储能科学与技术 ›› 2023, Vol. 12 ›› Issue (11): 3508-3518.doi: 10.19799/j.cnki.2095-4239.2023.0458

• 储能测试与评价 • 上一篇    下一篇

基于容量增量曲线与GWO-GPR的锂离子电池SOH估计

王琛(), 闵永军()   

  1. 南京林业大学汽车与交通工程学院,江苏 南京 210037
  • 收稿日期:2023-07-03 修回日期:2023-08-13 出版日期:2023-11-05 发布日期:2023-11-16
  • 通讯作者: 闵永军 E-mail:wangchen12090598@126.com;yjmin@njfu.edu.cn
  • 作者简介:王琛(1998—),男,硕士研究生,研究方向为基于数据驱动的新能源汽车电池状态监测,E-mail:wangchen12090598@126.com

SOH estimation of lithium-ion batteries based on capacity increment curve and GWO-GPR

Chen WANG(), Yongjun MIN()   

  1. College of Automobile and Traffic Engineering, Nanjing Forestry University, Nanjing 210037, Jiangsu, China
  • Received:2023-07-03 Revised:2023-08-13 Online:2023-11-05 Published:2023-11-16
  • Contact: Yongjun MIN E-mail:wangchen12090598@126.com;yjmin@njfu.edu.cn

摘要:

电池健康状态(state of health, SOH)的准确估计是电池管理系统的关键技术之一,对保障电动汽车安全、可靠运行至关重要。针对当前高斯过程回归(gaussian process regression,GPR)中单一核函数泛化性能不足,超参数选取易陷入局部最优导致SOH估计精度较低的问题,提出一种灰狼优化算法(grey wolf optimization,GWO)和组合核函数改进GPR的SOH估计方法。首先,基于容量增量分析法提取用于表征电池老化的特征,对电池恒流充电的容量-电压曲线插值并以差分法计算容量增量(increment capacity,IC)曲线,应用Savitzky-Golay滤波平滑处理,提取峰值高度、峰值电压及峰面积作为健康特征;其次,引入多维尺度变换(multidimensional scaling, MDS)消除特征冗余性同时降低模型计算复杂度,利用Pearson系数验证所提健康特征与SOH的相关性;然后,结合SOH退化轨迹的非线性和电池容量再生的准周期性特点,将神经网络核函数与周期核函数组合作为GPR的协方差核函数,以GWO对组合核函数超参数的初值进行优化;最后,基于NASA电池数据集将所提方法与SVR、ELM、GPR模型作对比,检验GWO-GPR模型的准确性,估计结果的最大均方根误差(root mean square error,RMSE)和平均绝对误差(mean absolute error,MAE)分别为1.03%和0.5%,以第60、80、100个循环为估计起始点,验证模型的鲁棒性,结果显示最大RMSE控制在1.03%以内。

关键词: 锂离子电池, 健康状态, 容量增量曲线, 高斯过程回归, 灰狼优化算法

Abstract:

Accurate estimation of the battery state of health (SOH) is a critical technology in battery management systems, which is crucial for ensuring the safe and reliable operation of electric vehicles. To solve the problem of low SOH estimation accuracy due to insufficient generalization performance of a single kernel function in Gaussian process regression (GPR) and the tendency of hyperparameter selection to fall into local optimality, an SOH estimation method based on the grey wolf optimization algorithm (GWO) and a combined kernel function was proposed. First, the characteristics of battery aging were extracted using incremental capacity analysis (ICA) method. The capacity-voltage curve of constant-current charging of the battery was interpolated and the increment capacity (IC) curve was calculated using the difference method. The IC curve was smoothed using Savitzky-Golay filtering, and the peak height, voltage, and area were extracted as health features. Second, multidimensional scaling (MDS) was presented to eliminate feature redundancy and reduce the computational complexity of the model. The Pearson coefficient was used to verify the correlation between the proposed health features and SOH. Then, considering the nonlinearity of the SOH degradation trajectory and the quasi-periodicity of battery capacity regeneration, the combination of the neural network kernel function and periodic kernel function was used as the covariance kernel function of GPR, and the initial hyperparameters of the combined kernel function were optimized by the GWO method. Finally, the proposed method was compared with SVR, ELM, and GPR models based on the NASA battery data set to verify the accuracy of the GWO-GPR model. The 60th, 80th, and 100th cycles were used as estimation starting points to verify the robustness of the model.

Key words: lithium-ion battery, state of health, increment capacity curve, gaussian process regression, grey wolf optimization algorithm

中图分类号: