Multistep predictive controllers based on neural network

• Jorge Alaña Universidad del Zulia-Venezuela
• Carlos Vinante Universidad del Zulia-Venezuela

Abstract

Although the potential for modeling dynamic systems, intrinsic in the recurrent neural network arquitectures is superior to the dynamic representational capabilities of the traditional feedfoward networks, they have not been widely used for model based control of nonlinear systems; possibly due the difficulties with the training algorithms. In this paper, we present a multi step nonlinear predictive controller, using recurrent neural networks trained with a new simplified algorithm, to model nonlinear multi step predictors. The proposed controller was then evaluated using it to control a rather difficult, highly nonlinear CSTR with a first order irreversible exothermic reaction. Its performance was satisfactory in the whole range of operation, and considerably better than the one obtained with the same predictive controller structure but using as a model, for the multi step prediction, the traditional concatenated sequence of one step ahead prediction static feedfoward networks. Based on the simulation results, we also present an analysis of the effects penalizing the control actions on their deviations, on the closed loop steady state offsets, in the presence of imperfect predictors