Neural Network Predictive Control Of Nonlinear Time-delay Systems

Nonlinear time-delay systems usually appear in the industrial process, but are difficult to be controlled. Due to the universal approximation of neural network for arbitrary nonlinear mapping and the capability of the predictive technique for solving the problems of time-delay and uncertainty, neural network predictive control for nonlinear time-delay systems rapidly developed in recent years and becomes one of the most important means for nonlinear process control.With regards to the nonlinear time-delay systems, d-step-ahead predictive model of neural network predictive control is adopted. This dissertation realizes the RTRL (real time recurrent learning) algorithm of parallel model of neural network predictive control for nonlinear time-delay systems for the first time. It describes advantages of RTRL algorithm of parallel model, compared with BP algorithm of series-parallel model.This dissertation points out the impropriety of existing NARMA model correction strategy for time-delay systems. The neural network predictive control of time-delay systems was taken as an example to present the correct NARMA model correction methods.This dissertation introduces a new orthogonal neural network (ONN), and gives a general equation for determining the completeness processing elements (neuron) of multivariable input neural network for the first time. But the drawback of ONN is that the number of completeness processing elements will increase exponentially with the increasing of the input variables. Several new methods of trimming completeness processing elements are proposed to reduce the processing elements. By simulation experiment, it is validated that orthogonal neural network with these reduction-processing-elements methods has good performance, compared with BPNN.In this dissertation, we use ONN with some feedback loops to form recurrent neural network, called orthogonal recurrent neural network (ORNN). It is applied in time delay identification and input order identification for the first time. Simulation results show:1. In time delay identification, ORNN has excellent performance in convergence speed, training time, and approximate ability, compared with recurrent BPNN. 2. In input order identification, due to the ONN's ability of globalconvergence, ORNN can well identify the input order of time-delay systems, while recurrent BPNN can't.