Abstract:
In order to improve the computation efficiency of multi-step predictive control, a predictive control system based on Elman network with temporal differences (TD) method is proposed. Elman network is used to predict the system output of multi-step ahead directly and a new hybrid learning algorithm combining the TD method with standard back propagation (BP) algorithm to train the Elman network is put forward according to the intrinsic disadvantages of BP algorithm,which can not update network weights incrementally. To simplify computation, a single-value predictive control algorithm is used to realize the optimization of control input of the next step. Theoretical analysis and simulation results demonstrate that this method is suitable for fast real-time systems because of its characteristics of simple structure, small calculating amount and fast speed, and that it has some self-adaptability against changeable parameters of the system.