Comparative Machine Learning Approaches for Battery Capacity Prediction Using Random Forest and LSBoost Regression Algorithms

Accurate capacity prediction is critical for lithium-ion battery health management in electric vehicles and energy storage systems. This study compares four machine learning methods for battery capacity estimation using the NASA Prognostics Center of Excellence dataset, including Optimized Random Forest, Least Squares Boosting, Lasso regression, and Ridge regression. Fourteen features were engineered from voltage, current, temperature, and electrochemical impedance measurements across 447 training cycles from three batteries, with validation on 132 cycles from an independent battery. All models were tuned using Leave-One-Battery-Out cross-validation with fold-specific normalization to ensure unbiased performance evaluation. The LSBoost model achieved the best performance with test RMSE of 0.0794 Ah, MAE of 0.0751 Ah, and R 2 of 0.7351, representing 12.2% improvement in RMSE compared to Optimized Random Forest and 13.6% improvement compared to linear baselines. Feature importance analysis using permutation methods identified cycle number as the dominant predictor followed by mean power and current variability, while partial dependence analysis revealed the marginal contribution of individual features to capacity estimation. The 4.92% mean absolute percentage error meets industrial requirements for battery management systems, establishing gradient boosting as an effective and computationally efficient methodology for state-of-health estimation in practical applications.