Neural network based predictive control for nonlinear chemical process

The paper presents a neural network based predictive control (NPC) strategy to control nonlinear chemical process or system. Multilayer perceptron neural network (MLP) is chosen to represent a Nonlinear autoregressive with exogenous signal (NARX) model of a nonlinear process. Based on the identified neural model, a generalized predictive control (GPC) algorithm is implemented to control the composition in a continuous stirred tank reactor (CSTR), whose parameters are optimally determined by solving quadratic performance index using well known Levenberg-Marquardt and Quasi-Newton algorithm. Also an Instantaneous linearization based predictive control (IPC) strategy is discussed, in which an approximated linear model is extracted from nonlinear neural network by instantaneous linearization around operating points. The tracking performance of the NPC and IPC is tested using different amplitude step function as a reference signal on CSTR application and it is shown using simulation results, that the NPC strategy is more effective and robust than the IPC strategy.

1. Introduction

Most of the industrial processes including the chemical process industry are nonlinear in nature, but still control practitioners have been using linear control techniques to control such systems. This is partly due to the fact that, over the normal operating region, many of the nonlinear processes can be approximated by their linear models which are easier to use and also since the theory for the stability analysis of linear control systems is quite well developed [1]. However, using linearization, robustness cannot be guaranteed, especially when the parameters of the plant are uncertain or there is a noise or disturbance in the process [2]. It has been proved that certain NN architectures, such as the multi-layer perceptron (MLP) networks and the radial basis function (RBF) networks, can approximate any nonlinear function to a desirable accuracy given enough hidden layer nodes and suitable weighting factors. The ability of neural networks (NN) to model any nonlinear function to an arbitrary degree of accuracy have been frequently used for modeling of nonlinear chemical processes. In some control schemes NN is either trained to operate as a controller or as an indirect control, which utilizes the NN model of the plant or process. The example of the latter is the Generalized Predictive control (GPC), originally derived for linear process models [3]. Predictive control is an optimal control strategy using the concept of receding horizon that can be proven to stabilize processes in the presence of nonlinearities and constraints. It refers to a class of computer control algorithms that control the future behavior of a plant through the use of an explicit process model. The process model plays an important role in predictive control strategy. Predictive control based on linear models is acceptable when the process operates at a single set-point and the primary use of the controller is the rejection of disturbances. However, many processes especially chemical processes are often required to operate at different set-points depending on the specification of the product to be produced. These processes make transitions over the nonlinearity of the system, so linear predictive control frequently results in poor control performance in such cases. This imposes the way for utilization of nonlinear process models in predictive control.