储能科学与技术 ›› 2020, Vol. 9 ›› Issue (6): 1969-1975.doi: 10.19799/j.cnki.2095-4239.2020.0200

• 储能测试与评价 • 上一篇    下一篇

基于LSTM-DaNN的动力电池SOC估算方法

王一全1(), 黄碧雄1(), 严晓2, 刘新田1, 王影1, 刘双宇1, 徐华源1   

  1. 1.上海工程技术大学机械与汽车工程学院，上海 201600
    2.上海玫克生储能科技有限公司，上海 201600
    3.浙江华云信息科技有限公司，浙江 杭州 310051
  • 收稿日期:2020-06-02 修回日期:2020-07-16 出版日期:2020-11-05 发布日期:2020-10-28
  • 通讯作者: 黄碧雄 E-mail:w1qnick@163.com;hbxzc@hotmail.com
  • 作者简介:王一全(1996—)，男，硕士研究生，研究方向为动力电池状态预测算法与新能源汽车大数据，E-mail:w1qnick@163.com；

SOC estimation method of power battery based on LSTM-DaNN

Yiquan WANG1(), Bixiong HUANG1(), Xiao YAN2, Xintian LIU1, Ying WANG1, Shuangyu LIU1, Huayuan XU1   

  1. 1.School of Mechanical and Automotive Engineering, Shanghai University of Engineering Science, Shanghai 201600, China
    2.Shanghai Makesens Energy Storage Technology Co. , Ltd. , Shanghai 201600, China
    3.Zhejiang Huayun Information Technology Co. Ltd. , Hangzhou 310051, Zhejiang, China
  • Received:2020-06-02 Revised:2020-07-16 Online:2020-11-05 Published:2020-10-28
  • Contact: Bixiong HUANG E-mail:w1qnick@163.com;hbxzc@hotmail.com

RichHTML

12

PDF

468

摘要：

针对动力电池荷电状态（state of charge， SOC）的估算问题，利用长短期记忆（LSTM）循环神经网络建立SOC估算模型，以实验室恒流放电数据训练模型并测试，测试最大绝对误差为2.7%。进一步以FSEC赛车电池实测数据验证，最大测试误差为3.9%。但在工程应用时，考虑到实际运行过程中的环境复杂性以及不同驾驶习惯对动力电池造成的不一致性，需要根据车辆实际行驶工况数据对其进行训练与测试，但是由于该数据中的SOC直接由BMS报文解析而来，无法确定BMS内的SOC算法是否准确，故不能直接用作训练模型时的标签，此时需计算出正确的训练标签或借助已有标签的模型，在其基础上根据实际运行数据对其模型参数进行动态调整。为解决无标签数据的训练问题，本文采取第二种方法，首次提出将迁移学习中的领域自适应网络（DaNN）与LSTM组合形成LSTM-DaNN的SOC估算算法，利用有标签数据预先训练好LSTM模型，再将其模型参数迁移至LSTM-DaNN，最后综合有标签与无标签数据一起对LSTM-DaNN模型进行训练。测试结果表明LSTM-DaNN可以在没有实际行驶工况标签（SOC）的情况下完成训练，最大测试误差为4.8%，相比模型自适应调整前误差下降了14.1%，且保证绝对误差<5%，满足实际需求。

关键词: 荷电状态（SOC）, 长短期记忆（LSTM）, 迁移学习, 领域自适应网络（DaNN）

Abstract:

In this study, a long–short-term memory (LSTM) recurrent neural network is used to establish an estimation model of the state of charge (SOC) to estimate the SOC of power batteries. The model is trained and tested with a laboratory's constant current discharge data. The maximum absolute error is 2.7%. A further verification by the FSEC racing battery measured data shows a maximum test error of 3.9%. However, considering the complexity of the environment during the actual operation and the inconsistency caused by different driving habits to the power battery in engineering applications, training and testing must be performed according to the actual driving conditions of the vehicle. The SOC in driving conditions are directly translated from the battery management system (BMS) message; hence, we cannot affirm whether the SOC algorithm in the BMS is accurate. Consequently, the SOC in the driving conditions cannot be used as a label while training the model. At this time, the correct training label must be calculated or the existing model trained by the labeled data must be used. Its model parameters must then be dynamically adjusted based on the actual operating unlabeled data. This study takes the second method to solve the training problem of unlabeled data, proposing for the first time the combination of the domain adaptive neural network (DaNN) in transfer learning with LSTM to form the SOC estimation algorithm of LSTM-DaNN using the labeled data to train the LSTM model in advance, transfer its model parameters to LSTM-DaNN, and finally train the LSTM-DaNN model by combining the labeled and unlabeled data. The result shows that LSTM-DaNN can complete the training without the label of the actual driving condition (i.e., SOC) and ensure an absolute error of 4.8%. Compared with the model before the adaptive adjustment, the error decreases by 14.1%, and the absolute error is guaranteed to be <5%, meeting actual needs.

Key words: state of charge(SOC), long short-term memory(LSTM), transfer learning, domain adaptive neural network(DaNN)

中图分类号: