A Deep Learning-based Method For Predicting The Remaining Life Of Lithium-ion Batteries

In the field of energy storage,the popularity of lithium-ion batteries is increasing,however,as a complex electrochemical system,lithium-ion batteries are subject to safety degradation and performance deterioration during use,which can lead to accidents if hidden problems are not eliminated in time.The management system of Li-ion battery involves various research contents,among which,the Remaining Useful Life(RUL)research is an important part.Remaining life prediction of Li-ion batteries is an effort to predict the remaining Li-ion battery cycle cycles by analyzing the historical decline data of the batteries.The data-driven approach based on predicting the remaining life of Li-ion batteries does not require a pilot knowledge,and the remaining life can be predicted by analyzing the information in it by mining the features of the time series,and this approach has become a hot research topic in the field.In this paper,based on the analysis of the decline mechanism of lithium-ion battery,we establish the remaining life prediction method based on deep learning technology,and the main work done is as follows.(1)To address the long-term dependence problem of traditional Recurrent Neural Network(RNN)in predicting battery life,Long Short-Term Memory(LSTM)is introduced,and this problem is effectively solved by the gate structure of LSTM neural network.Based on the capacity as the health factor,we propose several new decline feature parameters,introduce the feature parameter selection algorithm based on Sequential Forward Selection(SFS),and combine the LSTM neural network to predict the remaining life of Li-ion battery.The experimental results show that the RUL prediction method combining SFS and LSTM neural network is compared with the RUL prediction method based on traditional RNN neural network,and the relative errors of RUL prediction of the former are 4%,3.5%,2.2%,and 1.8% at the prediction starting point of 100,120,140,and 160 cycles,respectively,while the latter are 6.7%,5.3%,4.4%,and 3.8%.(2)On the basis of the above work,for the shortcoming that the SFS algorithm can only add features but not reduce them in the process of selecting a subset of lithium-ion battery decline features,the sequential floating for-ward selection(SFFS)algorithm is applied to carry out the feature selection in both directions before and after,which can better select At the same time,the LSTM neural network in battery life prediction is further optimized by applying Root Mean Square Prop(RMSprop)technique to reduce the oscillation amplitude of parameters and improve the convergence speed,introducing Dropout technique to shield neurons and solve the computational complexity and overfitting problems,introducing Attention mechanism is introduced to update the feature weights and give more weight to parameters with high relevance,and Gated Recurrent Unit(GRU)is used to improve the basic structure of LSTM to further simplify the network and reduce the complexity.The experimental results show that the RUL prediction method combining SFFS and improved LSTM has higher accuracy than the RUL prediction method combining SFS and LSTM,and the relative errors of the former RUL prediction are 2.9%,2.3%,1.5%,and 0.7% at the prediction starting point of 100,120,140,and 160 cycles,respectively,while the latter are 3.9%,3.6%,2.3%,and 1.8% for the latter.In addition,the former still has higher prediction accuracy when using only 80% of the cycle data collected by the latter,and 10% lower relative to the latter in terms of RUL prediction relative error when the prediction starting point is 140 cycles.