Scholarlink

<p>State of charge (SOC) estimation of lithium-ion batteries remains a critical challenge for battery management systems due to nonlinear dynamics, chemistry-dependent characteristics, temperature variations, and aging-induced degradation. Conventional model-based methods offer physical interpretability and noise robustness, yet their adaptability is limited under changing operating conditions. Purely data-driven approaches, on the other hand, can learn complex nonlinear behavior but often suffer from reduced generalization under domain shift and require large amounts of labeled data. To overcome these limitations, this study proposes a domain-adaptive hybrid H∞–LSTM framework for robust SOC estimation. The proposed method combines a two-time-constant equivalent circuit model and an H∞ filter to generate a physically consistent auxiliary SOC estimate, which is then integrated into an LSTM network to enhance temporal modeling and improve robustness. The framework is pretrained on NMC battery data and subsequently adapted to NCA and LFP chemistries through single-stage transfer learning (TL) using limited target-domain data. Experimental results demonstrate that the proposed hybrid model achieves high estimation accuracy under cross-domain conditions, with RMSE values of 1.48% for NCA and 2.29% for LFP, while substantially outperforming plain LSTM models. The method also shows strong robustness across temperature variations from 0&#xa0;°C to 40&#xa0;°C, with RMSE values between 1.57% and 1.82%, and maintains reliable performance under aging conditions, yielding RMSE values of 1.60% at 90% SOH and 1.81% at 80% SOH. In addition, the proposed architecture remains computationally efficient, supporting its suitability for embedded and real-time battery management systems (BMS) implementation. These results indicate that the proposed H∞–LSTM framework is a practical and transferable solution for SOC estimation across diverse battery chemistries and operating conditions.</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

A hybrid H∞–LSTM framework with transfer learning for robust state of charge estimation under varying battery conditions

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Ebubekir Buğra Özarslan,
  • Senem Kurşun

摘要

State of charge (SOC) estimation of lithium-ion batteries remains a critical challenge for battery management systems due to nonlinear dynamics, chemistry-dependent characteristics, temperature variations, and aging-induced degradation. Conventional model-based methods offer physical interpretability and noise robustness, yet their adaptability is limited under changing operating conditions. Purely data-driven approaches, on the other hand, can learn complex nonlinear behavior but often suffer from reduced generalization under domain shift and require large amounts of labeled data. To overcome these limitations, this study proposes a domain-adaptive hybrid H∞–LSTM framework for robust SOC estimation. The proposed method combines a two-time-constant equivalent circuit model and an H∞ filter to generate a physically consistent auxiliary SOC estimate, which is then integrated into an LSTM network to enhance temporal modeling and improve robustness. The framework is pretrained on NMC battery data and subsequently adapted to NCA and LFP chemistries through single-stage transfer learning (TL) using limited target-domain data. Experimental results demonstrate that the proposed hybrid model achieves high estimation accuracy under cross-domain conditions, with RMSE values of 1.48% for NCA and 2.29% for LFP, while substantially outperforming plain LSTM models. The method also shows strong robustness across temperature variations from 0 °C to 40 °C, with RMSE values between 1.57% and 1.82%, and maintains reliable performance under aging conditions, yielding RMSE values of 1.60% at 90% SOH and 1.81% at 80% SOH. In addition, the proposed architecture remains computationally efficient, supporting its suitability for embedded and real-time battery management systems (BMS) implementation. These results indicate that the proposed H∞–LSTM framework is a practical and transferable solution for SOC estimation across diverse battery chemistries and operating conditions.