数据分布多样性对锂电池SOC预测的泛化影响  

Generalized impact of data distribution diversity on SOC prediction of lithium battery

作　　者：何林[1,2] 刘江岩 刘彬[1,2] 李夔宁[1,2] 代帅 HE Lin;LIU Jiangyan;LIU Bin;LI Kuining;DAI Shuai(Key Laboratory of Low-Grade Energy Utilization Technology and System of Ministry of Education,Chongqing 400044,China;College of Energy and Power Engineering,Chongqing University,Chongqing 400044,China)

机构地区：[1]重庆大学低品位能源利用技术及系统教育部重点实验室,重庆400044 [2]重庆大学能源与动力工程学院,重庆400044

出　　处：《储能科学与技术》2024年第5期1677-1687,共11页Energy Storage Science and Technology

摘　　要：数据驱动模型预测荷电状态(SOC)依赖高质量的实验数据,在应用于实际使用场景下的分布多样的锂电池组数据时会出现预测的准确性不稳定即泛化能力差的情况,限制了模型的实际应用。研究实际场景下的大规模数据的分布多样性对SOC预测模型的泛化性影响具有重要意义。因此,对32个锂电池组的实际运行数据集进行研究,采用经典算法与多输入多输出(MIMO)策略结合来预测多步SOC,对每份数据分别建立模型进行SOC预测,研究了不同算法的应用效果并分析了数据分布多样性对模型的泛化能力的影响规律。结果表明:对大规模的锂电池组数据,LR-MIMO模型训练精度普遍优于RF-MIMO、KNN-MIMO、LSTM-MIMO模型,其预测未来0.5 h的SOC的R^(2)一般在0.98及以上,MAPE基本低于0.05。与其他模型相比,LR-MIMO模型有优秀的预测性能,预测其他数据集的R^(2)基本在0.95以上。而KNN-MIMO模型的预测精度与RF-MIMO模型相当,R^(2)大致在0.7以上,LSTM-MIMO模型的预测性能因数据集不同存在较明显的差异;当数据满足SOC与电压的相关系数≥0.9、SOC和电压分布范围广、核密度曲线呈左偏趋势、分布较均匀时,可使模型训练精度提高。The prediction of state-of-charge(SOC)in batteries in data-driven models relies on high-quality experimental data.However,when considering lithium battery packs with diverse distributions in real-world scenarios,the accuracy of prediction of SOC in these packs becomes unstable,leading to poor generalization ability and limiting the practical application of the models.Therefore,it is of great significance to investigate the impact of the diverse distribution of large-scale datasets from actual operating scenarios on the generalization ability of SOC prediction models.Consequently,in this study,we conducted research on 32 datasets of operational data associated with lithium battery packs.Classic algorithms were combined with a multi-input multi-output(MIMO)strategy to predict multi-step SOC.Separate models were established for each dataset to analyze the application effects of different algorithms and the influence of data distribution diversity on model generalization ability.The results demonstrated that for large-scale lithium battery pack datasets,the LR-MIMO model generally exhibited higher training accuracy compared to the RF-MIMO,KNN-MIMO,and LSTM-MIMO models.The LR-MIMO model achieved R^(2) values above 0.98 and MAPE values below 0.05 for predicting SOC in the next half hour.Compared to other models,the LR-MIMO model demonstrated excellent predictive performance,with R^(2) values above 0.95 for predicting other datasets.The prediction accuracy of the KNN-MIMO model is comparable to that of the RF-MIMO model,with R^(2) values roughly above 0.7,while the prediction performance of the LSTM-MIMO model differs significantly compared to other models due to the use of different datasets.The accuracy of the model can be improved when the data satisfies certain conditions,such as a correlation coefficient between SOC and voltage exceeding 0.9,a wide range of SOC and voltage distributions,a left-skewed kernel density curve,and a relatively uniform distribution.

关 键 词：锂离子电池 荷电状态 数据驱动 分布多样性 泛化性 

分 类 号：TM912[电气工程—电力电子与电力传动]