Prediction Of Remaining Useful Life Of Lithium-ion Batteries Based On Deep Learnin

The lithium-ion battery(LIB)has become an indispensable part of our daily electronic equipment because of its high energy density and long cycle life.With the increase in global carbon dioxide emissions and the continuous development of LIBs,the electrification of the automotive industry has become a general trend.Electrification of transport infrastructure is one of the best choices to achieve sustainable energy development and low-carbon travel.However,with the continuous increase of battery charge/discharge cycle,LIBs will inevitably appear to age and capacity attenuation,and the internal resistance will increase,which brings potential safety risks.Because of the complexity of the battery degradation process,it is a serious challenge to accurately estimate how long the battery can last.In the actual operation of the battery,the battery material system and charging/discharging strategies are different,so it is difficult to estimate the damage to the battery capacity caused by each charge and discharge.Moreover,most of the current studies on LIB life prediction are based on specific cycle conditions,such as specific charge/discharge strategies and constant environment temperature.The practical application of LIBs has complex ambient temperatures and distinct user differences in charge/discharge habits.In the case of new energy vehicles,even if they have excellent thermal management systems,the temperature of the battery module will still be affected by the ambient temperature.In addition,the market positioning of the vehicle will also affect the aging curve of the battery module.This puts forward a severe test for the practicability and versatility of the battery remaining useful life(RUL)prediction model.Given this,this work proposes three deep-learning battery RUL prediction models based on directly measurable variables such as charging current,voltage,temperature,and historical discharge capacity,which effectively solves the interference caused by the random workload in the battery discharge process and enhances the robustness of the model.The research content of this work is divided into three parts:(1)A deep convolutional neural network(DCNN)was used to extract aging features from the original charging process variables,and the high correlation between historical and predicted capacity curves was fully considered by lagging the historical capacity data for input and incorporating a long and short-term memory neural network to model the RUL of the battery.The Bayesian optimization algorithm was used to search for the best structure and process hyper-parameters for the model.Prediction performance validation under different training strategies and prediction starting points was completed using the National Aeronautics and Space Administration(NASA)battery ageing dataset,demonstrating the feasibility of the proposed model.(2)A novel adaptive feature separable convolution(AFSC)algorithm was proposed,and a predictor which could be applied to both early life prediction and real-time RUL prediction of LIB was developed by combining the convolutional long short-term memory neural network(Conv LSTM).In addition,the embedding of the attention mechanism(AM)enhanced the model’s ability to extracted key features and effectively filtered useless information.By comparing the performance of the single subnet model,the sub-model without AM,and the full model,the role of each algorithm in the full model was explored.The excellent performance of the proposed model was demonstrated by comparing it with many published works.Finally,through the visualization of the attention learning weight and feature map,the feature processing process was explored,which enhances the interpretability of the model.(3)Introducing Transformer to the field of LIB RUL prediction,to our knowledge,few current studies have been performed in this area.In contrast to the first two parts of the work,Transformer eschewed the traditional LSTM networks and was a deep learning model based entirely on a self-attention mechanism,and a deepening of the AM,with a multi-headed attention mechanism that captureed multi-scale ageing features.The work on capacity estimation and RUL prediction based on a series of prediction starting points demonstrated the great potential of Transformer in RUL prediction.