SIMULATION STUDIES OF A NON-LINEAR OPTIMAL REGULATOR USING MULTILAYERED NEURAL NETWORKS

Abstract

In a classical linear optimal regulator design method, a mathematical model of plant dynamics must be formed by using knowledge of the plant dynamics. If the model equation is an accurate representation of the plant dynamics then the optimal control input will be generated by standard optimal regulator theory. However, in actual applications such knowledge is rarely complete, and it is difficult to express exactly real plant dynamics in mathematical equations. The real plant always includes model uncertainties such as parameter variations or non linearities that arecnown at the time of controller design. Hence, robustness against system uncertainties is indispensable to control system design. In this dissertation work, a non linear regulator problem is simulated that includes linear and non linear optimal techniques using neural networks. The LOR has an inherent robustness against a certain range of model uncertainities. However, non linear dynamics cannot be taken into consideration in the LOR design. Hence, non linear neural networks are used to overcome non-linearity and system uncertainties -present in the modelling of the plant dynamics. The salient features of this regulator are that it will use a priori knowledge of the plant dynamics as the system equation and the corresponding LOR to improve the control performance. The proposed regulator study utilises two MNN's for modelling and control. The MNN for modelling is used to obtain a more accurate model than the given mathematical equations. The MNN for control is used to overcome deficiencies by adding corrections to the linear coefficients of the LOR and by adding non linear effects on the LOR. Both the MNN's are simultaneously trained so that a desired response of the plant is obtained. Computer simulations are performed to show the applicability and limitations of the new regulator.