储能科学与技术 ›› 2020, Vol. 9 ›› Issue (5): 1548-1557.doi: 10.19799/j.cnki.2095-4239.2020.0132

• 储能系统与工程 • 上一篇    下一篇

基于特征组合堆叠融合集成学习的锂离子动力电池SOC估算

何瑛(), 钟根鹏, 陈翌()   

  1. 同济大学汽车学院，上海 201804
  • 收稿日期:2020-04-02 修回日期:2020-04-25 出版日期:2020-09-05 发布日期:2020-09-08
  • 通讯作者: 陈翌 E-mail:03101@tongji.edu.cn;chenyi63@tongji.edu.cn
  • 作者简介:何瑛（1979—），女，硕士，研究方向为车辆工程（汽车市场营销），E-mail：03101@tongji.edu.cn；

Research on SOC estimation of lithium-ion power battery based on feature combination and stacking fusion ensemble Learning

Ying HE(), Genpeng ZHONG, Yi CHEN()   

  1. School of Automotive Studies, Tongji University, Shanghai 201804, China
  • Received:2020-04-02 Revised:2020-04-25 Online:2020-09-05 Published:2020-09-08
  • Contact: Yi CHEN E-mail:03101@tongji.edu.cn;chenyi63@tongji.edu.cn

RichHTML

21

PDF

883

摘要：

关于电池SOC估算的研究多采用单体电池在理想实验条件下的充放电数据进行，这未必能适应真实复杂多变的行驶工况。针对此问题，本工作依托新能源汽车国家大数据联盟，利用数据驱动的方法，构建了瞬态、驾驶行为、衰老、时变、行驶里程和动态特征来从多个角度反映动力电池的工作状况。采用基于特征组合的堆叠融合集成学习方法，对实际复杂多变工况下的动力电池放电过程进行了SOC估算，并构建了1个基于瞬态特征的Xgboost Model 0参照组模型，5个基于特征组合的Xgboost Model 1、2、3、5、6模型以及1个线性模型Linear Model 4进行对比分析。结果表明：对比Model 1～6与参照组Model 0，基于特征组合的堆叠融合模型(堆叠模型)平均绝对误差(MAE)、均方误差(MSE)最小，分别为0.39和0.32，相较于参照组模型分别降低了38%和46%；同时其决定系数R2最高，达到0.9995，相较于参照组模型提高了2%。堆叠模型的泛化能力也表现良好，其准确性平均值和标准差分别达到98.89%和0.03%。对比Model 5、6与参照组Model 0，可知随着特征维度的增加，模型的MAE、MSE会减小，R2会增加，模型的性能变好。本研究有助于推动数据驱动方法在动力电池SOC估算的应用，对实际行驶的电动汽车SOC估算有一定指导和参考意义。

关键词: 动力电池, Xgboost, SOC估算, 集成学习, 特征组合

Abstract:

Studies on the estimation of the battery power state of charge (SOC) are primarily based on the charge and discharge data of a single cell under ideal experimental conditions, which may not correspond to the actual complex and variable driving conditions. In response to this problem, relying on the National Big Data Alliance of New Energy Vehicles and using data-driven methods, the transient, dynamic, and driving behavior and aging, dynamic, time-varying, and mileage features were derived. Using the method of stacking fusion and integration learning based on feature combination, an SOC estimation of the battery power discharge process under actual complex and variable working conditions was conducted. A reference XGBoost Model 0 based on transient features, five XGBoost Models (1, 2, 3, 5, and 6) based on feature combination, and a Linear Model 4 were constructed for comparison and analysis. Comparing Models 1-6 with the reference Model 0, the mean absolute error (MAE) and mean square error (MSE) of the stacked fusion model (stacked model) based on feature combination were the smallest, 0.39 and 0.32, respectively, which compared to Model 0, decreased by 38% and 46%, respectively; meanwhile, the coefficient of determination was the highest, reaching 0.9995, which was an increase of 2% compared to Model 0. The generalization ability of the stacked model also performed well, and the average and standard deviation of its accuracy reached 98.89% and 0.03%, respectively. Comparing Models 5 and 6 and the reference Model 0, it can be seen that, as the feature dimension increases, the MAE and MSE of the model decreases, the coefficient of determination increases, and the performance of the model improves. This study helps promote the application of data-driven methods in the estimation of the SOC of power batteries and provides guidance and a reference for the estimation of the SOC of actual electric vehicles.

Key words: power battery, Xgboost, SOC estimation, ensemble learning, feature combination

中图分类号: