储能科学与技术 ›› 2025, Vol. 14 ›› Issue (7): 2865-2874.doi: 10.19799/j.cnki.2095-4239.2025.0062

• 第十三届储能国际峰会暨展览会专辑 • 上一篇    下一篇

基于力-电-温度信号和CNN-BiLSTM模型的磷酸铁锂电池SOC估计

马昊远1(), 吴焱1, 王通1, 胡锦洋1, 李佳2, 黄钰期1()   

  1. 1.浙江大学能源工程学院，动力机械及车辆工程研究所，浙江 杭州 310027
    2.威睿电动汽车技术(宁波)有限公司，浙江 宁波 315336
  • 收稿日期:2025-01-18 修回日期:2025-02-17 出版日期:2025-07-28 发布日期:2025-07-11
  • 通讯作者: 黄钰期 E-mail:3163274617@qq.com;huangyuqi@zju.edu.cn
  • 作者简介:马昊远（2002—），男，硕士研究生，研究方向为锂电池状态估计，E-mail：3163274617@qq.com；
  • 基金资助:
    浙江省重点研发计划“尖兵”项目(2024C01061);浙江省自然科学基金资助项目(LD25E070002)

State of charge estimation of lithium iron phosphate batteries based on force-electric-temperature signals and a CNN-BiLSTM model

Haoyuan MA1(), Yan WU1, Tong WANG1, Jinyang HU1, Jia LI2, Yuqi HUANG1()   

  1. 1.School of Energy Engineering, Zhejiang University, Institute of Power Mechinery and Vehicular Engineering, Hangzhou 310027, Zhejiang, China
    2.VEERY EV Tech (Ningbo) Co. , Ltd. , Ningbo 315336, Zhejiang, China
  • Received:2025-01-18 Revised:2025-02-17 Online:2025-07-28 Published:2025-07-11
  • Contact: Yuqi HUANG E-mail:3163274617@qq.com;huangyuqi@zju.edu.cn

RichHTML

0

PDF

119

摘要：

锂电池的荷电状态(state of charge，SOC)是电池管理系统的重要参数，但其与电池内部复杂的电化学特性高度关联，无法直接测量。近年来，基于数据驱动的方法在SOC估计领域展现了极大的潜力，然而其对输入数据的精确性有较高要求。磷酸铁锂电池因存在电压平台问题，其电压波动和噪声会严重影响SOC估计的精度，本文针对这一问题，通过实验和数据驱动结合的方法，引入电池膨胀力作为新的输入维度，融合电池的电化学特性与机械特性，有效补偿了电压平台问题对SOC估计结果的影响。本研究在4种环境温度和2种动态电流测试工况下进行了实验，利用所得数据对神经网络模型进行训练和测试，以评估SOC估计精度并验证本方法的可行性和普适性。此外，本文还提出了一种基于卷积神经网络(convolutional neural network，CNN)和双向长短期记忆网络(bidirectional long short-term memory，Bi-LSTM)的混合模型，兼顾序列数据的局部模式与长期依赖关系，进一步提升SOC估计的可靠性。结果表明，本文提出的方法可以显著提高磷酸铁锂电池SOC估计精度，相比未引入膨胀力信号，均方根误差(root-mean-square error，RMSE)平均下降了43.82%。同时，CNN-BiLSTM模型相比其他常规神经网络模型，RMSE最多降低了53.88%。本研究为高精度SOC估计提供了新的思路，对提升电池管理系统的性能具有重要意义。

关键词: 磷酸铁锂电池, 荷电状态估计, 膨胀力, 数据驱动, 双向长短期记忆模型

Abstract:

The state of charge (SOC) of lithium batteries is a critical parameter for battery management systems. However, SOC cannot be measured directly because it is strongly coupled to the intricate electrochemical characteristics of the battery. Although data-driven methods recently demonstrated significant potential in SOC estimation, their accuracy depends greatly on the precision of input data. SOC estimation of lithium iron phosphate (LiFePO4) batteries is challenging because the battery exhibits voltage plateau characteristics: voltage fluctuations and noise substantially degrade the estimation reliability. To overcome this challenge, this study proposes a hybrid experimental and data-driven approach that incorporates battery expansion force as a novel input dimension, thereby synergizing the electrochemical and mechanical properties of the battery to mitigate the impact of voltage plateau on SOC estimation. Experiments were conducted at four environmental temperatures and under two dynamic current test conditions. The acquired data were used to train and validate neural network models for evaluating the SOC estimation accuracy and to verify the feasibility and robustness of the proposed method. Furthermore, a hybrid model combining a convolutional neural network (CNN) and a bidirectional long short-term memory (Bi-LSTM) network was developed to simultaneously capture local temporal patterns and long-term dependencies in sequential data, further enhancing the SOC estimation reliability. The results indicated that the proposed method remarkably improved the SOC estimation accuracy for LiFePO4 batteries, achieving an average reduction of 43.82% in the root-mean-square error (RMSE) compared to methods that did not incorporate expansion force signals. Moreover, the CNN-BiLSTM model outperformed conventional neural network models, achieving a maximum RMSE reduction of 53.88%. Thus, this study provides a novel perspective for high-precision SOC estimation and offers valuable knowledge for advancing the performance of battery management systems.

Key words: iron phosphate lithium battery, state of charge estimation, expansion force, data-driven, bi-directional long short-term memory model

中图分类号: