电子学报 ›› 2023, Vol. 51 ›› Issue (1): 50-56.DOI: 10.12263/DZXB.20211640

• 学术论文 • 上一篇    下一篇

基于Informer的电池荷电状态估算及其稀疏优化方法

何滢婕, 刘月峰, 边浩东, 郭威, 张小燕   

  1. 内蒙古科技大学信息工程学院,内蒙古 包头 014010
  • 收稿日期:2021-12-09 修回日期:2022-08-15 出版日期:2023-01-25
    • 通讯作者:
    • 刘月峰
    • 作者简介:
    • 何滢婕 女,1998年出生,山西太原人.现为内蒙古科技大学信息工程学院计算机系硕士研究生.从事电动汽车电池状态估计及深度学习方面的研究.E-mail: 1353517169@qq.com
      刘月峰(通讯作者) 男,1977年出生,博士.现为内蒙古科技大学副教授.主要研究方向为机器学习、知识图谱、图像处理、大数据分析与应用.E-mail: liuyuefeng01035@163.com
    • 基金资助:
    • 内蒙古纪检监察大数据实验室开放课题基金 (IMDBD20200220)

State-of-Charge Estimation of Lithium-ion Battery Based on Informer and Its Sparse Optimization Method

HE Ying-jie, LIU Yue-feng, BIAN Hao-dong, GUO Wei, ZHANG Xiao-yan   

  1. School of Information Engineering, Inner Mongolia University of Science and Technology, Baotou, Inner Mongolia 014010, China
  • Received:2021-12-09 Revised:2022-08-15 Online:2023-01-25 Published:2023-02-23
    • Corresponding author:
    • LIU Yue-feng
    • Supported by:
    • Inner Mongolia Discipline Inspection and Supervision Big Data Laboratory Open Project Fund (IMDBD20200220)

摘要:

准确估计电池荷电状态(State Of Charge,SOC)是延长电动汽车电池使用寿命,确保电动汽车行驶安全的重要基础.传统的深度学习估计方法存在并行化计算效率不高、训练时间长的问题.为此,利用基于自注意力机制的Informer模型来估计电池SOC.其降低了传统自注意力机制的时间复杂度、提高了硬件使用率、降低了训练时长,与其他深度学习方法相比估计更准确.然而Informer模型仍然存在体量大及参数冗余的问题,故提出稀疏优化方法.利用基于彩票假设的幅值迭代剪枝方法对Informer进行稀疏化处理,突出主导注意力特征,实现了在降低参数冗余的同时提升模型估计精度.在室温下,提出的稀疏化Informer模型估计电池SOC的平均绝对误差和均方根误差分别达到0.285 8%和0.383 0%,相比于Informer模型在平均绝对误差指标上估计精度提升了25%.并验证了其具备估计不同类型锂电池SOC的泛化能力.与循环神经网络、卷积神经网络这类传统的深度学习模型相比,本模型进行电池SOC估计时训练速度更快,估计准确性和稳定性更高.

关键词: 荷电状态, 锂离子电池, 深度学习, 编解码结构, 自注意力机制

Abstract:

Accurate estimation of the state of charge (SOC) is an important basis for extending the battery life and ensuring the safety of electric vehicles. Traditional deep learning estimation methods suffer from inefficient parallelization and lengthy training time. To this end, the Informer model based on the self-attention mechanism is used to estimate the battery SOC, which reduces the time complexity of the traditional self-attention and training time, increases the hardware usage. It is more accurate than other deep learning methods. However, the Informer model still has the problems of large and redundant parameters, therefore a sparse optimization method is proposed. The iterative magnitude pruning method based on the lottery ticket hypothesis is used to sparse the Informer, highlighting the dominant attention feature and improving model estimation accuracy while reducing parameter redundancy. The proposed sparse-Informer's root mean square error and mean absolute error are 0.383 0% and 0.285 8% at room temperature, respectively. The mean absolute error is decreased by 25% compared to Informer model. Additionally, the generalization ability to estimate the SOC across other lithium-ion battery types is confirmed. When performing battery SOC estimation, this model outperforms other established deep learning models in terms of training speed, estimation accuracy and stability.

Key words: state of charge, lithium-ion battery, deep learning, encoder and decoder, self-attention mechanism

中图分类号: