Evaluation of hybrid and standalone learning models for predicting lithium-ion battery capacity degradation

Abstract

The prediction of lithium-ion battery capacity degradation plays a vital role in ensuring safe and efficient operation in electric mobility and renewable energy applications. This paper evaluates standalone machine learning, deep learning, and hybrid models for battery capacity estimation. The evaluated ML models include random forest, gradient boosting, and extreme gradient boosting (XGBoost), while the DL model employs a multilayer perceptron. The hybrid framework combines DL based feature extraction with ensemble ML regression or classification. A real-world dataset comprising temperature, resistance, reactance, and battery type was preprocessed, scaled, and divided into training and testing subsets. Hyperparameter tuning, k-fold cross-validation, and uncertainty quantification were incorporated to improve reliability and reproducibility. Model performance was assessed using RMSE, MAE, and R² for regression and receiver operating characteristic–area under the curve (ROC-AUC) and F1-score for classification. ROC curves, calibration curves, metric-comparison charts, cycle-wise degradation plots, and residual analyses were used for evaluation. Results demonstrate that the hybrid model outperforms standalone approaches by reducing RMSE and improving calibration, reliability, uncertainty alignment, and interpretability. This also establishes its novelty over existing state of health (SOH) models and highlights future extensions involving LSTM-based temporal modeling and chemistry-adaptive transfer learning. Overall, hybrid modeling provides a promising solution for reliable predictive battery maintenance.