储能科学与技术 ›› 2023, Vol. 12 ›› Issue (2): 560-569.doi: 10.19799/j.cnki.2095-4239.2022.0611

• 储能测试与评价 • 上一篇    下一篇

基于间接健康指标的高斯过程回归对锂电池SOH预测

王瑞洁1,2(), 惠周利1(), 杨明1,2()   

  1. 1.中北大学数学学院
    2.信息探测与处理山西省重点实验室,山西 太原 030051
  • 收稿日期:2022-10-21 修回日期:2022-11-07 出版日期:2023-02-05 发布日期:2022-11-25
  • 通讯作者: 惠周利,杨明 E-mail:2286944596@qq.com;13994208298@139.com;hgsnje@163.com
  • 作者简介:王瑞洁(1999—),女,硕士研究生,研究方向为数据分析, E-mail:2286944596@qq.com
  • 基金资助:
    山西省青年科技研究基金项目(201901D211275);山西省基础研究计划资助项目(202103021224190);国家自然科学基金项目(61801437)

Gaussian process regression based on indirect health indicators for SOH estimation of lithium battery

Ruijie WANG1,2(), Zhouli HUI1(), Ming YANG1,2()   

  1. 1.School of Mathematics, North University of China
    2.Shanxi Key Laboratory of Signal Capturing & Processing, North University of China, Taiyuan 030051, Shanxi, China
  • Received:2022-10-21 Revised:2022-11-07 Online:2023-02-05 Published:2022-11-25
  • Contact: Zhouli HUI, Ming YANG E-mail:2286944596@qq.com;13994208298@139.com;hgsnje@163.com

摘要:

锂电池性能会随使用时间增加而逐步退化,若更换不及时,可能造成爆炸等严重事故。快速准确预测电池健康状态(state of health,SOH),对于锂电池系统管理和维护以及安全使用至关重要。本工作提出一种基于间接健康指标(health indicators,HIs)和高斯过程回归(Gaussian process regression,GPR)相结合预测锂电池SOH的机器学习模型。首先,通过分析锂电池放电过程,提取若干易于获得且适合动态操作的直接外部特征作为间接健康指标,并计算它们和SOH的相关性,最终筛选出平均放电电压、等压降放电时间、最高放电温度和平台期放电电压初始骤降值作为健康指标;其次,以上述健康指标作为输入特征,利用GPR算法建立锂电池退化模型,对NASA锂电池数据集进行预测,平均绝对误差(mean absolute error,MAE)不超过2%,均方根误差(root mean square error,RSME)控制在4%之内;最后,将本工作模型与其他常用机器学习模型进行比较,再将模型带入不同实验条件的电池中进行泛化性能分析,最大预测误差控制在6%之内,实验结果表明,本工作提出的间接健康指标和GPR模型具有相对较高的预测精度和优秀的泛化能力。

关键词: 健康指标, 健康状态, 高斯过程回归, 支持向量机回归

Abstract:

The performance of a lithium battery degrades gradually with increasing use time. If the replacement is not completed on time, serious accidents such as explosions may occur. Rapid and accurate prediction of battery state of health (SOH) is critical for lithium battery system management, maintenance, and safe use. In this paper, a machine learning model based on indirect His (health indicators) and GPR (Gaussian process regression) is proposed to predict the SOH of lithium batteries. First, the analysis of the lithium battery discharge process extracts some easily available and suitable for the direct external features of dynamic operations as indirect His, and their correlation with SOH, eventually selecting average discharge voltage, such as pressure drop discharge time, maximum discharge temperature, and platform stage discharge voltage initial plummet in value as the health index. Second, using the above mentioned His as input features, the GPR algorithm is used to establish a lithium battery degradation model, and the MAE (mean absolute error) is less than 2% for the prediction of NASA lithium battery datasets, while the RMSE is kept within 4%. Finally, the model is compared to other commonly used machine learning models, and then into multiple experimental conditions of battery model generalization performance analysis, control about 6% of the maximum error of the prediction. The experimental results show that the proposed indirect His and GPR have relatively higher prediction precision and good generalization ability.

Key words: health indicators, state of health, Gaussian process regression, support vector regression

中图分类号: