Interpretable hybrid GRU–XGBoost Ensemble framework for lithium-ion battery SoH estimation

Accurate estimation of the State of Health (SoH) of lithium-ion batteries is essential for ensuring the safety, reliability, and operational efficiency of battery management systems in electric vehicles and energy storage applications. However, battery degradation exhibits highly nonlinear and time-dependent behavior, making SoH prediction a challenging task. This study proposes an interpretable hybrid and ensemble learning framework for SoH estimation by integrating Gated Recurrent Units (GRU) with Extreme Gradient Boosting (XGBoost). The GRU network is employed to extract latent temporal degradation features from sequential battery measurements, while XGBoost performs nonlinear regression using the learned feature representations. In addition, a weighted ensemble strategy is introduced to combine standalone GRU and hybrid GRU–XGBoost predictions in order to improve robustness and prediction stability. The proposed framework is evaluated using the NASA lithium-ion battery dataset under a time-aware train–test split protocol. Experimental results demonstrate that the ensemble model outperforms standalone and hybrid baselines, achieving an MAE of 1.40, RMSE of 1.95, and R² of 0.9898. Furthermore, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are employed to provide global and local interpretability of the prediction process. The results indicate that the proposed framework improves prediction accuracy while maintaining model transparency, demonstrating strong potential for practical battery management system applications.