大语言模型驱动的锂离子电池SOH预测方法
CSTR:
作者:
作者单位:

(1.长安大学 机器人与具身智能学院, 710018 西安; 2.西安市新能源汽车智能高效充放电与安全充电系统集成共性技术研发平台, 710048 西安)

作者简介:

郭玉龙(2000—),男,硕士研究生;雷旭(1981—),男,正高级工程师,博士生导师

通讯作者:

中图分类号:

TM912

基金项目:

国家自然科学基金 (52232015);西安市共性技术研发平台集群攻关项目(24GXPT0002);陕西省秦创原“科学家+工程师”队伍建设项目(2023KXJ-297)


SOH prediction method of lithium-ion batteries driven by large language models
Author:
Affiliation:

(1.School of Robotics and Embodied Intelligence, Chang’an University, Xi’an 710018, China; 2.Xi’an Research and Development Platform for Integrating Smart, Efficient, and Secure Charging and Discharging Systems for New Energy Vehicles, Xi’an 710048, China)

Fund Project:

undefined

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为解决现有锂离子电池健康状态(SOH)预测方法在缺乏大量标注数据时精度不足、泛化能力差且依赖复杂人工特征工程的难题,提出一种改进大语言模型(LLM)对锂离子电池SOH预测的新方法——时序大语言模型(T-LLM)。首先,通过专门设计文本化策略,将数值型的电池健康指标数据转换为LLM能够理解的语言序列。此策略保留了数值的内在结构与相对关系,使LLM能够利用其强大的自回归结构和上下文理解能力捕捉电池的复杂衰减动态。其次,在NASA及牛津大学电池数据集上的实验验证表明,T-LLM展现出卓越的性能与泛化能力。在NASA数据集上,该方法的R2稳定在96%以上,显著优于ARIMA、CNN-LSTM等模型。最后,仅需30%的训练数据便可达到超过94%的预测精度,有效克服了传统方法对大规模标注数据的依赖。该方法还成功实现了对牛津数据集的零样本跨数据集预测,在未经微调的情况下R2依然达到88%以上,解决了传统模型泛化能力差的难题。研究结果表明,T-LLM方法不仅预测精度高,并且能够快速适配不同电池类型,为实现数据高效且泛化能力强的电池SOH精准预测提供了一种全新的解决方案。

    Abstract:

    To address the problems of insufficient accuracy, poor generalization ability, and reliance on complex manual feature engineering in existing state-of-health (SOH) prediction methods of lithium-ion batteries when lacking massive labeled data, this paper proposed a new method for predicting SOH of lithium-ion batteries based on improved large language model (LLM), named T-LLM. First, this paper specifically designed a textualization strategy to convert numerical battery health indicator data into language sequences understandable by LLM. This strategy retained the internal structure and relative relationships of numerical values, enabling the LLM to utilize its powerful autoregressive structure and context understanding ability to capture the complex degradation dynamics of batteries. Second, experimental verification on the NASA and University of Oxford battery datasets demonstrates that T-LLM exhibits excellent performance and generalization ability. On the NASA dataset, the R2 of this method stabilizes above 96%, significantly outperforming models such as ARIMA and CNN-LSTM. Finally, it achieves a prediction accuracy of over 94% with only 30% of the training data, effectively overcoming the dependence of traditional methods on large-scale labeled data. This method also successfully realizes zero-shot cross-dataset prediction on the Oxford dataset; the R2 reaches over 88% without fine-tuning, solving the problem of the poor generalization ability of traditional models. The research results demonstrate that the T-LLM method not only has high prediction accuracy but also quickly adapts to different battery types, providing a novel solution for achieving precise SOH prediction of batteries with high data efficiency and strong generalization ability.

    参考文献
    相似文献
    引证文献
引用本文

郭玉龙,樊星,石玲霞,王佳龙,程吉东,雷旭.大语言模型驱动的锂离子电池SOH预测方法[J].哈尔滨工业大学学报,2026,58(8):227. DOI:10.11918/202507044

复制
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2025-07-19
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2026-08-29
  • 出版日期:
文章二维码