摘要
准确地估计电池健康状态(SOH)对锂离子电池储能系统的安全高效运行十分重要,围绕电池老化特征提取依赖特定工况和实际电池充放电过程存在随机性的问题,该文提出一种基于分段特征提取和稀疏高斯过程回归(GPR)的电池健康状态估计方法。首先,按照荷电状态对数据进行分段,针对每一段数据分别提取了30种特征,并基于充放电过程数据各训练了10种模型;其次,对部分特征计算Shapley值,解释充放电模型的差异性;然后,分析对比了五种高斯过程回归模型的核函数的预测性能;最后,与其他四种机器学习模型进行对比。测试结果表明,在高倍率工况下,基于60%~70%荷电状态的放电数据片段的模型预测效果最佳,所有测试集的平均方均根误差为0.26%。利用充电数据估计SOH时,基于马特恩核函数的GPR模型取得了最佳的预测结果,而利用放电数据估计SOH时,基于有理二次核函数的GPR模型取得了最佳的预测结果。
Accurate state of health(SOH)estimation is crucial for the safe and efficient operation of lithium-ion battery energy storage systems.Machine learning is one of the most promising methods for battery SOH estimation by extracting features from historical cycling data and aging batteries.The charging and discharging voltages of the battery,which are easily measurable parameters,are influenced by both the thermodynamic and kinetic characteristics of the battery.Thus,many studies focused on extracting battery aging features from constant current or constant voltage charge curves.Because of the uncertain need of battery output,the change of discharge voltage value is uncertain,which makes feature extraction become difficult.Besides,there are several pathways for battery aging,and models only based on the charging process may not capture all aging features.Additionally,the randomness of the battery working start points for charging and discharging further challenges for feature extraction in practical applications.To address these issues,this study proposed a novel feature extraction and sparse Gaussian process regression(GPR)method based on random segments of the charging and discharging data.This approach extracted aging features from intermittent charge and discharge data segments,which reduced the need for historical data and enhanced the generalization capability for practical applications.In this study,five different kernel functions were applied to improve the accuracy of SOH estimation for various segments of charging or discharging data in three steps.Firstly,segment of data is built.A data processing algorithm was proposed to segment the intermittent charging and discharging data under different loads.And this data processing algorithm segment data based on the range of the state of charge(SOC).Then,features of SOH estimation are extracted.A method of feature engineering was proposed to extract aging features from the data segments,which applied various algorithms of digital signal processing,audio processing,and image processing technologies.Because battery aging contains different pathways,various features are extracted,including statistical features,evolutional features,and frequency domain-related features of voltage and current curves.Statistical features include skewness,kurtosis,root mean square,root sum square,and standard deviation of the voltage and current curves.Evolutional features include dynamic time warping,correlation coefficients,and gradients of the voltage curves.Frequency domain features include mid-reference level crossings,band power,frequency bands,mean frequency,and occupied bandwidth of the voltage curves.Finally,SOH estimation is set.The sparse Gaussian process regression models with various kernels were trained to estimate battery SOH based on charging and discharging data segments.In this study,58 cells cycling data was used for testing and validation.The test results revealed that the GPR model trained on discharge data within the SOC range from 60%to 70%was the best model for SOH estimation,which achieved an average root mean square error(RMSE)of 0.09%.While the GPR model within the SOC range from 0%to 10%was the best model of the charging data,which achieved an average RMSE of 3.95%.Among the five kernels,the best kernel for the GPR models based on charging data was the Matern 5/2 kernel,whereas the rational quadratic kernel was the best kernel for the GPR models based on discharging data.Compared to traditional machine learning methods,such as linear regression,regularized linear regression(RLR),support vector machine regression,and multilayer perceptron,GPR achieved the lowest average RMSE and mean absolute percentage error.The test results demonstrated that the proposed method can accurately estimate SOH from charge and discharge data segments.
作者
江银锋
宋文祥
施语
Jiang Yinfeng;Song Wenxiang;Shi Yu(School of Mechatronic Engineering and Automation,Shanghai University,Shanghai,200444,China)
出处
《电工技术学报》
北大核心
2025年第19期6359-6377,共19页
Transactions of China Electrotechnical Society
关键词
锂离子电池
机器学习
高斯过程回归
健康状态
电池管理
Lithium-ion batteries
machine learning
Gaussian process regression
state of health
battery management
