Study Of Predictive Control Based On Neural Network In Nonlinear System

With the successful application in the process industry, predictive control is one of the most promising control methods in academia and in engineering. It has received considerable attention in the last decade. Today, it is universal phenomenon for non-minimum phase nonlinear system, multivariable nonlinear system during many industrial processes. Much effort has been made to extend the predictive control algorithm to nonlinear systems. Over the past several years, neural networks have been widely applied to identification and control of nonlinear system. Theoretical works have been proven that, even with one hidden layer, neural networks can uniformly approximate any continuous function over a compact domain. Artificial neural networks based predictive control has attracted more and more attention.In this thesis, we will address some nonlinear predictive control problems based upon neural networks. Firstly, for neural network and predictive control, the recent theoretical accomplishments and applications are thoroughly surveyed. Secondly, based on neural networks, a series of works are developed on the following topics: neural network modeling, prediction and control for highly non-minimum phase system, multivariable nonlinear system, etc. Because DRNN (Dynamic Recurrent Neural Network) is a dynamic network of nonlinear system which can simulate dynamic system and process very well, in this thesis, DRNN is adopted.In this paper, by utilizing the conventional PID controller and predictive control theory, the method is proposed under the PID-type predictive cost function. And the method is using to nonlinear system of two types: one is SISO system; the other is MIMO system. To SISO system, firstly the multi-step predictive algorithms based on neural networks are investigated for nonlinear system. It is known that the prediction results of the direct multi-step predictor are more accurate than the recursive predictor. A new direct cutting-error multi-step prediction method is proposed. The smaller multi-step prediction errors can be obtained. Then feed-forward neural networks and local recurrent neural networks are used to model and tune the controller's parameters on-line, respectively. A linear predictive PID controller is put forward. And by result of emulation it is proved that the nonlinear system has well controlling capability. To MIMO system, based on the nonlinear PID controller of SISO system, several nonlinear PID controllers are adopted in parallel. Under the decoupling cost function, a decoupling control strategy is presented for nonlinearmultivariable system. Furthermore two predict control methods of nonlinear system are studied: recursive multi-step predictive control and multi-step predictive cost function control.