A standardized comparative framework for machine learning techniques in lithium-ion battery state of health estimation

Abstract

The accurate estimation of lithium-ion battery State of Health (SOH) is essential for enhancing performance, safety, and lifecycle management in modern energy systems. While numerous individual studies have explored machine learning approaches for SOH prediction, a systematic comparative analysis using consistent experimental protocols and rigorous cross-validation remains limited. This study addresses this gap by presenting the first comprehensive comparison of three advanced machine learning models—Extreme Gradient Boosting (XGBoost), Random Forest, and Support Vector Machine (SVM)—using a standardized experimental framework with NASA battery datasets. Our novel contribution lies in implementing a unified training-testing protocol using battery #5 for training and batteries #6, #7, and #18 for validation, combined with systematic hyperparameter optimization through grid search and k-fold cross-validation. Key improvements include: (1) first standardized one-to-many validation protocol ensuring cross-battery generalization assessment that eliminates the data splitting limitations of previous comparative studies, (2) unified hyperparameter optimization methodology applied identically across all algorithms, eliminating the confounding effects of inconsistent parameter tuning that have biased previous comparisons, and (3) establishment of quantitative performance benchmarks providing evidence-based model selection criteria for practical Battery Management System (BMS) applications. The XGBoost model achieved superior performance with MAE of 0.016 and MSE of 0.000347, establishing empirical benchmarks for model selection in battery health diagnostics through our systematic comparative methodology. This work provides the first standardized comparative framework for SOH estimation, offering evidence-based guidance for BMS implementations and advancing the field toward more rigorous and replicable research practices in battery prognostics.