Bayesian-optimized machine learning models for state of charge estimation of lithium-ion EV batteries: sensitivity analysis and XAI via SHAP

Precise state-of-charge (SOC) estimation is essential for reliable battery management systems (BMSs) in electric vehicles (EVs). This study presents a comprehensive comparative evaluation of four advanced ensemble machine learning (ML) models, namely CatBoost, Histogram-based Gradient Boosting (HGBoost), LightGBM, and AdaBoost, for SOC estimation using a real-world dataset acquired from a BMW i3 EV equipped with a 60 Ah lithium-ion battery pack. Model hyperparameters were optimized using Optuna-based Bayesian optimization, while model generalization capability was assessed through cross-validation and holdout testing. Performance of the investigated ML models was evaluated using root mean squared error (RMSE), mean absolute error (MAE), standard deviation of residuals (STD.DEV), maximum error (MAX), coefficient of determination ( \(R^2\) ), distribution-based error analysis, and noise sensitivity assessment. All investigated boosting models outperformed conventional benchmarks such as random forest and extreme learning machine. Among them, CatBoost achieved the highest prediction accuracy with an RMSE of 5.0386% and an MAE of 3.7333%, whereas AdaBoost demonstrated superior robustness with the lowest MAX error. Furthermore, Explainable Artificial Intelligence was incorporated using SHapley Additive exPlanations to quantify feature importance and enhance the transparency and interpretability of SOC prediction mechanisms.