储能科学与技术 ›› 2021, Vol. 10 ›› Issue (3): 1163-1176.doi: 10.19799/j.cnki.2095-4239.2021.0051

• 储能测试与评价 • 上一篇    下一篇

基于Bi-LSTM/Bi-GRU循环神经网络的锂电池SOC估计

朱元富1(), 贺文武1(), 李建兴1, 李有财2, 李培强1   

  1. 1.福建工程学院,福建 福州 350018
    2.福建星云电子股份有限公司,福建 福州 350015
  • 收稿日期:2021-02-02 修回日期:2021-02-06 出版日期:2021-05-05 发布日期:2021-04-30
  • 通讯作者: 贺文武 E-mail:yuanfuzhu0130@gmail.com;hwwhbb@163.com
  • 作者简介:朱元富(1997—),男,硕士研究生,研究方向为储能技术与应用,E-mail:yuanfuzhu0130@gmail.com
  • 基金资助:
    国家重点研发计划项目(2018YFB0905304);福建省区域发展项目(2018H4005);福建省自然科学基金项目(2020J01891)

SOC estimation for Li-ion batteries based on Bi-LSTM and Bi-GRU

Yuanfu ZHU1(), Wenwu HE1(), Jianxing LI1, Youcai LI2, Peiqiang LI1   

  1. 1.Fujian University of Technology, Fuzhou 350018
    2.Fujian Nebula Electronics Co. Ltd. , Fuzhou 350015, Fujian, China
  • Received:2021-02-02 Revised:2021-02-06 Online:2021-05-05 Published:2021-04-30
  • Contact: Wenwu HE E-mail:yuanfuzhu0130@gmail.com;hwwhbb@163.com

摘要:

锂电池的荷电状态(state-of-charge, SOC)涉及的物理特性或电化学特性高度复杂,其值一般难以直接测量,基于深度神经网络等新方法的SOC估计近期为相关研究者所关注。为进一步提升SOC估计性能,有效捕获锂电池SOC的动态物理特性,缓解深度神经网络模型容易发生的梯度消失与梯度爆炸等问题,本文引入双向学习策略,基于双向长短期记忆循环神经网络(bidirectional long short-term memory, Bi-LSTM)以及双向门控循环单元网络(bidirectional gated recurrent unit, Bi-GRU)估计锂电池的SOC取值。双向循环神经网络SOC估计模型由输入层、隐藏层和输出层组成。输入层输入电池电压、电流与温度序列;隐藏层在正向LSTM/GRU 层的基础上增加反向LSTM/GRU 层,引入逆序信息,基于输入序列上下文所含信息整体上学习、表征电池特性序列与SOC序列之间的内在关联;输出层输出模型的估计值。所拟模型使用Python 语言结合TensorFlow 后端在Keras 框架中实现,并基于基准数据集在3种温度条件下结合多种工况进行性能分析。结果表明,双向学习策略能有效提升锂电池SOC的估计性能,较之单向学习模型具有更高的估计精度与鲁棒性。与构造电池等效模型等方法的思路不同,所拟方法基于数据驱动学习锂电池SOC的非线性特性,将易于观测的锂电池特性序列数据映射为待估计的SOC取值,为锂电池SOC估计提供了可能的新思路。

关键词: 锂电池, SOC估计, Bi-LSTM网络, Bi-GRU网络

Abstract:

The direct measurement of the state of charge (SOC) for lithium-ion batteries, whose physical characteristics or the electrochemical characteristics are highly complex, presents difficulty. Methods, such as deep neural networks, have recently received wide attention and have been exploited to estimate battery SOC. To improve the estimation performance of SOC, capture accurately the interior dynamic characteristics of the batteries, and alleviate the vanishing and exploding gradient problem in neural networks, in this paper, we introduce a bidirectional learning strategy and develop two specific methods, i.e., the estimation of the SOC with the bidirectional long short-term memory (Bi-LSTM) networks and the bidirectional gated recurrent unit (Bi-GRU) networks. Both the proposed bidirectional networks consist of three parts: the input, hidden, and output layers. The input layer accepts the sequence of measurements, such as the voltages, currents, and temperatures. The hidden layer consists of two LSTM/GRU sub-layers; one processes the forward information, and the other processes the backward information to learn the input-output sequence mapping form the context of the sequence. The output layer outputs the learned SOC. Python, TensorFlow, and Keras are used to develop the concrete models, and the resultant ones are tested and analyzed on the benchmark data sets, where three temperature and nine working conditions are considered. The results show that the proposed bidirectional learning methods can outperform their competitors, with higher accuracy and better robustness. Being different from the idea of constructing battery equivalent models, the proposed method is data-driven and intends to estimate SOC with easily obtainable measurements, such as the voltages, currents, and temperatures, and provides a potential method for SOC estimation.

Key words: lithium-ion battery, SOC estimation, Bi-LSTM network, Bi-GRU network

中图分类号: