Early Prediction Of Remaining Useful Life Of Lithium-ion Batteries Based On Machine Learning

In recent years,with the intensification of greenhouse effect,countries have begun to transform their energy structure.Lithium-ion battery has become the preferred energy storage device in various fields by virtue of its many advantages.The performance of lithium-ion batteries will decrease with the increase of use time,and there may even be safety risks if they continue to be used in the later stage of degradation.Therefore,it is necessary to replace the lithium-ion batteries in time before their performance degrades to the End of Life(EOL),so as to ensure the safety and stability of the entire battery system.Therefore,the prediction of Remaining Useful Life(RUL)of lithium-ion battery is of great significance.The data-driven method has the advantages of simple model establishment and few internal parameters,so it has been widely used in the research of RUL prediction of lithium-ion batteries.To achieve accurate Li-ion battery RUL prediction,it usually requires 40-70% data for the entire life cycle of Li-ion batteries.Early prediction technology is a technology that directly predicts battery failure life by using early data.The implementation of this technology can greatly improve the efficiency of production verification and technical verification of lithium-ion batteries.Recently,lithium-ion battery RUL early forecast progress rapidly.According to different life prediction methods,the early RUL prediction of lithium-ion battery can be divided into two categories,namely,based on time series model and based on early degradation feature method.However,the predictive performance of time series model is close to the limit on long time series,and there is less degradation information hidden in the early degradation characteristics.Therefore,it is difficult to achieve the early prediction of high-precision RUL of lithium-ion batteries by the above two methods.In this thesis,two kinds of methods based on time series model and early degradation feature are integrated.A novel method of Particle Swarm Optimization based on Long ShortTerm Memory(LSTM)feature prediction was proposed.PSO)Support Vector Regression(SVR)model.The high precision and early prediction of RUL of Li-ion battery is realized by this method.The main research contents and results of this thesis are as follows:(1)Analyze the characteristics of lithium-ion battery data.The feature engineering of lithium-ion battery data was carried out to construct the pre-middle,middle and late interval features,and the correlation between each interval feature and battery life was analyzed by using Pearson correlation.The analysis results showed that the characteristics in the middle and late period had the highest correlation with the battery life.(2)The PSO-SVR prediction model of the remaining life of lithium-ion batteries was established.The public data of lithium-ion batteries were used for experimental verification,and the prediction results under 46 different working conditions showed that the PSO-SVR model could predict the battery under different working conditions with high accuracy,which laid a model foundation for the construction of the subsequent early prediction model.(3)A PSO-SVR model based on LSTM feature prediction method(LSTM-PSO-SVR)is proposed to realize the early prediction of RUL.For the early prediction of lithium-ion battery RUL,degradation characteristics in the middle and late period cannot be directly obtained from historical data.Therefore,it is necessary to establish a feature prediction model based on LSTM.In order to verify the performance of the LSTM-PSO-SVR model proposed in this thesis,the open data set of lithium-ion batteries was used for experimental verification and compared with a variety of early prediction methods.The results of battery prediction in four working conditions show that the root mean square error(RMSE)of the method used in this thesis is 37 cycles.Compared with the prediction results of the early feature-driven PSO-SVR model and the LSTM-PSO-SVR model without step optimization,The RMSE of LSTM-PSO-SVR model reduced 76 and 30 cycles,and the mean absolute error(MAE)decreased by 52.63% and 66.67%,respectively,both of which were the highest among the three methods.The accuracy of the LSTM-PSO-SVR method and the good comprehensive performance of the model are further illustrated by comparing the results reported in the literature.The above results show that the prediction features of the middle and late period can realize the high precision and early prediction of RUL of lithium-ion batteries,which has a high utilization value.