53 - Pid
53 - Pid
103–115
www.elsevier.comrlocaterpowtec
a
Department of Electrical Engineering, LaÕal UniÕersity, Quebec, Canada G1K 7P4
b
Department of Mining and Metallurgy, LaÕal UniÕersity, Quebec, Canada G1K 7P4
Accepted 20 September 1999
Abstract
A conventional grinding circuit consisting of one open-loop rod mill and one closed-loop ball mill is essentially a two-input= two-
output system, assuming that the classifier pump box level is controlled by a local loop. The inputs are the ore and water feed rates and
the outputs are the product fineness and the circulating load. The design problem is to find a control algorithm and a tuning procedure
which satisfy specified servo and regulatory robust performances. A first approach is to use decentralized PID controllers and systematic
tuning methods which take into account loop interactions. Another technique consists of adding decouplers or pseudo-decouplers to the
decentralized controllers. Finally, the design of a fully multivariable controller is a possible option. To face the problem of performance
robustness related to change of process dynamics, two options are studied. A design criterion involving the minimization of a penalized
quadratic function on a future trajectory can be used. A second alternative is to track process dynamics changes using adaptive process
modelling. The paper will present a comparison of these various strategies, for a simulated grinding circuit. A benchmark test, involving a
sequence of disturbances Žgrindability, feed size distribution, change of cyclone number . . . . and setpoint changes, is used to compare the
performances of the controllers. q 2000 Elsevier Science S.A. All rights reserved.
Keywords: Grinding circuit; Distributed PID; Internal model control; Multivariable predictive control; Distributed adaptive predictive control
0032-5910r00r$ - see front matter q 2000 Elsevier Science S.A. All rights reserved.
PII: S 0 0 3 2 - 5 9 1 0 Ž 9 9 . 0 0 2 0 7 - 7
104 A. Pomerleau et al.r Powder Technology 108 (2000) 103–115
The first studied algorithm is the standard decentralized process and the time evolution of the parameters. It also
PID controller which is usually the reference to compare solves the tuning problems of the distributed control algo-
different algorithms. PID controllers present some draw- rithm.
backs. First, their performances are limited by the proper- The time specification on the circulating load has been
ties of the fixed structure. Second, the inputroutput pair- arbitrarily fixed to 25 min. This corresponds approxima-
ing problem has to be solved for multivariable systems. tively to a dynamic similar to the open-loop function
Finally, an efficient tuning method has to be used to obtain between the CL and RMF. The specification on the size
the desired dynamics of the controlled loops. distribution has been fixed to 5 min. Here, since the
The second studied algorithm consists of distributed open-loop transfer functions relating the size distribution
internal model controllers applied to the process preceded to the RMF or to the PBW have zeroes, it is more difficult
by a partial decoupler. The decoupler partially solves the to relate the specification to the open-loop time constants.
pairing problem and facilitates the tuning. The internal
model control structure which is model-based eliminates
the problems of a fixed structure. 5. Pairing of controlled and manipulated variables
The third studied algorithm is a full multivariable opti-
mal predictive controller. It totally solves the pairing prob- Distributed control consists of using n SISO to control
lem and takes into account the compromise between the an n = n process as shown in Fig. 4. The main difficulties
performances and the manipulated variables energy through of distributed control lie in the choice of the proper
a quadratic criterion. input–output pairing and in the fact that the tuning of each
Finally, a distributed adaptive predictive controller is SISO controller depends on the tuning parameters of the
used. Compared to the other algorithms, the adaptive part second SISO controller. When manipulated variables are
permits to take into account the nonlinearities of the not properly selected, interactions between controlled and
manipulated variables can result in undesirable control direct transfer functions. The presence of zeros in the
loop interactions, leading to poor control performances. transfer functions also degrades the performance of RGA.
The well-known Bristol’s relative gain array ŽRGA. w6x Consequently, a dynamic method must be employed.
being based only on steady-state information does not The basic and simple idea of the RDG ŽRelative Dy-
consider the process dynamics and is inappropriate for namic Gain. method is to extend the calculation of the
many processes. This is especially the case when dynamics Bristol’s relative gain for all frequencies. The method used
in transversal transfer functions are fast compared to direct here is the generalized dynamic relative gain w7x. For the
transfer functions and present smaller static gain than process under study, the RDG frequency responses are
given in Fig. 7. For the dynamics required by the specifica- The system can be drawn as depicted in Fig. 8. For each
tion, the output Y1 , is paired with U1. closed loop, the control objectives are: Ž1. no permanent
error and Ž2. a second-order set point tracking response
with specified dynamics. Thus, the corresponding closed-
loop transfer functions can be selected as follows:
6. Algorithm 1: Decentralized extended PIDs
Y1 Ž s . 1 y t 01 s
s
To tune the controllers, a frequency-based method, in Z1 Ž s . Ž 1 q t 11 s . Ž 1 q t 21 s .
which the open-loop characteristics are deduced from the
Y2 Ž s . 1 y t 02 s
closed-loop dynamics is used w8x. First, for tuning a con- s
troller, it is imperative to know the dynamics of the Z2 Ž s . Ž 1 q t 12 s . Ž 1 q t 22 s .
process to be controlled. With monovariable processes, the
method consists of opening the loop and evaluating the The open-loop characteristics deduced from the tracking
transfer function seen by the controller. For multivariable closed-loop specifications are:
processes, the loop with the controller under consideration
is opened while all the other loops are kept closed. For a G 0 L 1 Ž s . s GC1 Ž s . G 1 Ž s .
2 = 2 process, the transfer function seen by the controller 1 y t 01 s
of the first loop, GC1Ž s ., is: s
s Ž t 11 q t 21 q t 01 q t 11t 21 s .
G 12 Ž s . G 21 Ž s . GC 2 Ž s .
G 1 Ž s . s G 11 Ž s . y . G 0 L 2 Ž s . s GC 2 Ž s . G 2 Ž s .
1 q GC 2 Ž s . G 22 Ž s .
1 y t 02 s
The transfer function seen by the second controller s
s Ž t 12 q t 22 q t 02 q t 12t 22 s .
GC2 Ž s . is:
The transfer functions GC1Ž s . and GC2 Ž s . of the con-
G 12 Ž s . G 21 Ž s . GC1 Ž s . trollers are obtained by solving the previous equations.
G 2 Ž s . s G 22 Ž s . y .
1 q GC1 Ž s . G 11 Ž s . Due to their quadratic form, it leads to two Bode plots for
both controllers. Knowing the required sign of the con-
As can be seen, there exists an interdependence be- troller gain and the necessary presence of an integrator, the
tween GC1Ž s . and GC2 Ž s .. Indeed, their individual tuning appropriate solution can be selected without any difficulty.
depends on the choice of the other controller. Conse- The last step in tuning the controllers consists of defin-
quently, to evaluate the unknowns GC1Ž s . and GC2 Ž s ., ing the transfer functions GCP1Ž s . and GCP2 Ž s . that best
further information is necessary. This supplementary infor- approximate the frequency response GC1Ž j v . and GC2 Ž j v .,
mation is given by the closed-loop specifications which while ensuring a stable closed-loop system. The structure
can be translated into open-loop char-acteristics. of each controller is imposed and an evaluation of the
parameters is performed using a constrained objective Under the assumption that the system is satisfactorily
function. The following structure is used: decoupled for neglecting the remaining coupling for tuning
consideration, the transfer functions for the system under
K C Ž 1 q Ti s . Ž 1 q Td1 s q Td 2 s 2 .
GCP i Ž s . s . study are given by
Ti s Ž 1 q Tf 1 s q Tf 2 s 2 .
0.6925 ey5 s
G1Ž s . s
The resulting controllers are: Ž 1 q 18 s .
2
1.28 Ž 1 q 85.4 s . Ž 1 q 19.15s q 73.4 s . Ž 1 y 2.5s . Ž 1 q 2 s . Ž 1 q 21s . ey7 s
GC1 Ž s . s q 1.974
85.4 s Ž 1 q 17.81 s q 90.6 s . Ž 1 q s . Ž 1 q 28 s . Ž 1 q 92 s .
0.32 Ž 1 q 2.4 s . Ž 1 q 4.77s q 9.06 s 2 . Ž 1 y 3.5s .
GC 2 Ž s . s =
2.4 s Ž 1 q 6.4 s q 7.8 s 2 . Ž 1 q 4.5s . Ž 1 q s .
0.393 Ž 1 q 92 s . eys
G2 Ž s . s
Ž 1 q 2 s . Ž 1 q 21s .
7. Control algorithm 2: algebraic multivariable control Ž 1 y 3.5s . Ž 1 q 18 s .
q 1.120
Multivariable algebraic control can be done either with Ž 1 q 4.5s . Ž 1 q s .
a full multivariable controller or by using a decoupler in Ž 1 y 2.5s . ey8 s
cascade with monovariable controllers. The latter one, as =
Ž 1 q s . Ž 1 q 28 s .
shown in Fig. 9, is used here. The methodology consists
of: These transfer functions can be approximated by:
– evaluating the decoupler, 2.666ey5 s
– tuning the monovariable controllers on the model G1Ž s . s
Ž 1 q 85s .
including the multivariable process model and the de-
1.513 Ž 1 q 5s . eys
coupler. G2 Ž s . s 2
For many processes Že.g., grinding circuit., it is impos- Ž 1 q 3s .
sible to realize a perfect decoupler. This happens with The distributed internal model controllers are then tuned
non-minimal phase processes or with systems with longer into these transfer functions.
delays in the direct transfer functions. For the system
under study, the decouplers are:
G 12 Ž s . 8. Control algorithm 3: multivariable state-space inter-
D 12 Ž s . s y nal model predictive control
G 22 Ž s .
Here the controller is designed to control the process
1.633 Ž 1 y 2.5s . Ž 1 q 2 s . Ž 1 q 21 s .
s y7 s
e model and not the process itself. The difference between
0.393 Ž 1 q 28 s . Ž 1 q 92 s . the model and the process outputs, respectively, y and y P
G 21 Ž s . submitted to the same control action u, is then interpreted
D 21 Ž s . s y as an overall estimation d of the disturbance d. This
G 11 Ž s . estimated value embeds in fact both disturbances and
0.475 Ž 1 y 3.5s . modelling errors. The value d is used to modify the
s Ž 1 q 18 s . setpoint Žreference value r .. This biased setpoint is the
0.69 Ž 1 q 4.5s . Ž 1 q s .
desired value yd that must be tracked by the process
D 12 Ž s . realizes a perfect decoupler, while D 21Ž s . is a model. The controller calculates u in such a way to cancel
pseudo-decoupler. the difference e between yd and the model output ym .
Fig. 10 summarizes the basic principles of this control also be viewed as filters which modify the dynamics of
architecture. It is generally named Internal Model Control setpoint and disturbance changes.
ŽIMC.. Although the control actions weighting factors can be
At time i, the sequence of control actions to be applied used to constrain the variations of these control actions,
in the future Ž u i , u iq1 , . . . u iqHy1 ., where u k is the vector strict inequality constraints can be added to the criterion.
of control actions at time k and H the control horizon, is They can be formulated as:
selected such that the following quadratic performance u min F u F u max < D u < F D max
criterion is minimized: where D u is the incremental variation of u. Also, inequal-
iqH ity constraints can be formulated on the process model
Ji Ž u i . . . u iqHy1 . s Ý e kT Q e k outputs y or even on the process model states.
ksiq1
The resulting algorithm is a controller which uses the
iqH
state vector x of the process model and the reference
q Ý Ž u ky 1 y u ky2 . T trajectory yd .
ksiq1
For the case under study, we have:
=R Ž u ky 1 y u ky2 .
0.1 0 1 0
Rs Qs
In the equation, e k is the vector of differences between 0 0.1 0 1
the desired output yd and the model output y at time k in 1
the future. The matrix Q is used to scale the variables and 0
1 q 25s
to define the relative importance of the errors for the Md s
1
different output values. The matrix R has the same func- 0
tion, i.e., to weigh the incremental variations of the control 1 q 5s
actions. 1
It is important to note that the control variable penalty 0
1 q 25s
term in the criterion Ji is expressed with respect to the Mr s
1
increment of u and not to its variation around its nominal 0
value. This formulation warrants the cancellation of the 1 q 5s
steady-state deviation of the process output to the setpoint.
At each time i, the criterion Ji is minimized with
9. Adaptive predictive control algorithm
respect to u i , . . . u iqHy1 , then at the next time i q 1, again
the criterion Jiq1 is minimized. This is why this strategy Industrial processes are generally characterized by non-
is also called a moving horizon control. linear time-varying dynamics w9x. Most of nonlinear sys-
The control scheme of Fig. 10 can be improved by tems, however, may be modelled by locally linear approxi-
adding reference models which allow an easy tuning of the mations. The input–output signals are assumed to be large
closed loop dynamics. Reference models can be added to enough to avoid small signal nonlinearities Žvalve stiction,
specify the setpoint tracking dynamics and the disturbance etc.., but not too large to reach large signal nonlinearities
rejection dynamics. The reference models Md and Mr can Žhysteresis, saturation, etc...
The plant input–output behaviour which is to be adap- where A, B and C are polynomials in the backward shift
tively modelled is described by the following disturbed operator of order na, nb and nc.
linear discrete model: The parameter adaptation mechanism is the heart of the
adaptive control system: the performance of the control
A Ž qy1 . y Ž t . s qyk B Ž qy1 . u Ž t . q Õ Ž t . algorithm depends directly on the quality of this mecha-
C Ž qy1 . nism. The presented method deals with parameter estima-
ÕŽ t. s j Ž t. tion based on the information collected on the process, i.e.,
D Ž qy1 . the input–output measurements.
112 A. Pomerleau et al.r Powder Technology 108 (2000) 103–115
T
For identification purposes, the process output is ex- u s w a1 ,a2 , . . . ,a n a ,b 0 ,b 1 , . . . ,bn b x
pressed in terms of a parameter vector u and a regressor or
data vector: f Ž t . s yy Ž t y 1 . , . . . ,y y Ž t y na . ,
y Ž t . s u Tf Ž t . q Õ Ž t . u Ž t y k . , . . . , u Ž t y k y nb .
T
A. Pomerleau et al.r Powder Technology 108 (2000) 103–115 113
It is always advisable to provide the parameter estima- noise and unmodelled dynamics. This ensures smooth
tor with well-conditioned signals. The signal conditioning behaviour of parameter estimation and reduces their varia-
is provided by scaling, filtering and data normalization. tions.
Low-pass filtering of the measurements may be used in The long-range predictive control objective consists of
order to reduce the high frequency modes due to both minimizing the sum of squares of the errors between
predicted and desired output trajectories with an additional subject to D uŽ t q j y 1. s 0 when j ) Nu, where E is the
term weighting projected control increment: mathematical conditional expectation given information up
Ny to time t, w Ž t . is the desired output, Ny the maximum
JsE
½ Ý Ž yŽ tqj. ywŽ tqj. .
jsNm
2
prediction horizon, Nm the minimum prediction horizon,
Nu is the control horizon and l is a weighting upon future
Nu control increments. This criterion is roughly the same as
ql Ý
jsNm
2
D u Ž t q j y 1 . rt
5 the previous one except that it is monovariable and weight-
ing factors are limited only to one l.
A. Pomerleau et al.r Powder Technology 108 (2000) 103–115 115
A benchmark has been used to test the various control algorithms. It consists of the following:
Time Žmin. Event
50 CL setpoint from 191.7 to 211.7 trh
225 SD setpoint from 76.9% to 75.65%
300 CL setpoint from 211.7 to 231.7 trh
from 75.65% to 74.4%
575 Ore feed from 100% hard coarse to 50% hard coarser50% soft fine
825 Number of hydrocyclones from 4 to 3
1075 Diameter of apex from 4.64 to 4.44 cm
1300 CL setpoint from 231.7 to 171.7 trh
1500 SD setpoint from 74.4% to 75.65%
Based on the observations of the results ŽFigs. 11–14., – Algebraic tuning methods may require more know-
the following conclusions can be drawn: how than optimal controllers, but are easier to implement
– For setpoint changes, the algorithms that have a decou- on industrial DCS.
pler eliminate totally or partially the disturbance on the – Predictive controller and algebraic controllers per-
variable that was not imposed a setpoint change as was form equally in a stochastic environment if there is no
expected. noise model available.
– The algebraic internal model and the multivariable – Adaptive controllers perform better than fixed con-
predictive controller perform equally good. The advan- trollers for parametric disturbances or soft nonlinearities. It
tage of the former is to be easily implementable on must be noted though that identification is very difficult in
DCS Ždistributed control system. that has bloc pro- regulation where external disturbances act on the process.
grammation available. It may require though better They also have the advantage of facilitating the tuning
know-how from the designer. principally for distributed controllers.
– The adaptive controller performs better than the fixed
controllers in face of parameter variations occurring
closed to the inputs of the process Žore feed change.,
but no better for parameter variations near the output References
Žnumber of hydrocyclones..
w1x D. Hodouin, Y. Marcotte, A. Pomerleau, F. Flament, Predictive
Control of Grinding Circuit: An Evaluation by Dynamic Simulation.
w2x A. Desbiens, A. Pomerleau, K. Najim, Int. J. Miner. Process. 41
11. Conclusion Ž1994. 17–31.
w3x A. Desbiens, K. Najim, A. Pomerleau, D. Hodouin, Distributed
From the observations of the results on the grinding Partial State Reference Model Adaptive Control — Practical Aspects
process, which can be extrapolated to any process from a and Application to a Grinding Circuit, Optimal Control Applications
control point of view, one can conclude the following. and Methods 18 Ž1997. 29–47.
w4x F. Flament, R. del Villar, R. Lanthier, Computer aided design of a
– Fixed structure controllers Že.g., PID. perform as well control strategy for an industrial grinding circuit, in: R. Poulin, R.C.T.
as model-based controllers for processes where the delay is Pakalnis, A.L. Mular ŽEds.., Proc. of the 2nd Can. Conf. on Com-
relatively small compared to the dominant time constant puter Applications in the Mineral Industry, Vol. 1, 1991, pp. 337–348.
Ž u - 5T . since the models are rarely of order higher than w5x D. Hodouin, Y. Dube, ´ R. Lanthier, Stochastic simulation of filtering
two. and control strategies for grinding circuit, Int. J. Miner. Process. 22
Ž1988. 261–274.
– Distributed controllers perform as well as multivari- w6x E.H. Bristol, On a new measure of interaction for multivariable
able controllers in regulation if the coupling of the process process control, IEEE Trans. Autom. Control AC 11 Ž1966. 133.
is taken into account in the design and the proper pairing is w7x H.P. Huang, M. Ohshima, I. Hashimoto, J. Proc. Cont. 4 Ž1994.
used for the dynamics required. 15–17.
w8x A. Desbiens, A. Pomerleau, D. Hodouin, Frequency based tuning of
– Algebraic multivariable controllers perform as well as
SISO Controllers for two by two Processes, IEE Proc. on Control
optimal multivariable controllers and they have exactly the Theory and Applications 43 Ž1996. 49–56.
same limitations Že.g., perfect decoupling might be impos- w9x K. Najim, Process Modelling and Control in Chemical Engineering,
sible.. Marcel Dekker, New York, 1989, 520 pp.