Scholarlink

<p>As electric transportation and energy storage systems continue to scale up, lithium-ion battery health management imposes increasingly stringent requirements on cross-temperature State of Health (SOH) Estimation. However, temperature variations alter charging kinetics and the statistical scale of health features, inducing a systematic bias in the feature-to-SOH mapping learned in the Source temperature domain when applied to a Target temperature domain. Meanwhile, full-lifetime labeled data are often scarce in the Target temperature domain, making costly retraining-from-scratch impractical for rapid deployment. To address these challenges, a Temporal Graph Network–Long Short-Term Memory Network (TGL-Net) is developed for cross-temperature SOH Estimation, together with a transfer learning paradigm of Source-domain Pre-training followed by Few-Cell Fine-Tuning in the Target temperature domain for fast adaptation. TGL-Net takes a cycle-level 16-dimensional health-feature sequence as input. A Conv1d module first extracts short-horizon degradation patterns while suppressing local fluctuations; a chain-structured temporal Subgraph is then constructed via a Sliding Window, and 2 TimeGNN layers encode neighboring-cycle dependencies through message passing; subsequently, 2 stacked LSTM layers model mid-to-long-term degradation trends. A Linear output head produces the SOH estimate for the end-of-window Cycle, enabling full-curve Estimation over the battery lifetime. Experiments on 25&#xa0;°C test batteries show that TGL-Net achieves a Mean Absolute Error of 0.0075, a Root Mean Squared Error of 0.0106, and an R-squared of 0.9961, with an inference time of approximately 0.033&#xa0;s per Battery and a model size of about 0.737&#xa0;MB. Moreover, after Fine-Tuning with only a small number of Batteries in the 15&#xa0;°C/35°C Target temperature domain, Estimation accuracy is markedly improved compared with zero-shot direct transfer. Systematic evaluation using a 1008-run hyperparameter sweep, ablation studies, and comparisons with multiple baseline methods indicates that stable performance can be maintained over a broad hyperparameter range, together with favorable cross-temperature Transferability and Data efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-temperature state-of-health estimation of lithium-ion batteries via few-cell fine-tuning using a temporal graph network with sliding-window subgraph modeling

  • 期刊论文

  • 发表日期：

    2026年5月1日

查看全文
  • Jifei Zhao,
  • Yanan Zhou,
  • Ye Yuan,
  • Wenwen Dou,
  • Yaqi Liang,
  • Jitian Song,
  • Yuqian Fan

摘要

As electric transportation and energy storage systems continue to scale up, lithium-ion battery health management imposes increasingly stringent requirements on cross-temperature State of Health (SOH) Estimation. However, temperature variations alter charging kinetics and the statistical scale of health features, inducing a systematic bias in the feature-to-SOH mapping learned in the Source temperature domain when applied to a Target temperature domain. Meanwhile, full-lifetime labeled data are often scarce in the Target temperature domain, making costly retraining-from-scratch impractical for rapid deployment. To address these challenges, a Temporal Graph Network–Long Short-Term Memory Network (TGL-Net) is developed for cross-temperature SOH Estimation, together with a transfer learning paradigm of Source-domain Pre-training followed by Few-Cell Fine-Tuning in the Target temperature domain for fast adaptation. TGL-Net takes a cycle-level 16-dimensional health-feature sequence as input. A Conv1d module first extracts short-horizon degradation patterns while suppressing local fluctuations; a chain-structured temporal Subgraph is then constructed via a Sliding Window, and 2 TimeGNN layers encode neighboring-cycle dependencies through message passing; subsequently, 2 stacked LSTM layers model mid-to-long-term degradation trends. A Linear output head produces the SOH estimate for the end-of-window Cycle, enabling full-curve Estimation over the battery lifetime. Experiments on 25 °C test batteries show that TGL-Net achieves a Mean Absolute Error of 0.0075, a Root Mean Squared Error of 0.0106, and an R-squared of 0.9961, with an inference time of approximately 0.033 s per Battery and a model size of about 0.737 MB. Moreover, after Fine-Tuning with only a small number of Batteries in the 15 °C/35°C Target temperature domain, Estimation accuracy is markedly improved compared with zero-shot direct transfer. Systematic evaluation using a 1008-run hyperparameter sweep, ablation studies, and comparisons with multiple baseline methods indicates that stable performance can be maintained over a broad hyperparameter range, together with favorable cross-temperature Transferability and Data efficiency.