Data-Based Approach To Feedback-Feedforward Controller Design From Closed-Loop Plant Data
Data-Based Approach To Feedback-Feedforward Controller Design From Closed-Loop Plant Data
Data-Based Approach To Feedback-Feedforward Controller Design From Closed-Loop Plant Data
ABSTRACT: A novel data-based method for designing the feedback and feedforward controllers is proposed.
The controller design directly exploits closed-loop plant data and does not require identifying any process model.
Based on the closed-loop response data to a set-point change, the feedback PID controller is first designed such
that the resulting feedback loop behaves as closely as possible to the prescribed reference model for set-point
tracking. Subsequently, based on the closed-loop response data to a measurable disturbance input, a feedforward
lead-lag compensator is designed to allow the effects of the load disturbance to be minimized effectively.
Simulation examples have demonstrated the effectiveness of the proposed overall methodology.
Keywords: Process Control; Data-Based Controller Design; Feedforward Control; Controller Tuning.
1 Introduction
Feedforward control from measurable disturbances provides a possibility to take control actions before
disturbance response has occurred in the process output. This fact makes it a very powerful complement to the
feedback controller. The design of feedback and feedforward controllers requires knowledge of how the
controlled variable responds to changes in the manipulated variable and the disturbance variable. In the model-
based design methods, this knowledge is represented by low-order process and disturbance models which must
be derived before designing the controllers. Open-loop experiments are typically required to collect appropriate
plant data for identifying these models. This may be time consuming and may upset routine process operations.
Closed-loop identification techniques are the main alternatives that can be used to overcome this problem
(Veronesi and Visioli, 2014). Nevertheless, the major problem of model-based design methods is that
identifying the models from plant data involves approximations. Because the performance of feedback and
feedforward controllers depends on the model accuracy, the effectiveness of model-based design methods
degrades for higher-order process dynamics owing to inevitable modeling errors.
An attractive approach for alleviating the drawback of plant-model mismatch is to design the controllers
from plant data directly without resorting to the identified models. The use of data-based (model-free) approach
for controller design is mainly developed for the feedback controllers (e.g., Jeng et al. (2014) and the references
therein), whereas the application of data-based design method to the feedforward compensator has received little
attention. In this study, we developed a novel data-based method for designing the feedback and feedforward
controllers based on the idea of virtual reference feedback tuning (VRFT) (Campi et al., 2002). The controller
design directly exploits closed-loop plant data and does not require identifying process and disturbance models.
Thus, it is possible to use routine operating data without terminating the operating control loop, which is
obviously preferable in industrial settings. Because the integrals of plant data are used in the regression
equations for calculating the controller parameters, the proposed method is robust against measurement noise.
The necessary modifications to deal with unrealizable feedforward compensators are discussed. The proposed
method provides an overall design of the feedback-feedforward control system that shows satisfactory
performance for both set-point tracking and compensation of measurable disturbances.
Figure 1. (a) Feedback control system. (b) Reference model for feedback controller tuning.
The available output data y(t) are used, and the reference signal in Figure 1(b) can be obtained as
R( s ) = T ( s )1 Y ( s ) where R ( s ) is termed as the virtual reference signal because it does not actually exist.
Because Y(s) is considered to be the desired output of the closed-loop system when the reference signal is set to
R ( s ) , the corresponding PID controllers output (the virtual controller output) can be calculated by
1
U ( s ) = GC ( s ) R ( s ) Y ( s ) = K C 1 + + D s T ( s )1 1 Y ( s ) (1)
Is
When the process is subject to the input signal u(t), it generates y(t). Therefore, a PID controller that shapes the
closed-loop transfer function to the reference model generates u(t) or its Laplace transform U(s) when the error
signal is given by R ( s ) Y ( s ) . Because u(t) is known, the problem of controller design is equivalent to
minimizing the difference between the available u(t) and u (t ) (the time-domain signal of U ( s ) given in eq.(1)).
The reference model should satisfy T(0) = 1 for offset-free tracking, and it is specified as a typical form:
1
T (s) = e s (2)
s +1
where is related to the apparent time delay of the process, and is an adjustable parameter to manage the
trade-off between control performance and system robustness. Substituting eq.(2) into eq.(1) and performing
rearrangement yields
K 1 1
U ( s ) = K C ( s + 1) e s 1 Y ( s ) + C + e s Y ( s ) + K C D ( s 2 + s ) e s s Y ( s ) (3)
I s s
Expressing eq.(3) in time-domain and performing integration three times with respect to t yields
u ( ) = K C y ( + ) + y ( + ) y ( )
(3) ( 2) (3) (3)
[ 0, t ] [ 0, t ]
y ( ) ; k = [ K C K C D ]
t ( 2) ( 2)
D (t ) = y ( + ) d + y ( + ) KC I
T
0 [ 0, t ] [ 0, t ]
Then eq.(4) can be expressed as (t ) = (t )T k . The aforementioned minimization problem (i.e., minimizing the
difference between u(t) and u (t ) ) is equivalent to minimizing the difference between (t ) and (t ) . Choose ti,
i = 1, 2, , N such that 0 < t1 < t2 < < t N . The minimization problem is formulated as a linear regression:
min J = = k
2 2
2 2
(7)
k
where
= [ (t1 ) (t 2 ) (t N ) ] ; = (t1 ) (t2 ) (t N ) ; = [ (t1 ) (t2 ) (t N ) ] (8)
T T T
Calculation of depends on the values of and , which are not known a priori because the process is
unidentified. For a given value of , the parameter can be selected to match a designer-specified robustness
level in terms of the maximum sensitivity (MS) using the following criterion (Jeng et al., 2014)
2
The 7th International Symposium on Design, Operation and Control of Chemical Processes (PSE ASIA 2016)
0.7289 M S + 1.555
= ; 1.2 M S 2.0 (9)
M S 1.006
Thus, when a desired value of MS is specified, the minimization problem given in eq.(7) depends implicitly only
on ; therefore, the optimal solution k* becomes a function of . To determine the optimal for controller tuning,
the tuning problem is further formulated as
min min J = k
2
2
(10)
k
For a given , the inner minimization in eq.(10) (eq.(7)) can be solved by the least-squares method as follows:
( )
1
k ( ) = T T (11)
The minimization problem given in eq.(10) then becomes the following problem:
2
min J = k ( ) (12)
2
The smallest J() is identified in a specified range of (i.e., J(*)), and its corresponding solution k*(*) is used
to obtain the PID controller parameters. Given a searching range for , this one-dimensional minimization
problem can be readily solved using standard optimization algorithms such as the golden-section search.
uC ( ) = a u ( ) + b
(4) ( 4) (3) (3) (3) ( 4)
u ( ) uC ( ) d ( ) + c d ( ) (17)
[ 0, t ] [ 0,t ] [0,t ] [0,t ] [0,t ] [0, t ]
T
(19)
[0,t ] d ( ) [ 0,t ] d ( ) ; p = [ a b c ]
(t ) = uC ( ) u ( )
(3) (3) T (3) ( 4)
[0, t ] [0, t]
Then eq.(17) can be expressed as (t ) = (t )T p . Minimizing the difference between u(t) and u (t ) is equivalent
to minimizing the difference between (t ) and (t ) . Choose ti, i = 1, 2, , N such that 0 < t1 < t2 < < t N . The
minimization problem is formulated as a linear regression:
min J f = = p
2 2
2 2
(20)
p
where
= [ (t1 ) (t2 ) (t N ) ] ; = [ (t1 ) (t 2 ) (t N ) ] ; = [ (t1 ) (t2 ) (t N )] (21)
T T T
Calculation of depends on the values of (or f), which is not known a priori. Therefore, the optimal
solution p* becomes a function of . To determine the optimal for designing Gf(s), the design problem is
further formulated as
min min J f = p
2
2
(22)
p
For a given , the inner minimization in eq.(22) (eq.(20)) can be solved by the least-squares method as follows:
( )
1
p ( ) = T T (23)
The minimization problem given in eq.(22) then becomes the following problem:
2
min J f = p ( ) (24)
2
The smallest Jf ( ) is identified in a specified range of (i.e., Jf ( * ) ), and its corresponding solution p* ( * )
and f = * are the parameters of the feedforward compensator.
4
The 7th International Symposium on Design, Operation and Control of Chemical Processes (PSE ASIA 2016)
Step 2. If f (0) < 0, set f = 0, and then reduce the value of a, if a(0) > 0, by the following rule (Guzman and
Hagglund, 2011) to take the unrealizable delay into account. Denote the current parameters of a and f as a(1)
and f (1), respectively.
f(0)
a = max a (0) + , 0 (25)
1.7
Step 3. If a (1) < b ( c ) , set a = b ( c ) , and then set f = max f(1) a + a (1) , 0 . ( )
4 Simulation Examples
4.1 Example 1
The following system is considered:
e 2 s e 4 s
G (s) = ; Gd ( s ) = (26)
( s + 1) 3
0.5s + 1
Initially, the PID parameters are KC = 1.067 , I = 4 , and D = 0.75 and no feedforward action is employed.
The PID controller is retuned using the proposed method based on the set-point unit step response of the initial
control system. Choosing a desired MS of 1.65, the optimal is * = 2.86 ( = 1.57), and the corresponding PID
parameters are KC = 0.708 , I = 3.10 , and D = 0.936 . Then, a feedforward compensator is designed based on
the response of the initial control system to a unit step input in the disturbance, and the results are a = 0.0094 ,
b = 1.70 , c = 1.0 , and f = 1.16. The compensator is unrealizable because of a < 0; therefore, it is modified
according to the proposed procedure, which yields a = 0.294 , b = 1.70 , c = 1.0 , and f = 0.856. Note that the
design parameter was chosen as 5.78 to allow the high-frequency gain of the proposed feedforward
compensator equal to that of the feedforward compensator designed by Veronesi and Visioli (2014) for a fair
comparison. Figure 3 shows the performance improvement that results from the retuning of the PID controller
and from the use of the feedforward compensator. The integrated absolute error (IAE) improves from IAE = 6.0
(initial tuning) to IAE = 5.20 for the set-point response and from IAE = 5.11 to IAE = 0.520 for the disturbance
rejection. The result of feedback-feedforward control obtained from the design of Veronesi and Visioli (2014)
(based on the identified models) is also shown; the IAE for the set-point and disturbance response are 5.60 and
1.98, respectively. Obviously, the proposed method provides superior performance for the disturbance rejection.
Set-point Disturbance
1.5 1
1 0.5
y
0.5 0
0 -0.5
0 5 10 15 20 25 0 5 10 15 20 25
Proposed
2 0
Initial
Veronesi & Visioli
1.5 -2
u
1 -4
0.5 -6
0 5 10 15 20 25 0 5 10 15 20 25
time time
CBs = 0.552 mol/L. The dilution rate (F/V) is considered as the manipulated variable to control the concentration
of Component B (CB) around the steady state, and the feed concentration is considered as a measurable
disturbance. A time delay of 0.2 min in the measurements of concentration is assumed. Notably, the process
exhibits inverse response behavior. We consider the initial control system ( KC = 0.5 , I = 0.5 , and D = 0.1 )
where a static feedforward compensator Gf(s) = 0.05 is already in place. Figure 4 shows the system responses
to a step change of 0.1 mol/L in the set-point of CB and a unit step disturbance in the feed concentration. To
simulate realistic conditions, Gaussian white noise with a standard deviation of 0.005 was added to the
measurements to represent measurement noise. The proposed method is applied to retune the PID controller and
feedforward compensator by using the noisy plant data. Based on the set-point response and a desired MS of 1.4,
the optimal is * = 0.61 ( = 0.83), and the corresponding PID parameters are KC = 0.962 , I = 1.06 , and
D = 0.334 . Then, the feedforward compensator is designed by using the disturbance response, and the results
are a = 0.534 , b = 0.044 , c = 0.111 , and f = 0.47 . The compensator is unrealizable because of f < 0;
therefore, it is modified according to Step 2 of the proposed procedure, which yields a = 0.257 , b = 0.044 ,
c = 0.111 , and f = 0. Figure 5 shows the performance improvement that results from the retuning of the PID
controller and feedforward compensator. Indeed, the integrated absolute errors are IAE = 0.233 and IAE = 0.103
with the initial tuning and IAE = 0.162 and IAE = 0.032 after the application of the proposed method for the set-
point and disturbance response, respectively.
Set-point Disturbance Set-point Disturbance
0.6 0.62 0.6
0.55 Proposed
0.6 Initial
0.55
CB (mol/L)
0.58
CB (mol/L)
CB (mol/L)
CB (mol/L)
0.58 0.5
0.5
0.56
0.45 0.56
0.45
0.54
V/F (min-1)
V/F (min-1)
Figure 4. Noisy closed-loop plant data for Example 2. Figure 5. Closed-loop response for Example 2.
5 Conclusions
We have proposed a data-based method for the systematic design of feedback-feedforward control systems.
By using the set-point response data of an existing (possibly underperforming) control system, a new feedback
PID controller is first designed, which can be used to improve the control performance for set-point tracking.
Subsequently, by using the disturbance response data of the same initial control system, a feedforward lead-lag
compensator is designed, which can be added to the feedback system to effectively compensate for the
measurable disturbance. The effectiveness of the proposed method has been demonstrated by simulation results.
Acknowledgements
The authors thank the MOST of Taiwan for supporting this research under grant 104-2221-E-027-105.
References
Campi, M. C., Lecchini, A., and Savaresi, S. M., Virtual Reference Feedback Tuning: A Direct Approach for
the Design of Feedback Controllers, Automatica, 38, 1337-1346 (2002).
Guzman, J. L., and Hagglund, T., Simple Tuning Rules for Feedforward Compensators, J. Process Control, 21,
92-102 (2011).
Jeng, J. C., Tseng, W. L., and Chiu, M. S., A One-Step Tuning Method for PID Controllers with Robustness
Specification Using Plant Step-Response Data, Chem. Eng. Res. Des., 92, 545-558 (2014).
Jeng, J. C., A Model-Free Direct Synthesis Method for PI/PID Controller Design Based on Disturbance
Rejection, Chemometrics and Intelligent Laboratory Systems, 147, 14-29 (2015).
Veronesi, M., and Visioli, A., Automatic Tuning of Feedforward Controllers for Disturbance Rejection, Ind.
Eng. Chem. Res., 53, 2764-2770 (2014).