You are on page 1of 4

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

Iterative learning control integrated with model predictive control for


real-time disturbance rejection of batch processes

Se-Kyu Oh and Jong Min Lee


School of Chemical and Biological Engineering, Institute of Chemical Processes, Seoul
National University, 08826 Seoul, Republic of Korea
Corresponding Authors E-mail: jongmin@snu.ac.kr
ABSTRACT: In this paper, iterative learning control (ILC) is integrated with a model predictive control (MPC)
technique to reject real-time disturbances. The proposed scheme is called iterative learning model predictive con-
trol (ILMPC). In the proposed ILMPC scheme, delta input formulation is used for offset-free control. Furthermore,
we use a generalized objective function to independently tune weighting factors of manipulated variable change
with respect to both the time index and batch horizons. The structure of the proposed ILMPC is similar to that of
the conventional MPC. Therefore, the proposed algorithm is easy to tune and convenient to apply.

Keywords: Iterative learning control; Model predictive control; Offset-free control; Disturbance rejection

1 Introduction
Iterative learning control (ILC) is an effective control technique for reference tracking problem of batch processes.
Model predictive control (MPC) is a popular optimization-based control strategy based on a prediction model.
General ILC can achieve exponential or monotonic convergence along the batch index; however, cannot reject
real-time disturbances because ILC is basically open-loop control within a batch. On the other hand, MPC can
reject real-time disturbances; however, cannot show convergence along the batch index. In other words, MPC
shows the same tracking performance for all batches. ILC should be combined with MPC to reject real-time
disturbances of multivariable batch processes.
In most studies of ILC integrated with MPC, often referred to as iterative learning model predictive control
(ILMPC), the state vector consists of the entire error sequences of a batch (Lee et al. , 2000; Xiong et al. , 2005;
Liu and Kong , 2013). Therefore, the control calculation might not be performed within a sample time. Several
studies combine a time-wise feedback controller and batch-wise feed-forward controller separately (Chin et al. ,
2004; Lu et al. , 2015). This two-stage approach requires two optimization steps and thus has difficulties in system
analysis and parameter tuning. In addition, all the papers described above cannot tune weighting factor of the rate
of input change vector with respect to the time index independently.
In this paper, we propose a generalized ILMPC scheme, which uses the same parameters used in general MPC,
guarantees offset-free control, can reject real-time disturbance, and can tune weighting factor of the rate of input
change vector with respect to the time index independently.

2 Preliminary
2.1 Delta input formulation
We consider the following linear discrete time-invariant system.
xko (t + 1) = Ao xko (t) + Bo uk (t)
uk (t) = uk (t 1) + uk (t) (1)
yk (t) = Co xko (t)
where xko (t) Rnx is the state vector; uk (t) Rnu is the input vector; yk (t) Rny is the output vector; k is the batch
index; t is the time index. Delta input formulation uses the control increment ( uk (t)) instead of the control signal
(uk (t)). The following augmented state-space model is delta input formulation.
 o   o
A Bo
 o   o
xk (t + 1) xk (t) B
= + uk (t)
uk (t) 0 I uk (t 1) I
 o  (2)
 o  x (t)
yk (t) = C 0 k
uk (t 1)
Delta input formulation can be expressed as the following form:
xk (t + 1) = Axk (t) + B uk (t)
(3)
yk (t) = Cxk (t)

1
July 2427, 2016, Tokyo, Japan

We can obtain the following double delta formulation for ILMPC controller using the difference between the
state-space model of kth batch and k 1th batch.

xk (t + 1) = Axk (t) + B uk (t)


(4)
yk (t) = Cxk (t)

where means the increment operator with respect to the batch index, that is, xk (t) = xk (t) xk1 (t).

2.2 Prediction model


The state-space model of Eq. (4) can be rewritten as the following lifted system because ILMPC considers finite
time intervals [0, N] where N is the terminal time.

yk = G uk + Fxk (0) (5)

Using Eq. (5), we can obtain the following prediction model.

y kp (t + 1|t) = yk1
p
(t + 1) + G pm um
k (t) + Fm xk (t|t) (6)

where
CB 0 0 CA
CAB CB 0 CA2
G pm = , Fm = . (7)

.. .. .. ..
. . . . ..
CA p1 B CA p2 B CA pm B CA p
T
y kp (t + 1|t) = yk (t + 1|t)T yk (t + 2|t)T yk (t + p|t)T


p T
(t + 1) = yk1 (t + 1)T yk1 (t + 2)T yk1 (t + p)T

yk1 (8)
T
um T uk (t + 1)T uk (t + m 1)T

k (t) = uk (t)

where p is the prediction horizon and m is the control horizon. We use state estimates xk (t|t) instead of xk (t)
because we assume that all the states are not measurable. We also define the error vector as ek (t) = r(t) yk (t)
where r(t) is the reference trajectory.

3 Iterative learning model predictive controller


We use the following objective function to design the ILMPC controller.
1
min J = { kekp (t + 1|t)k2Q + k um 2 m 2
k (t)kR + kuk (t)kS } (9)
um
k (t)
2

To obtain the solution, each input term should be expressed with respect to um
k (t) as follows:

um m m
k (t) = IL uk (t) + IL uk1 (t) Im uk (t 1)
(10)
um m
k (t) = IL uk (t) Im uk (t 1)

where
I 0 0 0 0 I
I I 0 0 0 0

0 I I 0 0 0
n mnu m
IL = . .. R u , Im = . Rnu mnu (11)

.. .. . . ..
.. . . . . . ..

0 0 0 I 0 0
0 0 0 I I 0
The following analytical solution can be obtained using Eqs. (6), (9), and (10).
1
uk (t) =Im T IL T GTpm QG pm IL + IL T RIL + S
p  (12)
IL T GTpm Q ek1 (t + 1) + G pm Im uk (t 1) Fm xk (t|t) + IL T R Im uk (t 1) IL um
 
k1 (t)

Convergence proof can be found in Lee et al. (2000). In the proposed algorithm, um k (t) goes to the zero as
k for all time t but ekp (t) goes to the very small value e (t) as k because of the soft constraint about
p
umk (t). If the weighting factor R is zero, the error ek (t) also goes to the zero as k . However, the input
trajectory can show excessive variation. We will present a rigorous convergence proof of the proposed algorithm
in the future work.

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

Reference trajectory
(a) Output trajectory (b)
Unknown disturbance
34 40 34 40

30 30 30 30
u1

u1
y1

y1
20 20
24 24
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Time Time Time Time
34 40 34 40

30 30 30 30
u2

u2
y2

y2

20 20
24 24
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Time Time Time Time
34 40 34 40

30 30 30 30
u3

u3
y3

y3

20 20
24 24
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Time Time Time Time
34 40 34 40

30 30 30 30
u8
u8

y8
y8

20 20
24 24
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Time Time Time Time
34 40 34 40

30 30 30 30
u9
u9

y9
y9

20 20
24 24
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
Time Time Time Time

Figure 1: The performance of the proposed ILMPC. (a) Q = I, R = 0.01I, S = 0.01I. (b) Q = I, R = 0, S = 0.01I.

3
July 2427, 2016, Tokyo, Japan

4 Illustrative example
We consider a cooling jacket temperature (T j ) control of a nonlinear batch reactor. A second-order exothermic
reaction A B occurs. It is assumed that T j is directly manipulated.
 
dT UA HV E
= (T T j ) k0 exp CA2
dt MC p MC p RT
  (13)
dCA E
= k0 exp CA2
dt RT
where T and CA are the state variables, T is the output variable, and T j is the input variable. The following
parameters were used for plant:
UA
= 0.09 (l/min)
MC p
HV
= 1.64 (K l/mol)
MC p
k0 = 2.53 1019 (l/mol min) (14)
E
= 13, 550 (K)
R
T (0) = 25 (oC)
CA (0) = 0.9 (mol/l)
We obtained the following linear discrete-time model using the least squares method with step input of 26 o C and
the sampling interval of 1 min. We assumed that the system is second-order.
   
0.9153 0.0416 1
x(t + 1) = x(t) + u(t)
0.0313 0 0 (15)
 
y(t) = 0.0436 1.3600 x(t)

We assume that there exist two output disturbances. One is repetitive disturbance for all batches, and the other is
non-repetitive disturbance at the 9th batch. Disturbance model is 1/(2s + 1); and disturbance input is unit step. The
prediction horizon and the control horizon are 80 and 10, respectively. In Fig. 1, the output trajectory converges to
the reference trajectory under the disturbances. The case of Fig. 1 (a) uses the weighting factor of the rate of input
change with respect to the time index (R). Therefore, smooth input change is maintained. The case of Fig. 2 (b)
only uses the weighting factor of the rate of input change with respect to the batch index (S). In the early batches
of the case (b), the rate of input change is similar to the case (a). At the 8 th batch, input change of the case (b) is
more aggressive than input change of case (a) because the input trajectory in the case (b) converges to the optimal
input trajectory for perfect tracking. However, it can lead to impractical input trajectory when the output trajectory
perfectly converges to the reference trajectory including the angular points. Hence, the weighting factor of the rate
of input change with respect to the time index is required to obtain an practical input trajectory.

Acknowledgements
This research was supported by Basic Science Research Program through the National Research Foundation of
Korea (NRF) funded by the Ministry of Science, ICT & Future Planning (2015R1A1A1A05001310)

References
Lee J.H., Lee K.S., Kim W.C., Model-based iterative learning control with a quadratic criterion for time-varying
linear systems, Automatica, 36, 641-657 (2000)
Xiong Z.H., Zhang J., Wang X., Xu Y.M., Tracking control for batch processes through integrating batch-to-batch
iterative learning control and within-batch on-line control, Industrial & Engineering Chemistry Research, 44,
3983-3992 (2005)
Liu X.J., and Kong X.B., Nonlinear fuzzy model predictive iterative learning control for drum-type boiler-turbine
system, Journal of Process Control, 23, 1023-1040 (2013)
Chin I., Qin S.J., Lee K.S., Cho M., A two-stage iterative learning control technique combined with real-time
feedback for independent disturbance rejection, Automatica, 40, 1913-1922 (2004)
Lu J.Y., Cao Z.X., Wang Z., Gao F.R.,A two-stage design of two-dimensional model predictive iterative learning
control for nonrepetitive disturbance attenuation, Industrial & Engineering Chemistry Research, 54, 5683-5689
(2015)

You might also like