Physics-Informed Explainable Machine Learning for Hybrid State of Health and Remaining Useful Life Estimation of Lithium-Ion Batteries

Abstract

Accurate health assessment is essential for the reliable and cost-effective operation of lithium-ion battery energy storage systems. This study presents a physics-informed explainable machine-learning framework for hybrid State-of-Health (SoH) estimation and Remaining Useful Life (RUL) prediction under heterogeneous simulated operating and degradation conditions. A multi-cell degradation model represents capacity fade, resistance growth, thermal and operational stress, and nonlinear aging acceleration. Hybrid SoH combines capacity- and resistance-related health information, while end of life (EOL) and RUL are defined exclusively by an 80% capacity-retention criterion. Generalization is evaluated using cell-wise and leave-one-cell-out validation, with right-censored cells excluded from exact RUL error calculation. In the primary evaluation, Gradient Boosting achieves a SoH RMSE of 0.340 percentage points and R2 = 0.9979, while Random Forest provides the lower RUL RMSE of 239.5 cycles with R2 = 0.9278. Whole-cell bootstrap analysis, simulated distribution-shift testing, and expanded feature ablation further assess uncertainty and robustness. Explainability analysis indicates strong model associations with cumulative electrothermal stress, charging-stage behavior, and pulse resistance. The results demonstrate strong simulation-based predictive capability while revealing substantial inter-cell variability in RUL generalization and uncertainty. Independent experimental validation remains necessary to establish transferability to physical battery systems.