储能科学与技术 ›› 2022, Vol. 11 ›› Issue (1): 228-239.doi: 10.19799/j.cnki.2095-4239.2021.0373

• 储能测试与评价 • 上一篇    下一篇

基于滑动窗口和LSTM神经网络的锂离子电池建模方法

张少凤1, 张清勇1(), 杨叶森2, 苏义鑫1, 熊斌宇1   

  1. 1.武汉理工大学自动化学院,湖北 武汉 430072
    2.新加坡南洋理工大学电气与电子工程学院,新加坡 639798
  • 收稿日期:2021-07-26 修回日期:2021-08-19 出版日期:2022-01-05 发布日期:2022-01-10
  • 通讯作者: 张清勇 E-mail:qyzhang@whut.edu.cn
  • 作者简介:张少凤(1998—),女,硕士研究生,主要研究方向为机器学习,E-mail:zhangshaofeng@wuht.?edu.?cn;
  • 基金资助:
    国家自然科学基金项目(2020619)

Lithium-ion battery model based on sliding window and long short term memory neural network

Shaofeng ZHANG1, Qingyong ZHANG1(), Yesen YANG2, Yixin SU1, Binyu XIONG1   

  1. 1.School of Automation, Wuhan University of Technology, Wuhan 430072, Hubei, China
    2.School of Electrical and Electronic Engineering (EEE), Nanyang Technological University, Singapore 639798
  • Received:2021-07-26 Revised:2021-08-19 Online:2022-01-05 Published:2022-01-10
  • Contact: Qingyong ZHANG E-mail:qyzhang@whut.edu.cn

摘要:

为提高锂离子电池在复杂工况下的预测能力和建模精度,提出一种基于滑动窗口和长短时记忆(long short term memory,LSTM)神经网络的锂离子电池建模方法。首先建立了基于神经网络的锂离子电池模型,确定了神经网络的基本结构,通过LSTM层、向量拼接层和全连接层分别实现了时序特征提取、特征融合和回归预测。然后提出了滑动窗口的输入向量处理方法,滑动窗口每次向前推进一个时间点,通过限制时间窗口内所能处理的最大信元数对数据量进行限制,为多个LSTM层的并行计算和深隐层的拼接层和全连接层预留了计算量的裕度,实现了对模型中循环网络层深度的优化选择。为解决模型在多工况下运行的泛化问题,提出使用离线数据集的预训练和在线数据的参数修正的训练方法,通过大量离线数据集的反复训练,使模型学习电池的共性部分;再使用部分在线数据,对网络参数进行调整,将其应用于预测中。最后使用恒流/恒压、随机电流脉冲、大功率脉冲等多个工况的数据分别进行测试。结果表明,基于长短时记忆神经网络的建模方法能够准确预测电池输出电压和荷电状态。

关键词: 锂离子电池, 模型, 神经网络, 长短时间记忆, 多时序特征提取, 滑动窗口

Abstract:

This study proposes a lithium-ion battery model based on the sliding window and long short-term memory (LSTM) neural network to improve the model accuracy under complex working conditions. First, a lithium-ion battery model based on the LSTM neural network is established. Next, the basic structure of the neural network is determined. The time series feature extraction, feature fusion, and regression prediction are realized by combining the LSTM, vector splicer, and full connection layers. A sliding window input vector processing method is then proposed. The sliding window is advanced one time point at a time, and the data volume is limited by restricting the maximum number of letter elements within the time window. A computational margin is reserved for the parallel computation of the multiple LSTM layers and the deep hidden layers of the splicing and fully connected layers. Subsequently, the optimal selection of the depth of the recurrent network layer in the model is achieved. A training method using offline data set pre-training and online data parameter modification is proposed to solve the generalization problem under various complex working conditions. The model learns the common parts of the battery through the repetitive training of a large number of offline data sets. The network parameters are adjusted and used in the prediction by using a part of the online data. Finally, the datasets of the constant current/constant voltage, random current pulse, high-power pulse, and other working condition test profiles are applied for validation. The results show that the proposed modeling method can accurately predict the battery's output voltage and state of charge.

Key words: lithium-ion battery, battery model, neural network, long short term memory, feature extraction, sliding window

中图分类号: