Nonlinear Model Predictive Control Using State-space Recurrent Multi-dimensional Taylor Network

In this paper, we propose a discrete-time infinite horizon nonlinear model predictive control (QIH-NMPC) based on state-space recurrent multi-dimensional Taylor network (WIN). The purpose of this paper is to construct a state-space RMTN to be used as internal predictive model for QIH-NMPC and train this network by backpropagation through time (BPTT) algorithm. The multi-dimensional Taylor network (MTN) differs from the existing neural network (NN) on its structure and dynamic performance. RMTN gains advantages over recurrent neural network (RNN) on its training efficiency and ease of use, thus it reduces the on-line nonlinear optimization burden and enhances the efficiency of computation. The stability of closed loop system is guaranteed via Lyapunov stability theory. Finally, a numeric example is given to illustrate the effectiveness of the proposed design approach.