Comprehensive Machine Learning for Lithium-Ion Battery State-of-Health Estimation Using Group-Wise Cross-Validation

The application of Machine Learning (ML) models in the estimation of Lithium-ion Batteries' (LIBs) State-of-health (SOH) has been expanding, and prior studies report high scores from train-test splits, likely inflated by the split. Although some works mention data leakage, explicit treatment of leakage and its impact on reported performance has received limited attention in the battery SOH literature. To address this gap, an evaluation of seven machine-learning models for SOH estimation was conducted using a group-wise cross-validation protocol that keeps all cycles from the same cell in a single fold. Based on the aging data, random train-test splits yielded an R2 of up to 0.9999 and an RMSE of 0.0815. With group-wise cross-validation, most models achieved an R2 of around 0.9100 and an RMSE of 3, resulting in a reduction of up to 9% in optimistic bias. Overall, the results support the adoption of cell-wise GroupKFold as the standard evaluation protocol for reliable out-of-cell generalization.