Remaining Useful Life Prediction Of Lithium-ion Battery Based On Improved Particle Filter

Lithium-ion batteries have the characteristics of reusability,high energy density,and long cycle life,and are widely used in electronic products,electric vehicles,and aerospace and other fields.Due to the loss of internal chemical reactions and the complexity of the external environment,the battery performance will gradually decrease until it fails,which may lead to major safety accidents.Predicting the remaining useful life(RUL)and arranging maintenance in advance is an effective way to solve the problems caused by battery failure.This article has carried out the following work on the lithium-ion battery RUL prediction method:First,based on the non-linear and non-Gaussian characteristics of lithium-ion battery capacity degradation,a double exponential degradation model is selected as the prediction model.In order to obtain reliable model parameters,after the relevance vector points are obtained by the relevance vector machine(RVM)that can obtain the sparse model,the non-linear least squares method is used to fit these characteristic points to obtain the initial model.Secondly,to solve the problem of insufficient prediction accuracy of the traditional particle filter algorithm,the support vector regression optimized by the genetic algorithm(GA-SVR)is used to improve the unscented particle filter(UPF)algorithm,and a fusion algorithm of UPF and GA-SVR(UPF-GA-SVR)is proposed to improve the accuracy of prediction.UPF algorithm generates particle filter suggested density function through unscented Kalman filter,which can effectively reduce the problems of particle degradation and lack of particle diversity,and greatly improve filtering accuracy;GA-SVR uses GA algorithm to optimize SVR and predict measurement error,solve the problem that there is no real measurement value after the prediction point,and the filter cannot update the state vector,and improve the prediction accuracy.Finally,based on the RVM initialization model,the NASA battery data set is selected,the UPF-GA-SVR algorithm and other common methods are simulated and compared.The experimental results show that based on the RVM initialization model,the UPF-GA-SVR algorithm has higher filtering accuracy and RUL prediction accuracy in the case of strong process noise and observation noise.