Nonlinear model predictive control based on multiple local linear models

A long range nonlinear predictive control strategy using multiple local linear models is proposed. The multiple local linear models are identified through recurrent neuro-fuzzy network training. In this modelling approach, the process operation is partitioned into several fuzzy operating regions. Within each region, a local linear model is used to represent the process. The global model output is obtained through the centre of gravity defuzzification which is essentially the interpolation of local model outputs. Based upon the multiple local linear models, a nonlinear model based controller is developed by combining several local linear model based predictive controllers which usually have analytical solutions. Control actions obtained based on local incremental models contain inherent integral actions eliminating static offsets in a natural way. The techniques are demonstrated by applying to pH control in a continuous stirred tank reactor.

1. Introduction

Model predictive control (MPC) technologies have been recognised as efficient means to improve operating efficiency and profitability and have been widely applied in chemical process control (Morari and Lee, 1999); Qin and Badgwell, 1997; Rawlings et al., 1994). In most of the MPC schemes a linear model of the process is assumed. If the process dynamics are relatively linear within the operating region, then the use of a linear model based control algorithm leads to good performance. However, in the situations where the process is highly nonlinear, the linearity assumption may well be detrimental to control system robustness and performance, An alternative approach is to develop nonlinear model predictive control strategies. The nonlinear models employed vary in complexity from first principles descriptions to simplified mechanistic models and empirical models developed from process input output data. Neural networks have been shown to possess good function approximation capability and have been applied to process modelling by many researchers (e.g. Bhat and McAvoy, 1990; Bulsari, 1995). Most of the reported neural network based controllers are developed in the framework of MPC (e.g. Pottmann and Seborg, 1997; Saint-Donat et al., 1991). The key to the successful implementation of nonlinear MPC is an accurate nonlinear model and an efficient optimisation algorithm.