Neural network based direct optimizing predictive control with on-line PID gradient optimization

In this paper, a neural network model-based predictive control has been developed to solve problems of nonlinear process control. In the proposed control scheme, a neural network model with recurrent connections is employed to describe nonlinear dynamic processes. Based on the neural network model, a nonlinear d-step-ahead predictor is constructed. The nonlinear predictive control is directly formulated as an on-line nonlinear programming problem (NLP). To improve the performance of the back-propagation algorithm, a PID instantaneous gradient descent optimisation algorithm, as motivated by the Proportional-Integral-Differential (PID) control strategy, is proposed for the on-line NLP. The applications of the nonlinear predictive control scheme to nonlinear processes including a continuous-stirred-tank-reactor (CSTR) is finally presented.