Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

The 7th International Symposium on Design, Operation and Control of Chemical Processes (PSE ASIA 2016)

Data-driven PID Gain Tuning from Regulatory Control Data Based on


Generalized Minimum Variance Evaluation

Ryoko YOKOYAMAa∗ , and Shiro MASUDAa


a Department System Design, Tokyo Metropolitan University, 191-0065 Hino, Tokyo, JAPAN
∗ Corresponding Author’s E-mail: smasuda@tmu.ac.jp
ABSTRACT: The data-driven PID gain tuning directly tunes PID gains for disturbance attenuation based
on generalized minimum variance evaluation from closed-loop data. Since the approach is derived based on the
Controlled Auto-Regressive Integrated Moving Average (CARIMA) model, the PID gain tuning has only been
treated for stochastic disturbances. From the viewpoint of industrial applications to real processes, disturbance
attenuation for both deterministic signals and stochastic signals should be examined. The present work proposes the
data-driven PID gain tuning method when both deterministic and stochastic disturbances are fed to the controlled
system. The proposed method employs a CARIMA model with deterministic disturbances as a process model, and
gives a comparison result between the model-based cost criterion incorporated with a sinusoidal disturbance and
the iteratively updated data-driven cost criterion. The result implies that the data-driven PID gain tuning can almost
achieve similar control performance to the model-based PID gain tuning. Finally, a numerical example shows that
the proposed data-driven PID gain tuning successfully suppresses both colored noise and a sinusoidal disturbance.

Keywords: PID Control; Data-driven Control Parameter Tuning; Generalized Minimum Variance Control

1 Introduction
The Proportional-Integral-Derivative (PID) control plays an important role for keeping safe and efficient operation
in industrial process systems. The PID gain tuning is crucial in order to improve control performance of the
PID control. Thus, many researches on PID gain tuning have been studied so far. The Ziegler-Nichols and the
related methods are well-known, and have been widely used (Åstrom et al. (2006)). However, these approaches
often suffer from insufficient control performance because definite control performance criteria have not been
considered. In addition, regulatory control data without any plant test are not tractable by the conventional PID
gain tuning methods.
On the other hand, we have recently proposed a PID gain tuning method based on generalized minimum vari-
ance evaluation (Yokoyama et al., (2015)). The method can derive the PID gains for reducing the variance of
the generalized output from regulatory control data disturbed by colored noise. The PID gains are obtained by
optimizing the data-driven cost criterion without a process model. The analytical result on the performance im-
provement has been provided by clarifying the relation between the data-driven cost criterion and the model-based
one. However, since the earlier work has formulated the data-driven PID gain tuning based on the Controlled
Auto-Regressive Integrated Moving Average (CARIMA) model, the stochastic disturbance modeled by colored
noise has only been treated. From the viewpoint of industrial applications to real processes, deterministic distur-
bances should be suppressed as well as stochastic disturbances.
The present work considers a CARIMA model with deterministic disturbances as a process model, and analyzes
how a sinusoidal disturbance gives influence on the variance of the generalized output. Then, after deriving the
model-based cost criterion incorporated with a sinusoidal disturbance, it is compared with the iteratively updated
data-driven cost criterion. From the analytical study, it follows that the data-driven cost criterion has the same
value as the model-based one in the case where the PID gains obtained by the iteratively updated data-driven PID
gain tuning converge to a fixed point. The effectiveness of the proposed method is demonstrated through numerical
simulation using a CARIMA model with a sinusoidal disturbance.

2 Data-driven generalized minimum variance control


Consider the following CARIMA model (Fig. 1) as a controlled process

∆A(z−1 )y(t) = B(z−1 )z−km ∆u(t) + D(z−1 )w(t), (1)



 A(z−1 ) = 1 + a1 z−1 + · · · + an z−n
B(z−1 ) = b0 + b1 z−1 + · · · + bm z−m

D(z−1 ) = 1 + d1 z−1 + · · · + dl z−l
where u(t) and y(t) denote the input and the output, respectively. w(t) is white noise with zero mean and variance
σw2 . z−1 is the backward shift operator defined as z−1 y(t) = y(t − 1), and ∆ stands for 1 − z−1 . km denotes the delay
time of the controlled process.

1
July 24–27, 2016, Tokyo, Japan

w(t ) D( z −1 )
∆A( z −1 )

∆u (t ) B( z −1 ) − k ++ y (t )
z m
∆A( z −1 )
Figure 1: Block diagram of CARIMA model

The generalized minimum variance(GMV) control aims to minimize the variance of the generalized output
defined as (2).
ϕ (t ) = P(z−1 )y(t + km ) + Q(z−1 )∆u(t).
{+ km−1 (2)
P(z ) = 1 + p1 z−1 + · · · + pn p z−n p
Q(z−1 ) = q0 + q1 z−1 + · · · + qnq z−nq
The criterion of GMV control can be expressed as
[ ]
J = E ϕ 2 (t + km ) , (3)
where E [·] denotes the expectation operator.
To derive the GMV control law, the following Diophantine equations are now introduced.
P(z−1 )D(z−1 ) = ∆A(z−1 )E ∗ (z−1 ) + z−km F ∗ (z−1 ), (4)
∗ −1 ∗ −1 −1 −1 −1
S (z ) = E (z )B(z ) + Q(z )D(z ). (5)
From (4) and (5), the generalized output ϕ (t + km ) can be calculated as
F ∗ (z−1 ) S∗ (z−1 )
ϕ (t + km ) = y(t) + ∆u(t) + E ∗ (z−1 )w(t + km ). (6)
D(z−1 ) D(z−1 )
Noting that the degree of E ∗ (z−1 ) is km − 1, E ∗ (z−1 )w(t + km ) turns out to be composed of the data at future time
from t + 1. Since the first and second terms of the right-hand side of (6) are composed of the signals up to the
present time t, these terms are uncorrelated to the third term of the right-hand side of (6). Thus, the criterion
representing the variance of ϕ (t + km ) can be expressed as
[ ∗ −1 ]
F (z ) S∗ (z−1 ) 2
J = Var −1
y(t) + −1
∆u(t) + E ∗ (z−1 ) 2 σw2 , (7)
D(z ) D(z )
where Var [·] and ∥·∥2 denote the variance of signals and H2 norm of discrete-time transfer functions, respectively.
The cost criterion described as (7) clearly derives the GMV control law, which is represented as
F ∗ (z−1 )
∆u(t) = y(t), (8)
S∗ (z−1 )
2
and the GMV is, then, expressed as E ∗ (z−1 ) 2 σw2 .
As is shown above, the GMV control law (8) is derived using the process model (1). On the other hand, the
data-driven control law is derived from the initial process input and output data instead of the process model.
Subsequently, the derivation of the data-driven GMV cost criterion will briefly be shown.
It is assumed that the process parameters, the coefficients of the polynomial A(z−1 ) and B(z−1 ), are unknown,
and the disturbance model D(z−1 ) is known. In addition, the initial process input and output data ∆u0 (t) and y0 (t)
are assumed to be obtained from the closed-loop system, where the initial controller stabilizes the closed-loop
system. Noting that the closed-loop data are generated by white noise w0 (t) in the regulatory control. By using
∆u0 (t) and y0 (t), the generalized output ϕ0 (t + km ) is constructed as
ϕ0 (t + km ) = P(z−1 )y0 (t + km ) + Q(z−1 )∆u0 (t). (9)
Using the Diophantine equations (4) and (5), the generalized output shown in (9) can also be represented as
F ∗ (z−1 ) S∗ (z−1 )
ϕ0 (t + km ) = y0 (t) + ∆u0 (t) + E ∗ (z−1 )w0 (t). (10)
D(z−1 ) D(z−1 )
In the following, the coefficients of F ∗ (z−1 ) and S∗ (z−1 ) are denoted as ρF∗ and ρS∗ , respectively. For deriving the
data-driven GMV cost criterion using (10), ρF∗ and ρS∗ are replaced with unknown parameter vectors ρF and ρS ,
respectively. The following residual term is defined as
F(ρF ) S(ρS )
ϕres (t + km ) = ϕ0 (t + km ) − y0 (t) − ∆u0 (t). (11)
D D

2
The 7th International Symposium on Design, Operation and Control of Chemical Processes (PSE ASIA 2016)

From the equation (11), the argument z−1 is omitted for simplicity of notations. The cost criterion for the data-
driven GMV control is introduced as the variance of ϕres (t + km ) in (11).
[ 2 ]
Jdata = E ϕres (t + km ) . (12)

The cost criterion (12) is optimized in the case of ρF = ρF∗ and ρS = ρS∗ , which leads to the optimal value of the
cost criterion ∥E ∗ ∥22 σw2 . Thus, ρF∗ and ρS∗ optimize both the data-driven GMV cost criterion and the model-based
GMV cost criterion. However, the earlier work (Ando et al. (2014)) pointed out that the optimal variables for the
data-driven GMV cost criterion are not determined uniquely. Hence, the earlier work (Ando et al. (2014)) has
modified the cost criterion so that the optimization variables become the process parameters.
Instead of such the change of the optimization variables, the present work formulates the data-driven cost
criterion restricting the controller structure to a PID controller. The restriction of the control structure contributes
to not only bring the uniqueness of the optimal variables, but also facilitate applying the data-driven approach to
industrial processes. The detail design procedure and analysis are given in the next section.

3 Data-driven PID gain tuning based on GMV evaluation


For a PID controller, consider the following controller structure.
( )
Ts Td 2
∆u(t) = −K p + ∆ + ∆ y(t),
Ti Ts
L(ρL )
=− y(t), (13)
1
where K p , Ti , and Td are the proportional gain, the integral time, and the derivative time, respectively. Ts is the
sampling interval. ρL is the coefficient vector of the polynomial L, which equivalently represents PID gains.
Comparing the controller structure between (8) and (13), it follows that the polynomial F and S are corresponding
to the polynomial L and 1, respectively. From the relation, the data-driven cost criterion restricted to the PID
controller is introduced as

[ 2 ]
JPID = E ϕPID (t + km ) . (14)
L(ρL ) 1
ϕPID = ϕ0 (t + km ) − y0 (t) − ∆u0 (t). (15)
D D
In the earlier work (Yokoyama et al., (2015)), the relationship between the model-based criterion and the data-
driven criterion was examined. From that relationship, it is revealed that the data-driven PID gain tuning can almost
achieve similar control performance to the model-based PID gain tuning.

4 The PID gain tuning based on GMV evaluation with deterministic disturbances
4.1 The data-driven cost criterion with deterministic disturbances
This section proposes a data-driven PID gain tuning method based on GMV evaluation with deterministic distur-
bances. To begin with, we extend the CARIMA model to the case where deterministic disturbances are fed to the
controlled process.
Consider the controlled system shown in Fig. 2. The system can be described as the following equation,

Bz−km D
y(t) = ∆u(t) + w(t) + v(t), (16)
∆A ∆A
where v(t) is a deterministic disturbance. From (16) and the Diophantine equation (4), the km -step output prediction

v(t )
w(t ) D( z ) −1 +
+
∆A( z −1 )

∆u (t ) B ( z −1 ) − k m + y (t )
+
z
∆A( z −1 )
Figure 2: Block diagram of controlled system Figure 3: Block diagram of closed-loop system with de-
terministic disturbances

3
July 24–27, 2016, Tokyo, Japan

filtered by P can be expressed as


∆AE ∗ F∗
Py(t + km ) = y(t + km ) + y(t),
D D
E ∗B F∗ ∆AE ∗
= ∆u(t) + y(t) + v(t + km ) + E ∗ w(t + km ). (17)
D D D
Using (5) and (17), the generalized output can be expressed as
S∗ F∗ ∆AE ∗
ϕ (t + km ) = ∆u(t) + y(t) + v(t + km ) + E ∗ w(t + km ). (18)
D D D
Noting that the last term and the sum of the other three terms in the right-hand side of (18) are uncorrelated with
each other, the variance of ϕ (t + km ) is represented as
[ ∗ ]
S F∗ ∆AE ∗
Var [ϕ (t + km )] = Var ∆u(t) + y(t) + v(t + km ) + ∥E ∗ ∥22 σw2 . (19)
D D D
Using (19), we define the model-based cost criterion with deterministic disturbances in the case of the PID con-
troller (13) as follows. [ ∗ ]
F − S∗ L(ρL ) ∆AE ∗
JM = Var y(t) + v(t + km ) (20)
D D
Noting that the closed-output y(t) is generated by disturbance signals w(t) and v(t), we get
D ∆A
y(t) = w(t) + v(t). (21)
∆A + BL(ρL )z−km ∆A + BL(ρL )z−km
Substituting (21) into (20), the model-based criterion JM is
[ ]
F ∗ − S∗ L(ρL ) ∆A (F ∗ − S∗ L(ρL )) ∆AE ∗
JM = Var w(t) + v(t) + v(t + km ) ,
∆A + BL(ρL )z−km D (∆A + BL(ρL )z−km ) D
[ ( ) ]
F ∗ − S∗ L(ρL ) ∆A Pzkm − QL(ρL )
= Var w(t) + v(t) . (22)
∆A + BL(ρL )z−km ∆A + BL(ρL )z−km

The optimal PID gains can be derived by minimizing the model-based cost criterion. Let the deterministic distur-
bance v(t) be a sinusoidal signal with the period ω20πT s and the amplitude l, which leads to

( ) 2
F ∗ − S∗ L(ρL ) 2 2 l 2 ∆A P(e jω0 Ts )km − QL(ρL )
JM =
∆A + BL(ρL )z−km σw + 2 ∆A + BL(ρL )(e jω0 Ts )−km . (23)
2

where the second term of the right-hand side of (23) is the square of the magnitude of the frequency response at
ω0 .
The model-based cost criterion (23) uses the disturbance model in order to take account of the influence of
the deterministic disturbance. In contrast to the model-based cost criterion, the data-driven cost criterion (14) can
deal with the influence of the deterministic disturbance because the collected data contain the information on the
disturbance. In the next section, we analyze this point.

4.2 Analysis of the cost criterion


Let the initial input and output data ∆u0 (t) and y0 (t) be collected in the presence of the white noise w0 (t) and the
deterministic disturbance v0 (t) shown in Fig. 3. Thus, ϕPID (t + km ) can be expressed as

F ∗ − L(ρL ) − L(ρL0 )(S∗ − 1) ∆A PDzkm − L(ρL ) − L(ρL0 )(QD − 1)


ϕPID (t + km ) = w0 (t) + v0 (t) + E ∗ w0 (t + km ),
∆A + BL(ρL )z−km
0 D ∆A + BL(ρL0 )z−km
(24)
where ρL0 is an initial control parameter vector which equivalently represents an initial PID gains. When v0 (t) is
the sinusoidal signal, the variance of ϕPID (t + km ) becomes

F − L(ρL ) − L(ρL0 )(S∗ − 1) 2 2
Var [ϕPID (t + km )] =

σw

∆A + BL(ρL0 )z−km 2
2
2
l ∆A PD(e j ω 0 Ts ) − L(ρL ) − L(ρL0 )(QD − 1)
km
∗ 2 2
+
2 D ∆A + BL(ρL0 )(e jω0 Ts )−km + ∥E ∥2 σw . (25)

4
The 7th International Symposium on Design, Operation and Control of Chemical Processes (PSE ASIA 2016)

Obviously, the minimization of JPID in (14) is equivalent to the minimization of JD defined as



F − L(ρL ) − L(ρL0 )(S∗ − 1) 2 2 l 2 ∆A PD(e jω0 Ts )km − L(ρL ) − L(ρL0 )(QD − 1) 2
JD =
σw +

. (26)
∆A + BL(ρL0 )z−km 2
2 D ∆A + BL(ρL0 )(e jω0 Ts )−km

JD appears to be different from JM mainly because JD depends on the initial parameter vector ρL , but these cost
criteria have the interesting relation.
First, when ρL = ρL0 , JD and JM are described as
( ) 2
F ∗ − S∗ L(ρL0 ) 2 2 l 2 ∆A P(e jω0 Ts )km − QL(ρL0 )
JM = JD =
∆A + BL(ρ 0 )z−km σw + 2 ∆A + BL(ρ 0 )(e jω0 Ts )−km . (27)
L 2 L

In addition, when ρL converges to the fixed point ρL∗ through the iteratively updated data-driven PID gain tuning,
JD and JM at the fixed point are equivalent as is shown in (28)
( ) 2
F ∗ − S∗ L(ρL∗ ) 2 2 l 2 ∆A P(e jω0 Ts )km − QL(ρL∗ )
JM = JD =
∆A + BL(ρ ∗ )z−km σw + 2 ∆A + BL(ρ ∗ )(e jω0 Ts )−km . (28)
L 2 L

The relations between JD and JM in (27) and (28) show that the data-driven PID gain tuning can achieve the same
control performance as the model-based PID gain tuning when the parameter vector obtained by the iteratively
updated data-driven PID gain tuning converges to the fixed point even in the presence of a sinusoidal disturbance.
Hence, if the fixed point makes the data-driven cost criterion smaller, it also makes the model-based cost criterion
smaller.

5 Numerical example
In order to validate the proposed method with a sinusoidal disturbance, a simulation example is presented in this
section. The iterative data-driven PID tuning was performed to investigate whether the iterative tuning converges
on the neighborhood of the optimal PID gains derived by using the model-based criterion.
Consider the following continuous-time process model:
10s + 1
G(s) = e−2s . (29)
(7s + 1)(3s + 1)(0.5s + 1)
G(s) was discretized by sampling interval Ts = 1.0.
B(z−1 ) −2
G(z−1 ) = z . (30)
A(z−1 )
{
A(z−1 ) = 1 − 1.719z−1 + 0.835z−2 − 0.084z−3
B(z−1 ) = 0.235 − 0.105z−1 − 0.098z−2
The numerical experiment considers the controlled process (1) where the polynomial A, B and the delay time
km are given as (30), and the disturbance model is D = 1 − 0.9z−1 . w(t) is zero-mean Gaussian white noise with
σw2 = 1, and v(t) is a sinusoidal disturbance with the amplitude 3, and the period 4π .
The initial input and output data {u0 , y0 } of 5000 samples were collected with the initial PID gains set at
[ ] [ ]
K p Ti Td = 0.1 1.0 0.1 . (31)
The initial variances of the generalized output, input, and output were calculated as
[ ] [ ]
Var[ϕ ] Var[u] Var[y] = 45.49 0.634 56.02 . (32)
Using the initial input and output data, the data-driven GMV cost function for PID gain tuning defined as
(2), (14), (15) was designed. Here, the weighting functions are P = Q = 1. In the numerical experiment, the
iterative PID gain tuning was carried out until 20 times. The iterative PID gain tuning repeats the series of calcu-
lations: running the simulation experiment after implementing the PID gains, collecting the process input-output
data, and calculating the PID gains by optimizing the data-driven cost criterion derived from the latest collected
process input-output data. The PID gains by optimizing the model-based cost criterion (23) was also calculated
for comparison. To calculate the optimization of the cost criteria, ’fminsearch.m’ in the Optimization Toolbox,
MATLAB/Simulink Ver. 8.6.0.267246 (R2015b) was empolyed.
Table 1 summarizes the PID gains and the variance of generalized outputs in the case of the iterative PID
gain tuning and the model-based PID gain tuning. Fig. 4 shows the simulation results of the input signals and
the output signals in the case of the iterative PID gain tuning and the model-based PID gain tuning. From the
simulation results, we can see that the PID gains converge to the fixed point by the iterative PID gain tuning, and
the variance of the generalized output at the fixed point belongs to the neighborhood of the optimal value based on
the model-based cost criterion. Thus, it follows that the data-driven PID gain tuning works well in the presence of
both the stochastic signals and a sinusoidal disturbance.

5
July 24–27, 2016, Tokyo, Japan

Table 1: Achieved PID gains and variance of the closed-loop data by iterative tuning
Iteration Kp Ti Td Var [ϕ ] Var [u] Var [y]
Initial 0.1 1.0 0.1 45.489 0.634 56.019
1st 0.791 4.313 1.697 15.493 8.426 18.953
2nd 1.416 5.807 0.440 9.216 11.889 14.167
5th 1.035 2.667 0.034 5.966 11.358 20.737
10th 1.046 2.261 0.427 5.174 11.797 16.914
20th 0.998 2.082 0.576 5.163 11.467 16.606
Model-based solution 1.002 2.150 0.479 5.146 11.346 17.275

10
10
5
Output

Input
0 0

-5
-10
-10
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Step Step

(a) Output data (b) Input data


Figure 4: Comparison of the initial data and updated data (red-line: initial, green-line: one-shot tunig, blue-line:
iterative tuning)

6 Conclusion
This paper extended the data-driven PID gain tuning method for disturbance attenuation based on GMV evaluation
to the case where a sinusoidal disturbance is fed to the controlled process as well as stochastic disturbances.
The comparison study between the data-driven cost criterion and the model-based cost criterion with a sinusoidal
disturbance was presented. From the analytical study, it turned out that the data-driven cost criterion has the same
value as the model-based one in the case where the PID gains obtained by the iteratively updated data-driven PID
gain tuning converge to a fixed point. A numerical example showed that the data-driven PID gain tuning worked
well even when a sinusoidal disturbance was fed to the controlled process in addition to stochastic disturbances.
Since the way how data-driven cost criterion is costructed from the closed-loop system is not changed even in the
presence of a sinusoidal disturbance, we can say that it has the ablity to derive the adequate PID gains from the
closed-loop data.
For the future work, there remains the study of the convergence property of the iterative tuning using the data-
driven PID gain tuning. The application of the proposed method to the industrial process is also an important
issue.

Acknowledgements
This work was partially supported by JSPS KAKENHI Grant Number 25420446.

References
Åstrom, K. and Hägglund T., Advanced PID Control, ISA (2006)
Ando, K., Masuda S., and Kano M., Data-driven generalized minimum variance regulatory control, 2014 European
Control Conference, 418-423 (2014)
Yokoyama, R., Masuda S., and Kano M., M., Data-driven Generalized Minimum Variance Regulatory Control for
Model-free PID Gain Tuning, IEEE International Conference on Control Applications (CCA), 82-87 (2015)

You might also like