Feature Selection for Battery Lifetime Prediction Using Explainable Machine Learning

Accurate battery lifetime estimation is crucial for health management and system safety. Data-driven research yields extensive feature sets, yet optimal feature selection is often impeded by conventional methods that are computationally demanding, dependent on prior knowledge, and incapable of valuing feature interactions. In this paper, we propose a two-stage feature selection approach that combines recursive feature elimination (RFE) with SHAP (SHapley Additive exPlanations)-informed valuation to improve cross-validated prediction performance. Leveraging seven statistical metrics to quantify feature contributions, our framework achieves a 1.03% absolute improvement in MAPE—a 7.58% relative gain—and surpasses the performance of standard filter and embedded methods. Furthermore, our interpretability analysis uncovers the relationships between key feature values, charging cycles, and percentile windows, providing deeper insights into the underlying battery degradation mechanisms.