable moves in l/min.
! !
Fig. 8. (a), (c), and (e) show the unconstrained system response to a step
!
change in the setpoint from pH7.0 pH5.5 pH7.0 for Wiener MPC, Fig. 10. (a) and (c) show the system response to a step change in the setpoint
linear MPC and PID, respectively. (b), (d), and (f) show the corresponding from pH7.0 pH8.5 for Wiener MPC and linear MPC, respectively, when
manipulated variable moves in l/min. a hard output constraint of pH8.6 is imposed on the first predicted output. (b)
and (d) show the corresponding manipulated variable moves in l/min.
Unconstrained Control
Figs. 79, show the results produced for setpoint changes output and placed on the second sampled predicted output
and unmeasured disturbances in the unconstrained case. It instead, and the results are shown in Fig. 11. These results
can be seen from these results that the WMPC algorithm demonstrate that placement of the constraint on the second
offers marked improvement over the linear MPC and PID by sampled predicted output is clearly a more effective method
reducing the offset and the oscillatory behavior. of constraint handling for this system as the WMPC does
constrain the system within the pH 8.6 bound and the
linear MPC allows only a small violation. This violation of the
Constrained Control
constraint by the linear MPC is probably a direct consequence
To investigate the effect of hard constraints on the system, of the linear model not being as accurate in this region as the
the setpoint was once again changed from pH 7 to pH Wiener model.
8.5, but this time with a hard constraint of pH 8.6 on the first The effects of constraining the control in an attempt to
sampled predicted output. The results of this implementation decrease the effects of a disturbance were also considered and
are shown in Fig. 10, where it can be seen that both the no significant change was observed in the performance of the
Wiener and linear MPC failed to constrain the process within WMPC algorithm.
the pH 8.6 bound. One can argue that this is due to the
controller allowing a manipulation of the base flowrate which
did not violate the constraint after one time interval, but set the Unconstrained Control with Model Mismatch
process on a trajectory which made violation of this constraint A difficulty often experienced when attempting to control
unavoidable by the second time interval. In response to this, pH is variation in the composition of the process stream
the constraint was removed from the first sampled predicted being neutralized. Different levels of buffering about different
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
NORQUAY et al.: APPLICATION OF WIENER MODEL PREDICTIVE CONTROL 443
! !
Fig. 13. (a), (c), and (e) show the unconstrained system response to a step
!
Fig. 11. (a) and (c) show the system response to a step change in the setpoint
from pH7.0 pH8.5 for Wiener MPC and linear MPC, respectively, when change in the setpoint from pH7.0 pH8.5 pH7.0 for Wiener MPC, linear
a hard output constraint of pH8.6 is imposed on the second predicted output. MPC and PID, respectively, when there is no pH7 buffer added to the acid
(b) and (d) show the corresponding manipulated variable moves in l/min. stream. (b), (d), and (f) show the corresponding manipulated variable moves
in l/min.
! !
Fig. 12. Titration curve for the pH neutralization system without pH7 buffer Fig. 14. (a), (c), and (e) show the unconstrained system response to a step
in the acid stream (solid line) and with pH7 buffer in the acid stream (dotted change in the acid stream flowrate from 1.0 0.5 1.0 l/min for Wienr
line). MPC, linear MPC and PID, respectively, when there is no pH7 buffer added to
the acid stream. (b), (d), and (f) show the corresponding manipulated variable
moves in l/min.
pH levels can play havoc with even the most accurately
designed predictive control system. It was therefore considered results for the unconstrained case are shown in Figs. 13 and 14.
important to include the effects of such a mismatch. For These results show that while there is definitely a deterioration
truly severe changes in composition where the titration curve in the performance of each of the controllers, all three remain
is drastically changed, no nonadaptive, empirically based stable and are able to maintain the required setpoint. Despite
controller will perform well, but that is not the situation the increase in oscillation within the closed loop, the WMPC
considered here. Here, we consider mismatch but not the type still performs better compared to both the linear MPC and the
of radical buffering mismatch which results in a system bearing PID control.
little resemblance to the original. In this case, the mismatch is
a reduction of buffering about the operating point. This is often
a difficult disturbance for a controller to deal with as it may Constrained Control with Model Mismatch
cause significant increase in the process gain, a modeling error While the performance of the controllers was shown to be
which can result in instability or at least impractical levels good for the unconstrained case, we wished to examine the
of oscillation in many a controller. To investigate this, no effect that this modeling error had on the constraint handling
potassium phosphate monobasic/sodium hydroxide buffering of the model predictive controllers. The same setpoint change
was used in the acid solution, resulting in the titration curve was performed as for the fully buffered case, with a constraint
shown in Fig. 12. The increase in gain about the operating on the second sampled predicted output. The results are shown
point is clearly visible from this graph. in Fig. 15, where it is observed that the mismatch does have a
First, the same setpoint and acid flowrate changes were significant effect on the ability of the controllers to work within
performed previously for the fully buffered system and the the constraints. The system violates the constraint when under
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
444 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 7, NO. 4, JULY 1999
! !
Fig. 16. (a), (c), and (e) show the unconstrained system response to a step
!
Fig. 15. (a) and (c) show the system response to a step change in the setpoint
change in the setpoint from pH7.0 pH8.5 pH7.0 for Wiener MPC with
from pH7.0 pH8.5 for Wiener MPC and linear MPC, respectively, when
there is no pH7 buffer added to the acid stream and a hard output constraint
= 0:5; = 0:6; and = 0:8; respectively, when there is no pH7 buffer
added to the acid stream. (b), (d), and (f) show the corresponding manipulated
of pH8.6 is imposed on the second predicted output. (b) and (d) show the
variable moves in l/min.
corresponding manipulated variable moves in l/min.
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.
NORQUAY et al.: APPLICATION OF WIENER MODEL PREDICTIVE CONTROL 445
[4] S. A. Billings and W. S. F. Voon, Correlation based model validity Sandra J. Norquay received the Bachelor of Science degree in applied
tests for nonlinear models, Int. J. Control, vol. 44, pp. 235244, 1986. mathematics and bachelor of engineering in chemical engineering at Monash
[5] H. S. Chiang and L. D. Durbin, Gain-adaptive control applied to a University, Australia, and the Ph.D. degree from the University of Sydney,
heat exchange process using a first-order plus deadtime compensator, Australia. During her postgraduate study she spent five months as a Research
in Proc. Amer. Contr. Conf., San Francisco, CA, 1983, pp. 7277. Student at the University of California, Davis. Her thesis is titled Nonlinear
[6] R. Fletcher, Practical Methods of Optimization: Vol. 2, Constrained Model Predictive Control Based on Wiener and Hammerstein Structures.
Optimization. New York: Wiley, 1981. After completing her degree in mid-1997 she began work as a Process
[7] K. P. Fruzzetti, A. Palazoglu, and K. A. McDonald, Nonlinear model Systems Engineer with Orica Australia.
predictive control using Hammerstein models, J. Process Control, vol.
7, pp. 3141, 1997.
[8] C. E. Garcia and M. Morari, Internal model control. 1. A unifying
review and some new results, Ind. Eng. Chem. Process Des. Dev., vol.
21, pp. 308323, 1982.
[9] C. E. Garcia, D. M. Prett, and M. Morari, Model predictive control: Ahmet Palazoglu received the B.S. degree in chemical engineering from
Theory and practiceA survey, Automatica, vol. 25, pp. 335348, Middle East Technical University, Turkey, in 1978, and the M.S. degree
1989. from Bogazici University, Turkey, also in chemical engineering, in 1980. He
[10] S. R. Gupta and D. R. Coughanowr, On-line gain identification of flow received the Ph.D. degree in chemical engineering from Rensselaer Polytechic
processes with application to adaptive pH control, AIChE J., vol. 24, Institute, Troy, NY, in 1984.
pp. 654664, 1978. He joined the faculty at the University of California, Davis, in 1994. He is
[11] R. C. Hall and D. E. Seborg, Modeling and self-tuning control
a Professor of Chemical Engineering and Materials Science at the University
of a multivariable pH neutralization process. Part I: Modeling and
of California, Davis. His research interest include nonlinear process control,
multiloop control, in Proc. Amer. Contr. Conf., Pittsburgh, PA, 1989,
distributed parameter systems, and robust control theory.
pp. 18221824.
[12] A. Kalafatis, N. Arifin, L. Wang, and W. R. Cluett, W.R., A new
approach to the identification of pH processes based on the Wiener
model, Chem. Eng. Science, vol. 50, pp. 36933701, 1995.
[13] S. J. Norquay, A. Palazoglu, and J. A Romagnoli, Model predictive
control using Wiener models, Chemical Engineering Science, vol. 53,
pp. 7584, 1998. Jose Alberto Romagnoli was born in Bahia Blanca, Argentina, in 1948. He
[14] S. J. Norquay, A. Palazoglu, and J. A. Romagnoli, Nonlinear model received the Bachelor degree and the Ph.D. degree in chemical engineering
predictive control based on Wiener and Hammerstein models, Ph.D. from the Universidad Nacional del Sur, Argentina, in 1973 and the University
dissertation, University of Sydney, Australia, 1997. of Minnesota, Mineapolis, in 1980, respectively.
[15] G. A. Pajunen, Identification of a pH process represented by a nonlinear Since 1980, he has been a Researcher for the CONICET (National Council
Wiener model, Adaptive. Syst. Control Signal Processing, pp. 9195, for Research in Science and Technology) and a Full Professor of Process
1983. Control in the Department of Chemical Engineering, Universidad del Sur.
[16] T. Proll and M. N. Karim, Model-predictive pH control using real-time During the academic year 1987 to 1988, he was a Visiting Associate Professor
NARX approach, AIChE J., vol. 40, pp. 269282, 1994. at University of Minnesota while on sabatical leave from the Universidad del
[17] F. G. Shinskey, Adaptive pH controller monitors nonlinear process, Sur. He was also a Visiting Associate Professor at University of California
Control Eng., Feb. Issue, pp. 5759, 1974. at Davis during the academic year 1988 to 1989. He holds now the Joint
[18] R. A. Wright and C. Kravaris, C., On-line identification and nonlinear Orica-University of Sydney Chair of Process Systems Engineering at Sydney
control of pH processes, in Proc. Amer. Contr. Conf., San Francisco, University. His current research interests include process dynamics and robust
CA, 1993, pp. 11671171. linear and nonlinear control.
Authorized licensed use limited to: Universidad Autonoma de Nuevo Leon. Downloaded on January 13, 2009 at 19:49 from IEEE Xplore. Restrictions apply.