Model-Based Iterative Learning Control for Batch Processes Using
Generalized Hinging Hyperplanes
Xiaodong Yu,
†,‡
Zhihua Xiong,
†,‡
Dexian Huang,*
,†,‡
and Yongheng Jiang
†,‡
†
Institute of Process Control and Engineering, Department of Automation, Tsinghua University, Beijing 100084, China
‡
Tsinghua National Laboratory for Information Science and Technology, Tsinghua University, Beijing 100084, China
ABSTRACT: A model-based iterative learning control (ILC) strategy using the generalized hinging hyperplanes (GHH) is
proposed to track the product quality trajectory in the batch process. As an empirical model with piecewise affine basis functions,
GHH is very suitable for constructing the dynamic model of batch processes, in which its gradient information can be easily
obtained due to the structure of GHH model. Based on the GHH, a quadratic-criterion-based ILC (Q-ILC) algorithm is
constructed, where the input trajectory for the next batch is updated by ILC law and the output tracking error can be gradually
reduced from batch to batch. The proposed strategy is demonstrated on a simulated typical batch reactor and compared with the
method based on neural networks. The simulation results show the convergence of the output tracking error and the robustness
of the proposed method under model−plant mismatches and unknown disturbances.
1. INTRODUCTION
Generally, batch processes run intermittently and are very
suitable for low-volume and high-value products, while
continuous processes are proper to make high-volume products
continuously.
1
A batch process usually performs a given task
repetitively over a fixed period of time (called a batch or trial),
and it has been paid more attention for the past decade because
it plays more important roles in the chemical industry.
2−4
Because of the repetitiveness and periodicity in nature, batch
processes are usually run based on the results and experiences
of the previous batches, which is in accordance with the basic
idea of iterative learning control (ILC) strategy. ILC updates
control signals for the next batch based on the information
from previous batches, and then the output trajectory can
converge asymptotically to the reference trajectory. The initial
explicit formulation of ILC was presented by Uchiyama in
1978.
5
Although many significant improvements of ILC have
been achieved in both industry and academia, these develop-
ments are mainly based on model-free approaches.
6
At the
beginning of the 21st century, model-based ILC algorithm with
a quadratic criterion (Q-ILC) for time-varying linear systems
was proposed.
7−9
Thereafter, the ILC strategy was fused with
model predictive control (MPC) to build an integrated control
technique for end-product and transient profile.
6
However,
these above-mentioned methods are mainly based on linear
models.
To extend the ILC to nonlinear batch processes, neural
networks (NN) and supported vector machine (SVM)
methods were used to build the process model and then ILC
law was formed based on the linearization of the nonlinear
model.
10−12
Zhang
11
proposed an ILC method using a feed-
forward neural network model, where the network was
linearized around the current batch and the control policy for
the next batch was updated by ILC based on the linearized
model. In our previous work,
13
we combined a special control-
affine feed-forward neural network with Q-ILC scheme for
batch processes, and with the help of the gradient information
obtained from the predictive model, the ILC law can be
obtained theoretically. However, it is still difficult to analyze the
convergence performance explicitly.
The hinging hyperplanes (HH) model was first introduced
by Breiman in 1993.
14
The hinge functions in the HH are used
as basis functions, rather than the sigmoid functions in NN. HH
becomes ridge construction with an additional linear term and
yields a type of piecewise-linear model.
15
Compared with other
nonlinear model approximation methods, there are several
advantages of the HH model.
14
For example, the upper bound
of the approximation error can be easily determined, and all
parameters employed in the HH model can be estimated using
a fast and efficient least-squares (LS) method. With the help of
this type of piecewise-linear model, most of the existing linear
analysis techniques become available.
16
By modifying the basis
functions of HH, Wang et al.
17
proposed a general model
structure, which is called generalized hinging hyperplanes
(GHH), and they also proved that GHH possessed more
flexibility of nonlinear black-box modeling. Because GHH is
also a type of piecewise-linear model and its performance is
more accurate than NN, GHH is used here to build the model
of nonlinear batch processes.
In this paper, by making some compromises between the
approximation capability of the model and the simplicity for
controller design, the GHH model-based ILC strategy (called
as GHH−ILC) for batch processes is proposed. With the
predictive model of batch processes constructed by GHH, the
input trajectory is updated by the quadratic-criterion-based ILC
with the help of the gradient information obtained easily from
the GHH model. Meanwhile, model predictions of GHH are
modified by the predictive errors of the immediately previous
batch, in order to improve the model accuracy.
Received: August 17, 2011
Revised: November 26, 2012
Accepted: November 26, 2012
Published: November 26, 2012
Article
pubs.acs.org/IECR
© 2012 American Chemical Society 1627 dx.doi.org/10.1021/ie201842a | Ind. Eng. Chem. Res. 2013, 52, 1627−1634