Combining Thermal–Electrochemical Modeling and Deep Learning: A Physics-Constrained GRU for State-of-Health Estimation of Li-Ion Cells

Battery health monitoring is essential for ensuring the safety, longevity, and efficiency of energy storage systems, particularly in critical applications where reliability is im-portant. Traditional methods for assessing battery degradation, such as Electrochemi-cal Impedance Spectroscopy (EIS), are effective but impractical for large-scale de-ployment due to their time-intensive nature. This study introduces a novel mod-el-based approach for estimating a critical indicator of battery aging, the internal re-sistance. Using the NASA battery dataset, specifically focusing on batteries number 5 and 7 with NCA chemistry, a comprehensive framework that integrates advanced predictive models, i.e. the Random Forest Regressor (RF), the XGBoost Regressor (XGBR), the Gated Recurrent Unit (GRU), and the Long Short-Term Memory (LSTM) networks, was developed. The models were evaluated using common regression met-rics, while hyperparameter tuning was accomplished to optimize performance. The results demonstrated that recurrent neural networks, particularly GRU and LSTM, ef-fectively capture the temporal dependencies inherent in battery aging, offering more accurate State of Health (SOH) predictions. This approach significantly improves computational efficiency and prediction accuracy, paving the way for practical appli-cations in Battery Management Systems (BMS).