储能科学与技术 ›› 2025, Vol. 14 ›› Issue (4): 1645-1653.doi: 10.19799/j.cnki.2095-4239.2024.0983

• 储能测试与评价 • 上一篇    下一篇

基于模态分解和深度学习的锂离子电池寿命预测

董作林1,2,3(), 宋金岩1,2,3(), 孟子迪1,2   

  1. 1.大连海洋大学信息工程学院
    2.设施渔业教育部重点实验室
    3.辽宁省海洋信息重点实验室,辽宁 大连 116023
  • 收稿日期:2024-10-21 修回日期:2024-11-23 出版日期:2025-04-28 发布日期:2025-05-20
  • 通讯作者: 宋金岩 E-mail:2900325200@qq.com;songjinyan@dlou.edu.cn
  • 作者简介:董作林(2000—),男,硕士研究生,研究方向为电池寿命预测,E-mail:2900325200@qq.com
  • 基金资助:
    辽宁省自然科学基金指导计划(20180550573);大连海洋大学校级教改项目(32),辽宁省教育厅项目(JL201918)

Lithium-ion battery life prediction based on mode decomposition and deep learning

Zuolin DONG1,2,3(), Jinyan SONG1,2,3(), Zidi MENG1,2   

  1. 1.College of Information Engineering, Dalian Ocean University
    2.Key Laboratory of Facilities Aquaculture, Ministry of Education
    3.Key Laboratory of Marine Information of Liaoning Province, Dalian 116023, Liaoning, China
  • Received:2024-10-21 Revised:2024-11-23 Online:2025-04-28 Published:2025-05-20
  • Contact: Jinyan SONG E-mail:2900325200@qq.com;songjinyan@dlou.edu.cn

摘要:

随着新能源汽车数量的快速增长,精准预测锂离子电池的剩余使用寿命(remaining useful life, RUL)对新能源汽车产业的持续发展起到了至关重要的作用。本工作提出了一种基于集成经验模态分解(ensemble empirical mode decomposition, EEMD)和深度学习的创新方法,旨在提升锂离子电池RUL的预测精度。首先,通过EEMD对电池容量数据进行多尺度分解,得到电池容量数据的全局退化趋势和局部随机波动分量。为了减轻噪声对模型预测精确度的干扰,引入去噪自编码器(denoising autoencoder, DAE)对随机波动分量进行降噪处理。随后,分别使用长短期记忆(long short-term memory, LSTM)网络和自注意力模型(Transformer)对全局退化趋势和降噪后的随机波动分量进行建模。最后为进一步提取各模态分量中存在的信息,采用随机森林(random forest, RF)算法计算各分量的重要性权重,根据得到的权重值对预测结果加权重构。本工作在美国国家航空航天局(National Aeronautics and Space Administration, NASA)公开的电池数据集上分别使用40%和60%的历史数据进行实验,结果表明所提出的方法在精度和有效性方面均优于现有方法,验证了其在锂离子电池RUL预测中的应用潜力。

关键词: 锂离子电池, 寿命预测, 深度学习, 长短期记忆, 随机森林

Abstract:

With the rapid growth in the electric vehicle (EV) adoption, accurately predicting the remaining useful life (RUL) of lithium-ion batteries has become critical for the sustained development of the EV industry. This paper proposes an innovative approach that integrates ensemble mode decomposition (EEMD) and deep learning to improve RUL prediction accuracy for lithium-ion batteries. The proposed method begins with EEMD, which performs multiscale decomposition of battery capacity data. This process separates the global degradation trend from local random fluctuation components. To mitigate the impact of noise on model prediction accuracy, a denoising autoencoder (DAE) is introduced to remove noise from the random fluctuation components. Subsequently, long short-term memory (LSTM) networks and the transformer model are applied to model the global degradation trend and the denoised random fluctuations, respectively. To further refine predictions, a random forest (RF) algorithm calculates the importance weights of each mode component, enabling a weighted reconstruction of the prediction results. Experiments were conducted on a public battery dataset provided by the National Aeronautics and Space Administration (NASA), leveraging 40% and 60% of the historical battery data. The results demonstrate that the proposed method outperforms existing approaches in both accuracy and effectiveness, validating its potential for application in lithium-ion battery RUL prediction.

Key words: lithium-ion battery, life prediction, deep learning, long short-term memory, random forest

中图分类号: