Physics-Informed Neural Networks for Thermal Anomaly Prediction in Battery Energy Storage Systems

Battery Energy Storage Systems (BESSs) are increasingly deployed in grid-scale applications, electric mobility, and renewable integration, where safety, reliability, and longevity are critical. Thermal runaway remains one of the most severe failure modes in lithium-ion batteries, often triggered by complex interactions between electrochemical, thermal, and mechanical phenomena. This paper presents an extended hybrid Physics-Informed Neural Network (PINN) framework for thermal anomaly prediction and early detection of runaway precursors in BESS. The proposed architecture integrates governing physical laws, specifically the Bernardi heat generation equation and Fick’s diffusion law, within a deep learning pipeline composed of a physics module, a temporal Bi-LSTM, and an attention mechanism for explainability, which may represent an obstacle in the application of deep learning algorithms. Beyond the initial formulation, the extended version presented here provides a deeper theoretical background, an expanded methodological justification, a more comprehensive comparison with state-of-the-art approaches, and a detailed discussion on scalability, uncertainty, and deployment challenges. The results for synthetic yet physically consistent datasets represent a proof of concept of the PINN approach, which can achieve superior generalization, robustness to noise, and interpretability compared to purely data-driven baselines, achieving an accuracy above 90% and an AUC of 0.95. The framework contributes to proactive safety management in cyber-physical energy systems and establishes a foundation for real-time, physics-aware anomaly detection in safety-critical BESS applications, e.g., marine transportation contexts and port environments.