Multi-stage data-driven prediction with physics-constrained coordination for the remaining useful life of battery

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries (LIBs) is critical for energy storage system safety and reliability. However, existing methods face two fundamental challenges: physical inconsistency from decoupling of data-driven features from electrochemical mechanisms, and stage-adaptation issues due to the nonlinear capacity decay in early-cycle (0–100) and long-term aging (100–300 cycles) phases. To address these , we propose X-PABO, a physics-informed Bayesian optimization framework integrating Extreme Gradient Boosting (XGBoost) with Physics-Informed Neural Networks (PINNs). The framework introduces two key innovations: a Stage-Adaptation Index (PAI) that dynamically coordinates XGBoost (for rapid early-cycle trend capture) with PINNs (for long-term aging modeling under electrochemical constraints), and a multi-criteria physical-constraint fusion mechanism enforced via a composite loss function with adaptive weighting, ensuring four-dimensional consistency . Experimental results from the MIT dataset demonstrate that X-PABO achieves state-of-the-art performance, with a root mean square error (RMSE) of 86 cycles and a mean absolute percentage error (MAPE) of 7.7%, representing a 15–20% improvement over existing methods. Shapley Additive Explanations (SHAP) analysis validates the physical consistency of PINN-generated features and their dominant role in RUL prediction accuracy. The proposed framework provides a physically interpretable and computationally efficient solution for real-time battery health monitoring, supporting next-generation battery management systems (BMS).