采用全局健康因子和残差模型的锂离子电池健康状态估计  

Estimation of Lithium-Ion Battery State of Health Using Global Health Factor and Residual Model

作　　者：胡循泉 耿莉敏[1] 舒俊豪 张文博 巫春玲[2] 尉小龙 黄东 陈昊[1] HU Xunquan;GENG Limin;SHU Junhao;ZHANG Wenbo;WU Chunling;WEI Xiaolong;HUANG Dong;CHEN Hao(Shaanxi Key Laboratory of New Transportation Energy and Automotive Energy Saving,Chang’an University,Xi’an 710064,China;Xi’an Key Laboratory of Advanced Transport Power Machinery,Chang’an University,Xi’an 710064,China;Automotive Engineering Research Institute,Shaanxi Heavy Vehicle Co.,Ltd.,Xi’an 710200,China;Nuclear Industry 203 Research Institute,Xianyang,Shaanxi 712099,China)

机构地区：[1]长安大学陕西省交通新能源开发、应用与汽车节能重点实验室,710064西安 [2]长安大学西安市交通先进动力重点实验室,710064西安 [3]陕西重型汽车有限公司汽车工程研究院,710200西安 [4]核工业二○三研究所,712099陕西咸阳

出　　处：《西安交通大学学报》2025年第4期105-117,共13页Journal of Xi'an Jiaotong University

基　　金：陕西省自然科学基础研究计划资助项目(2023-JC-QN-0417);西安市科技计划资助项目(24ZDCYJSGG0048);咸阳市重点研发计划资助项目(L2023-ZDYF-SF-077)。

摘　　要：为准确估计锂离子电池的健康状态(SOH),提出了一种卷积神经网络-残差网络-双向门控循环单元-注意力机制(CNN-Residual-BiGRU-Attention)模型和微调估计方法。首先,采用分段近似聚合算法对电池容量增量和恒流充电曲线进行降维,构建全局健康因子;接着,利用卷积神经网络提取全局健康因子时序特征,通过注意力机制突出强相关特征,并引入残差网络保持信息完整性;最后,通过改进人工蜂群算法对模型超参数寻优,提升模型SOH估计精度。采用美国国家航空航天局和牛津大学锂离子电池数据集进行精度验证,结果表明:利用提出的微调估计方法,即使精度较差的卷积神经-长短期记忆模型,SOH估计结果的平均绝对误差e_( MAE)、平均绝对百分比误差e_( MAPE)和均方根误差e RMSE也均在2%以内;相较于卷积神经网络-双向门控循环单元-注意力机制模型,采用CNN-Residual-BiGRU-Attention模型对训练集比例为30%的同一电池SOH进行估计,得到的e_( MAE)、e_( MAPE)和e RMSE分别降低了41.86%、44.35%、42.11%;对训练集比例为40%的同类电池SOH进行估计,得到的e_( MAE)、e_( MAPE)和e RMSE分别降低了45.51%、45.93%、40.10%。该研究结果可为低比例训练集条件下准确估计锂离子电池的SOH提供理论参考。To accurately estimate the state of health(SOH)of lithium-ion batteries,a convolutional neural network-residual-BiGRU-attention(CNN-Residual-BiGRU-Attention)model and[HJ]fine-tuning estimation method are proposed.Firstly,a piecewise approximate aggregation algorithm is used to reduce the dimensionality of battery capacity increment and constant current charging curves to construct a global health factor.Secondly,the CNN is used to extract the temporal features of the global health factor,highlighting strong correlations through an attention mechanism,and the residual network is introduced to maintain information integrity.Finally,by improving the artificial bee colony algorithm for model hyperparameter optimization,the accuracy of SOH estimation is enhanced.The accuracy is validated using lithium-ion battery datasets from NASA and the University of Oxford.The results show that,even with a less accurate convolutional neural network-long short-term memory model,the proposed fine-tuning estimation method keeps the mean absolute error e_( MAE),mean absolute percentage error e_( MAPE),and root mean square error e RMSE of the SOH estimation results within 2%.Compared with the CNN-BiGRU-Attention model,when the CNN-Residual-BiGRU-Attention model is employed to estimate the SOH of the same battery with a training set ratio of 30%,the e_( MAE)、e_( MAPE) and e RMSE of the estimated results are reduced by 41.86%,44.35%and 42.11%,respectively.For estimating the SOH of the same type of batteries with a training set ratio of 40%,the e_( MAE)、e_( MAPE) and e RMSE of the estimated results are reduced by 45.51%,45.93%,and 40.10%,respectively.These research findings provide a theoretical reference for accurately estimating the SOH of lithium-ion batteries under low training set conditions.

关 键 词：锂离子电池 健康状态估计 全局健康因子 改进人工蜂群算法 残差 双向门控循环单元 

分 类 号：TM912.8[电气工程—电力电子与电力传动]