基于间接健康特征优化与多模型融合的锂电池SOH-RUL联合预测  

Joint Prediction of Lithium Battery State of Health and Remaining Useful Life Based on Indirect Health Features Optimization and Multi-Model Fusion

作　　者：蔡雨思 李泽文[1] 刘萍 夏向阳[1] 王文[1] Cai Yusi;Li Zewen;Liu Ping;Xia Xiangyang;Wang Wen(School of Electrical and Information Engineering Changsha Ligong University,Changsha 410000 China)

机构地区：[1]长沙理工大学电气与信息工程学院,长沙410000

出　　处：《电工技术学报》2024年第18期5883-5898,共16页Transactions of China Electrotechnical Society

基　　金：湖南省科技创新人才计划科技创新团队资助项目(2021RC4061)。

摘　　要：准确预测锂电池健康状态(SOH)与电池剩余使用寿命(RUL)对提高电池安全性能具有重要意义。而当前针对SOH和RUL的预测,存在着间接健康特征选取困难,以及使用数据驱动方法缺乏不确定性表达的问题。为此,该文提出一种基于间接健康特征优化与多模型融合的锂电池SOH-RUL联合预测方法。首先从充电电压曲线中采集多个健康特征,并通过特征并行融合方法和注意力机制进行优化处理得到间接健康特征(IHF)。然后引入贝叶斯模型平均(BMA)方法来解决预测过程中的不确定性问题,将其与支持向量回归(SVR)和长短期记忆神经网络(LSTM)相结合,构建SVR-BMA融合模型和LSTM-BMA融合模型,并分别进行SOH和RUL预测;通过自适应噪声完备集合经验模态分解(CEEMDAN)方法从SOH预测阶段的容量预测结果中提取出RUL预测的输入特征,以实现SOH和RUL的联合预测。最后利用CALCE数据集进行性能测试,实验结果表明,所提方法能有效提高SOH和RUL预测的准确性和可靠性。Lithium-ion batteries have become essential for new energy vehicles and energy storage power stations due to their high energy density,low self-discharge rate,and reliable cleanliness.However,lithium batteries gradually age with increasing cycle counts,leading to decreased battery performance.State of health(SOH)and remaining useful life(RUL)are important indicators for evaluating battery aging status,and accurate predictions of these metrics are crucial for the safe operation of energy storage systems.However,choosing suitable indirect health features(IHF)for SOH and RUL predictions is challenging.Data-driven models generate uncertainties,resulting in inaccurate predictions of SOH and RUL.Therefore,this paper proposes a joint prediction method for lithium battery SOH and RUL based on indirect health feature optimization and multi-model fusion.Firstly,multiple health factors(HF)are extracted from charging voltage curves,and IHF are obtained through feature concatenation and attention mechanism optimization,denoted as feature Co-HF.Then,Bayesian model averaging(BMA)is introduced to address uncertainties in the prediction process.With support vector regression(SVR)and long short-term memory(LSTM),SVR-BMA and LSTM-BMA models are constructed.Additionally,the input features for RUL prediction are extracted from SOH capacity predictions using the complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)method,enabling joint prediction of SOH and RUL.The proposed method is validated using the University of Maryland's CALCE lithium battery dataset.Compared with different HFs and predictions of each sub-model,the Co-HF capacity prediction results are closer to the measured values,and the tracking effect of capacity trends is ideal.The mean absolute error(MAE)and root mean squared error(RMSE)of SOH predictions using the proposed fusion model are 0.0081 and 0.0120,respectively,and the MAE and RMSE of RUL predictions are 2.56 and 4.63,respectively.The prediction errors are lower than those of each sub-model

关 键 词：电池健康状态 剩余使用寿命 间接健康特征 贝叶斯模型平均 支持向量回归 长短期记忆神经网络 

分 类 号：TM912[电气工程—电力电子与电力传动]