摘要
锂离子电池的剩余使用寿命(RUL)是健康管理中重要参数,其准确评估对于保证电池设备的安全稳定运行非常重要。该文提出一种数据预处理联合变分模态分解(VMD)、长短期记忆网络(LSTM)和高斯回归过程(GPR)的预测框架。首先选取充放电循环过程中的信息作为间接健康因子(HI),并通过核主元分析方法(KPCA)实现间接HI的特征提取,完成数据预处理;其次通过VMD-LSTM方法实现健康因子的分解、预测和重构,并将重构得到的数据应用于RUL预测的GPR模型,完成预测模型搭建;最后以NASA锂电池数据集作为算法测试数据,结果表明,所提取的健康因子能够准确跟踪锂电池的退化过程;所提预测方法能够准确地估计电池的剩余寿命,同时具有较高的可靠性和稳定性。
The performance of lithium-ion batteriescandeteriorate with the decrease of capacity and the increase of impedance during continuous charging and discharging process,which poses a risk of equipment and system failures,includingcatastrophic losses.Accurate and reliable prediction of the remaining useful life(RUL)of lithium-ion batteries is crucial.However,previous prediction methods mainly supportedpoint predictions without offering a clear mathematical representation of the confidence level of the prediction results.The noise and capacity rebound phenomena in the original data are ignored.Therefore,this paper proposes a prediction framework based on data pre-processing combined with variational mode decomposition(VMD),long and short-term memory network(LSTM),and Gaussian regression process(GPR).Firstly,anindirect health indicator(HI)reflecting the life degradation trend of lithium-ion batteriesisextracted from the charge/discharge curve.Pearson and Spearman correlation coefficients are used to verify the correlation between the extracted indirect HI and capacity.The kernel principal element analysis(KPCA)method reduces computational complexity by removingredundant components of indirect HI,transform them into fusion HI.Then,the VMD decomposition method decomposes the fusion HI into multiple modal components.Based on the central frequency of modal components and capacity-related coefficients,multiple modal components are divided into three parts:global attenuation,local regeneration,and other noise.Time series prediction of the global attenuation and local regeneration components are performed separately using LSTM neural networks to obtain the predicted values after the prediction starting point.The values of each component aresummed up to obtain the reconstructed HI.Finally,the reconstructed HI and capacity serve as the input and output of the GPR model RUL prediction,respectively.NASA lithium-ion battery public data set is used for experimental validation,and different prediction starting points are set for each battery.The starting point set is at about 40%of the total cycle.The experimental results show that the predicted capacity values closely track the battery aging trend and effectively estimate the local regeneration phenomenon during the aging process.The true capacity values generally fall within the 95%confidence interval of the predicted value,and the prediction effect improves as the training data increase.Regarding evaluation indexes for capacity prediction results,the maximum root mean square error,mean absolute error,and mean absolute percentage error are 2.98%,2.34%,and 1.81%,respectively.Errors in the remaining useful life prediction are all within 2 cycles.The following conclusions can be drawn from the experimental analysis:(1)The KPCA algorithm removesredundant information between indirect health indicators to reduce data complexity and achieve data pre-processing.(2)The VMD decomposition method minesintrinsic information,capturing the long-term downward trend,local regeneration,and noise component to remove implied noise in the data.(3)The GPR model supports probabilistic prediction and provides confidence intervals for capacity prediction results.
作者
李英顺
阚宏达
郭占男
王德彪
王铖
Li Yingshun;Kan Hongda;Guo Zhannan;Wang Debiao;Wang Cheng(School of Control Science and Engineering Dalian University of Technology,Dalian 116000 China;Shenyang Shunyi Science and Technology,Co.Shenyang,110000 China)
出处
《电工技术学报》
EI
CSCD
北大核心
2024年第10期3244-3258,共15页
Transactions of China Electrotechnical Society
基金
辽宁省“兴辽英才计划”资助项目(XLYC1903015)。
关键词
锂离子电池
剩余寿命
健康因子
变分模态分解
高斯回归过程
长短期记忆
Lithium-ion battery
remaining useful life
health indicator
variational mode decomposition
Gaussian regression process
long and short-term memory