Research Of Nonlinear Predictive Control Algorithm Based On Neural Networks And PSO

Nonlinear model predictive control has become a research hotspot in the field of control theory and industry.Neural networks can learn and adapt to the characteristics of nonlinear systems,so that neural network based nonlinear predictive control has received much attention from scholars.In this thesis,a nonlinear predictive control algorithm based on the Multi-step Predictive Error Index Function is proposed to solve the problem that the recurrent prediction decrease the prediction accuracy and the prediction model is difficult to adapt to the timevarying system which is existed in neural network based nonlinear predictive control.Doubleswarm Division and Cooperation Particle Swarm Optimization is proposed for the local optimal problem.Based on the improved pruning neural network model and the improved time window online modeling algorithm,an adaptive predictive control algorithm is proposed.The proposed algorithms have achieved satisfactory simulation results.The main contributions of this thesis are as follows:Concerning the poor multi-step prediction precision of neural network prediction model,according to the concept of control relevant identification,a Multi-step Predictive Error Index Function which is suitable for predictive control is proposed.Then,aiming at the over-fitting problem which is often appearing in neural networks,the regularization term is added to the objective function,and the Multi-step Predictive Error and Regularization Index Function is proposed,which improves the accuracy and generalization performance of neural network multi-step prediction to a certain extent.At the same time,an improved feedback correction method is proposed and further improves the predictive precision.Finally,based on the above neural network prediction model,an unconstrained neural network predictive controller designing method is given,and the simulation analysis is carried out on a discrete system without noise and with noise.And satisfactory results have been achieved.Aiming at the local optimal problem,particle swarm optimization algorithm is adopted to train recurrent neural network.In order to balance the global exploration and local exploitation ability of the particle swarm optimization algorithm,Double-swarm Division and Cooperation Particle Swarm Optimization is proposed from the perspective of sub-population and based on the common win-win cooperation phenomenon in nature,which further strengthens the algorithm's optimization performance,and the simulation results show the effectiveness of the proposed algorithm.Based on correlation analysis theory,and the improved particle swarm optimization algorithm is used to train the diagonal recurrent neural network,a pruning neural network modeling algorithm is designed.In order to make the prediction model can adapt to the timevarying working conditions of the controlled plant,an adaptive predictive control algorithm is proposed based on an improved time window online modeling algorithm.The simulation analysis is carried out on an anaerobic digestion model,and the final numerical simulation verifies the applicability and efficiency of the algorithm.