Scholarlink

<p>Accurate State of Charge (SOC) estimation holds considerable significance for battery management yet remains challenging due to complex electrochemical dynamics and sensing noise. To address the limitations of fixed-parameter equivalent circuit models (ECM) and the interpretability issues of black-box data-driven methods, this paper proposes a robust hybrid framework coupling a Physics-Informed Neural Network (PINN) with an Adaptive Extended Kalman Filter (AEKF). Contrasting with loosely integrated approaches, a differentiable Multi-Layer Perceptron (MLP) is embedded into the ECM topology to model the nonlinear Open-Circuit Voltage (OCV) relationship. Through a physics-constrained training strategy, the framework can simultaneously identify physically consistent circuit parameters and learn the OCV-SOC curve from dynamic data, eliminating the necessity for offline OCV calibration. Furthermore, the learned differentiable model is employed as the state evolution function for the AEKF, where the MLP’s analytical gradient is directly utilized for system linearization. An adaptive noise covariance update mechanism is further incorporated to compensate for model mismatches. Validations on real-world datasets demonstrated that the proposed method delivered a Mean Absolute Error (MAE) below 1% and outperformed existing physics-guided learning approaches in terms of accuracy and robustness against temperature variations and aging.</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

Real-world state of charge estimation for electric vehicle batteries using interpretable physics-informed neural networks

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Rongpu Huang,
  • Chunxiang Zhu,
  • Xingkai Zhou,
  • Changcheng Sun,
  • Zhou Hong

摘要

Accurate State of Charge (SOC) estimation holds considerable significance for battery management yet remains challenging due to complex electrochemical dynamics and sensing noise. To address the limitations of fixed-parameter equivalent circuit models (ECM) and the interpretability issues of black-box data-driven methods, this paper proposes a robust hybrid framework coupling a Physics-Informed Neural Network (PINN) with an Adaptive Extended Kalman Filter (AEKF). Contrasting with loosely integrated approaches, a differentiable Multi-Layer Perceptron (MLP) is embedded into the ECM topology to model the nonlinear Open-Circuit Voltage (OCV) relationship. Through a physics-constrained training strategy, the framework can simultaneously identify physically consistent circuit parameters and learn the OCV-SOC curve from dynamic data, eliminating the necessity for offline OCV calibration. Furthermore, the learned differentiable model is employed as the state evolution function for the AEKF, where the MLP’s analytical gradient is directly utilized for system linearization. An adaptive noise covariance update mechanism is further incorporated to compensate for model mismatches. Validations on real-world datasets demonstrated that the proposed method delivered a Mean Absolute Error (MAE) below 1% and outperformed existing physics-guided learning approaches in terms of accuracy and robustness against temperature variations and aging.