DEVELOPMENT OF A NEURO-CONTROLLER FOR NONLINEAR PROCESSES

Abstract

Neurocontrollers are seen to he one of the best tools for most of t/te complex and nonlinear chemical processes control problems. Marty new techniques for neural network based modeling and control of chemical processes are being proposed, studied and reported. from so want' research areas and the achievements in the field are encouraging and rewarding to tel oil, for the nonlinear processes modelling and cotltrol. Contribution on the basis of'suc/i issues was a target for the present t/resis. The study is conducted with three principal objectives in ,,und. The first and most important objective was to suggest and develop a neural network based controller, comparatively better, from the other well known neurocontrollers mainly with respect to its better performance for nonlinear processes control problem. The second objective is to study the ltenro_model representation mechanisms and process identification, including neural networks training algorithms for better process identification. The third and the final objective is to integrate the work and the results display in interactive mode. The objective of process control is to influence the behavior of dynamical process. The latter includes maintaining the outputs of systems at constant valrres(vegala(ioll) or forcing tbellt to follow prescribed time functions(tracking). The control problem is to determine the control inputs to the syste/lt using all available data, achieving fast and accurate process control even while assuring stability. One of the major difficulties in chemical processes control is of rtorrliitearity, trllcertainity, inseparability, and complexity of the processes themselves. And hence, the application of artificial neural networks for the modelling and control of such processes seems a proper choice and development. Artificial neural networks having the capability of' nonlinear mapping are still the better tools for the solution of such problems. Hence, this study is done only otr three nettrocontrollers sampling the others. These were: direct netlrocontroller, a single step predictive controller, and the adaptive self'ttrrling ncrrrocontroller, using the capability of backpropagation neural network to learn arbitrary aolrlilreariiv. lit this work, using computer programming in C++, simulation stut/p on the three uerrrocontrollers stated above, the strategy of their dynamic modelling and the capabilit' of neural lretworks.for such: performances are conducted, alid comparative results are reported based on an es.amn pie problem on co11ti11No11,x'ly stirred tank reactor. for the slimy on the idcnfifcatlolr capabilities of nerrrtll networks, the choice is given to bacirpropagatioll static and dynamic identification ahl,'orilhrl►, with a brief work on the Cascade Correlation also for the sake of comparison.