Research On Prediction Method Of Remaining Useful Life Of Lithium Battery Based On Deep Learning

In various fields such as transportation,lithium batteries have gained popularity as energy storage devices,military manufacturing,and aerospace,thanks to the progress and application of renewable energy.However,despite their outstanding characteristics of long life and fast discharge rate,the concern for battery safety in real-world scenarios has always been a crucial point.Especially lithium batteries with end-of-life can easily cause equipment failures or safety problems,so monitoring the health status and remaining life of lithium batteries has become particularly important.At present,the prediction of the remaining life of lithium batteries is mainly carried out through researchers employing modelbased,data-driven,and fusion techniques.In this study,a deep learning model should be applied to predict the remaining battery life of lithium batteries,in order to analyze the correlation between battery capacity and battery life.These are the main topics of the paper:To begin with,further analysis of the factors affecting the performance of lithium batteries is carried out in the paper,together with an introduction to their classification,basic structure,and operating characteristics.Subsequently,a charge-discharge cycle test was carried out on lithium iron phosphate batteries based on these factors,and an analysis of the experimental results data was carried out,and it was found that the capacity of lithium batteries had the same decay trend as the remaining life.In addition,NASA’s lithium battery dataset was introduced as a supplement to further verify capacity degradation characteristics.The next step involves introducing the fundamental elements of time series and deep learning,which provide the theoretical foundations needed to predict the remaining life by analyzing lithium battery datasets.Ascertaining the algorithmic aspects of deep learning,a CNN-LSTM model is developed.Among them,CNN neural networks have a strong feature extraction ability,but they are not very good at learning long-distance dependencies.Therefore,it is proposed to use LSTM neural networks as a complement,which can handle long-time series data well.Experimental data on the capacity decay of lithium iron phosphate batteries are taken as the research object.Before the data is fed into the deep learning model,Newton interpolation is used to complete the missing data,and then normalization is used to map the values between [0,1].It is convenient to speed up the convergence speed.The preprocessed historical battery capacity data is fed into the CNN-LSTM model,and the cycle period when the capacity reaches the failure threshold is predicted.The aim is to determine the remaining life of the lithium battery using the established model.To determine the superiority of the proposed model,four additional methods are used for comparison,and the estimation criterion used is the RMSE.These analyses revealed that the prediction performance of the CNN-LSTM model is more accurate.Finally,in view of the problem that lithium battery capacity cannot be directly obtained in practical scenarios,it is proposed to use indirect health factors to predict the remaining life.This part uses the NASA 18650 lithium battery dataset to study the trend of current,voltage and temperature changes in the discharge state,so as to extract the constant current discharge time,discharge cutoff voltage time and discharge time to the highest temperature as the capacity characterization,and through statistical indicators,it is proved that these three indirect health factors have a strong correlation with the actual capacity.In order to effectively utilize the data characteristics of indirect health factors,based on the CNN-LSTM model,an attention mechanism was proposed to predict the remaining life of lithium batteries.The attention mechanism can flexibly and accurately assign different weights to the input vector.Experimental results show that the prediction model with the introduction of the attention mechanism is more accurate,and it still maintains good stability for data prediction of small samples.