Neural Network Predictive Control Of Magnetic Resonance Wireless Power Transfer Syste

In recent years,Wireless Power Transfer(WPT)technology has become an important topic in the field of new energy and Power electronics.Compared with traditional technology,magnetic resonance radio energy transmission technology has great advantages in security,transmission distance,transmission efficiency and transmission load,so it has become a major development direction of radio energy transmission technology.This paper presents a method of constant voltage output of magnetic resonance radio energy transmission system based on neural network predictive control.Firstly,T and π equivalent circuits are used to analyze the input and output characteristics of the four basic topologies,and SS topologies with constant voltage input and constant voltage output are selected as the compensation network.Based on the principle of mutual inductance circuit,the mathematical models of output voltage,load power and transmission efficiency are established.The results show that under different coil coupling conditions,the system operating at natural resonant frequency can not maintain stable output voltage,high load,and can not adapt to practical applications.Second,using neural network model and predictive control method,a frequency regulation controller is designed at the transmitter end of the system,so that the constant load system can automatically track the resonant frequency under different coupling conditions,and keep the output voltage unchanged.The simulation results show that this method can adjust the resonance frequency of the system quickly and accurately,and make it reach the output voltage of stable load,so that it can keep stable voltage in a certain working area,and improve its transmission efficiency under high load.Thirdly,aiming at the problems of gradient descent method in neural network training,a momentum gradient algorithm is used to optimize the NN weight and threshold,and the current weight change direction is taken as the compromise of the current and previous gradient,so as to solve the possible oscillation problem of traditional gradient descent method in the process of parameter optimization and improve the convergence speed;Furthermore,aiming at the random selection of initial NN weight and threshold,the genetic algorithm and the quantum particle swarm optimization are proposed respectively.Using the global optimization ability of GA and QPSO algorithm,a set of optimal initial weights and thresholds are found before the training of neural network,which reduces the training time of neural network and improves the training accuracy.Finally,the feasibility of the method is verified by simulation with MATLAB software.