Research On Chaotic Time Series Prediction Of Motor System Based On Neural Network

Chaos theory is the qualitative study of unstable and aperiodic behavior in deterministic nonlinear dynamical systems.Chaotic time series are nonlinear time series generated by the evolution of chaotic systems over time,which has many characteristics such as non-periodicity,pseudo-randomness,fractal,long-term unpredictability,etc..And its coverage is wide,involving many natural sciences and social sciences.Permanent magnet synchronous motor(PMSM)has many advantages such as high power efficiency,reliable structure,small size,light weight and low energy consumption,and is widely used in industrial production and daily life.It is found that the PMSM system will have chaotic oscillations under certain conditions,which will have an important impact on the stable operation of the motor system.Therefore,it is of great significance to predict the chaotic time series of the motor system for proposing measures to ensure the stable operation of the motor.On the other hand,the Artificial Neural Network(ANN)model is widely used to solve nonlinear system problems due to its excellent memory ability,self-learning ability,and function approximation ability.At present,the Echo State Network(ESN)and the Long Short Term Memory Network(LSTM)are two artificial neural network models,which are widely used in the field of nonlinear systems.With the deepening of research,ESN model and LSTM model gradually exposed some shortcomings.This paper improves and optimizes the traditional ESN model and LSTM model to further improve the performance of artificial neural network model in the chaotic time series prediction of PMSM system.The details are shown as follows:(1)Firstly,this paper reviews the research background and significance of the issue,and then introduces the research status of chaos theory,artificial neural networks and PMSM systems.Finally,the relevant theoretical knowledge is introduced to provide theoretical support for the work of this paper.(2)Traditional ESN models require larger computational data sets and higher computational costs when facing more complex nonlinear systems.The current research points out that in some experiments,there is an urgent need for a method that can quickly process the chaotic time series prediction of nonlinear systems.In this paper,it is first proved that the traditional ESN network model can be reduced to a polynomial regression model by conducting perturbation ablation experiments on the key parameters of the traditional ESN model,which is called as a data-driven regularized regression model(D2R2).Then,the simplified model is applied to the chaotic time series prediction of PMSM system,and experiments show that D2R2 can realize the short-term dynamic prediction of PMSM system and reconstruct a long-term nonlinear dynamic system based on existing variable data.Finally,by introducing a comparison model for comparative experiments,it is found that D2R2 can achieve a significant improvement in computational efficiency at the expense of acceptable prediction accuracy.(3)Although traditional ESN models perform well in nonlinear prediction tasks,they still have many shortcomings.Recent studies have shown that the nonlinear vector autoregression(NVAR)with a specific construction is equivalent to the linear reservoir computing(RC)with a linear readout layer.Based on the latest research,this paper first proposes the Next Generation Reservoir Computing(NG-RC),which essentially replaces the reserve pool in traditional reserve pool computing with a nonlinear vector autoregressive process.Then,based on NG-RC,the chaotic time series prediction of PMSM system is predicted,and experimental simulation proves that NG-RC can realize the chaotic time series prediction of PMSM system,reconstruct a long-term nonlinear dynamic system based on existing data,and infer unknown variables through existing variable data.Finally,by comparing with the traditional RC model,the results show that NG-RC can achieve higher prediction accuracy with smaller data sets.At the same time,time complexity experiments show that NG-RC has higher computational efficiency than traditional RC models.(4)The LSTM model solves the problem of lack of long-term data correlation in the traditional recurrent neural network(RNN),but the LSTM model itself still has problems such as random selection of relevant parameters.Sparrow Search Algorithm(SSA)has been applied by researchers to optimize LSTM model parameters because of its good optimization ability,and has achieved certain research results.With the deepening of research,the standard SSA algorithm has gradually exposed the problems of slow convergence effect of the algorithm and easy to fall into local optimization.This paper combines various strategies to optimize the standard sparrow search algorithm.First,elite chaotic reverse learning strategy is used to generate the initial population,which enhance the quality of the initial individuals and the diversity of the population,and realize the exploration of more high-quality search areas to improve the local extremum escape ability and convergence performance of the algorithm.Then,the follower position update strategy is optimized through the chicken swarm algorithm,and then the Gauss-Cauchy variation is introduced to improve the local extreme escape ability and global search ability of the algorithm,and the optimized sparrow search algorithm is called the Improved Sparrow Search Algorithm(ISSA).The ISSA algorithm is applied to solve the parameter optimization problem of the LSTM(ISSA-LSTM).Finally,ISSA-LSTM is applied to PMSM system prediction.Simulation experiments and comparative experiments show that the ISSA-LSTM model not only realizes the chaotic time series prediction of PMSM system,but also has higher prediction accuracy than other models.