Hybrid machine learning framework for battery remaining useful life prediction with time-series modelling and statistical validation

Battery lifetime prediction plays a vital role in the reliability and safety of energy storage systems used in electric vehicles and renewable grids. However, the most existing data driven models either neglect redundant features or fail to generalize under complex operating conditions. This study presents a novel hybrid machine learning framework integrating Random Forest (RF) based feature selection with Long Short-Term Memory (LSTM) networks for accurate and robust prediction of the Remaining Useful Life (RUL) of lithium-ion batteries. RF is first applied to eliminate irrelevant or highly correlated features, thereby improving interpretability and reducing computational cost. The refined features are subsequently modeled by an LSTM to capture long term temporal dependencies in degradation. To further strengthen predictive consistency, Extreme Gradient Boost (XGBoost) is employed at the ensemble layer for nonlinear feature fusion and final RUL estimation. Experimental evaluation on the NASA battery dataset under varied charge/discharge rates and temperature conditions demonstrates that the proposed RF–LSTM–XGBoost model achieves 7–12% lower Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) and enhanced robustness across multiple trials compared to conventional LSTM and other baseline models. Statistical analysis including confidence intervals, Wilcoxon Signed-Rank tests, and Bland–Altman plots confirm the model’s reliability and stability under diverse operating conditions. The novelty of this study lies in the integration of interpretable feature selection with deep temporal modeling and ensemble learning, offering a scalable and computationally efficient framework for practical battery health monitoring systems.