汽车工程 ›› 2022, Vol. 44 ›› Issue (6): 868-878.doi: 10.19562/j.chinasae.qcgc.2022.06.008

所属专题: 新能源汽车技术-动力电池&燃料电池2022年

• • 上一篇    下一篇

动态条件下基于深度学习的锂电池容量估计

毕贵红,谢旭,蔡子龙(),骆钊,陈臣鹏,赵鑫   

  1. 昆明理工大学电力工程学院,昆明  650500
  • 收稿日期:2021-11-24 修回日期:2021-12-30 出版日期:2022-06-25 发布日期:2022-06-28
  • 通讯作者: 蔡子龙 E-mail:1250582439@qq.com
  • 基金资助:
    国家自然科学基金(51907084)

Capacity Estimation of Lithium-ion Battery Based on Deep Learning Under Dynamic Conditions

Guihong Bi,Xu Xie,Zilong Cai(),Zhao Luo,Chenpeng Chen,Xin Zhao   

  1. School of Electric Power Engineering,Kunming University of Science and Technology,Kunming  650500
  • Received:2021-11-24 Revised:2021-12-30 Online:2022-06-25 Published:2022-06-28
  • Contact: Zilong Cai E-mail:1250582439@qq.com

摘要:

锂离子电池在老化过程中,其内部呈现非线性的复杂变化,因此直接使用动态条件下的锂离子电池运行时段的数据(电流、电压和温度)进行电池健康状态的实时估计是一个具有挑战性的问题。本文中选取锂离子电池随机充放电数据,对动态数据的部分片段进行时频特征提取,组成时频特征矩阵作为输入,构建级联式卷积神经网络和门控循环单元容量估计模型,对输入数据进行内在特征提取,并进一步挖掘各时间序列中的相关特征,实现锂离子电池动态条件下的容量估计。利用美国航空航天局锂离子电池随机使用数据集进行实验验证的结果表明,该方法能在仅已知电池的额定容量的情况下,准确完成锂离子电池容量估计。最后,本文还分析了模型超参数设置、原始数据时序长度、网络输入和模型结构对容量估计精度的影响。

关键词: 锂离子电池, 容量估计, 卷积神经网络, 门控循环单元循环神经网络, 深度学习

Abstract:

In the aging course of lithium-ion battery, nonlinear complicated changes occur inside the battery, therefore directly using the operation data of lithium-ion battery such as current, voltage and temperature in certain time sections to conduct real-time estimation on the battery state of health under dynamic condition is a challenging issue. In this paper, the random charging and discharging data of lithium-ion battery are selected, the time and frequency features of some segments of dynamic data are extracted to compose time and frequency feature matrices as input, a cascaded convolutional neural network and a gated recurrent unit capacity estimation model are constructed to extract the intrinsic features of input data, and the related features of each time sequence are further explored to fulfill the estimation of battery capacity under dynamic condition. The results of experimental verification utilizing NASA's lithium-ion battery random use data set show that the method adopted can accurately estimate the capacity of lithium-ion battery under the condition of only the nominal capacity of battery is known. Finally, the effects of the setting of model’s hyper-parameters, the time-sequence length of raw data, network input and model structure on the accuracy of battery capacity estimation are analyzed.

Key words: lithium-ion battery, capacity estimation, CNN, GRU-RNN, deep learning