Application Of An Artificial Neural NetworkTo On-line Control Of A Process

Abstract

Although the development of linear control theory is well established, real industrial processes contain significant non-linearities that give limited credence to the optimal performance of controllers over a wide operational region. Consequently, the development of techniques that enable the design of a controller suitable for operation with a non-linear process would be beneficial. This paper describes the development and implementation of an on-line, one-step-ahead, optimal predictive controller incorporating a neural network model of the process. The scheme is based on a Multi-Layered Perceptron neural network as a modelling tool for a real non-linear, dual tank, liquid level process. The model validation techniques are described as well as the choice of network structure and topology. The ability of the trained neural network to represent both a simulation of the process, modelled from first principles, and the actual process is investigated. The implementation, of the optimal control algorithm, to both the simulation and the real process are described. Results are presented to illustrate the steady-state and transient performance of the control scheme.