Scholarlink

With the global shift to clean energy, lithium-ion batteries are vital for energy storage, necessitating long service lives and accurate State of Health (SOH) predictions for reliability. Current methods often overlook battery multi-field coupling and physical constraints, compromising accuracy. This paper introduces a hybrid approach combining physical constraints with a Graph Convolutional Network-Long Short-Term Memory (GCN-LSTM) model to enhance SOH prediction accuracy and robustness. We identify charging intervals as aging indicators through voltage platforms and uniformity processing. A physics-constrained GCN-LSTM framework is then developed for SOH estimation, validated via perturbation analysis. Tests on single-cell and battery pack datasets under various conditions show high accuracy, with mean absolute errors below 0.38%, and strong noise resistance, maintaining errors under 0.58% with 20% noise.

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

Robust State of Health Estimation for Energy Storage Batteries Based on Physical Constraints and Graph Convolutional Deep Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuxin He,
  • Zhongwei Deng,
  • Jiwei Wang,
  • Chunlin Jiang,
  • Wenhao Nie

摘要

With the global shift to clean energy, lithium-ion batteries are vital for energy storage, necessitating long service lives and accurate State of Health (SOH) predictions for reliability. Current methods often overlook battery multi-field coupling and physical constraints, compromising accuracy. This paper introduces a hybrid approach combining physical constraints with a Graph Convolutional Network-Long Short-Term Memory (GCN-LSTM) model to enhance SOH prediction accuracy and robustness. We identify charging intervals as aging indicators through voltage platforms and uniformity processing. A physics-constrained GCN-LSTM framework is then developed for SOH estimation, validated via perturbation analysis. Tests on single-cell and battery pack datasets under various conditions show high accuracy, with mean absolute errors below 0.38%, and strong noise resistance, maintaining errors under 0.58% with 20% noise.