Explainable AI-based prediction of lithium-ion battery aging: A comparative study at low and moderate temperatures

Abstract

This study investigates temperature-driven degradation mechanisms in lithium-ion batteries using machine learning and explainable artificial intelligence, comparing performance at 4 °C and 24 °C. A dataset of 7000 electrochemical impedance spectroscopy measurements demonstrates pronounced thermal effects on aging behavior. At 24 °C, capacity retention remains above 0.90 over extended cycling, gradually declining to 0.80–0.85, whereas operation at 4 °C shows accelerated degradation, with capacity dropping below 0.85 during early cycles and reaching 0.70–0.75 prematurely. Charge transfer resistance (Rct) below 50 mΩ corresponds to more than 80% remaining useful life (RUL), while Rct values exceeding 100 mΩ indicate critical degradation. SHapley Additive exPlanations (SHAP) analysis identifies capacity (±0.35), Rct (−0.40 to −0.05), and temperature (±0.20) as dominant predictive features. Among eight regression models evaluated, Random Forest achieves the best performance with R2 = 0.933, MAE = 13.36 cycles, and RMSE = 36.62 cycles, followed by Light GBM and XG Boost, which significantly outperform linear models. Temperature–capacity interactions reveal non-monotonic effects: moderate temperatures of 20–25 °C enhance RUL, whereas extreme conditions of 0–5 °C and 30–40 °C accelerate degradation. This framework integrates electrochemical insights with data-driven modeling, supporting advanced battery management and thermal optimization strategies. These findings improve predictive reliability, enable early fault detection, and guide robust lifetime-aware control policies for practical deployment