Data-Driven Capacity Estimation of Commercial Lithium-Ion Batteries Using Full-Charge Voltage Relaxation Statistics

Accurate capacity estimation is essential for lithium-ion battery health monitoring and safe operation, yet conventional capacity measurements based on complete charge–discharge tests are difficult to implement online. This study investigates full-charge voltage relaxation statistics for data-driven capacity estimation of commercial 18650 NCA cells. Six descriptors, namely maximum, minimum, mean, variance, skewness, and excess kurtosis, were extracted from each relaxation curve and combined with the cycle number, temperature, and charge/discharge rate. Grouped three-fold cross-validation based on battery identity was applied to prevent information leakage among repeated measurements from the same cell. Elastic Net, support vector regression, random forest, LightGBM, and XGBoost were evaluated under four feature-ablation settings. Relaxation voltage statistics markedly improved estimation accuracy over models using only cycle and operating information. After the cycle number was removed, models incorporating relaxation features and operating variables retained high accuracy, indicating independent degradation-related information in the relaxation response. In the full-feature setting, XGBoost achieved the best performance, with an RMSE of 33.07 mAh, R2 of 0.9745, MAPE of 0.847%, and bias of 0.679 mAh. These findings support relaxation voltage statistics as compact and interpretable features for capacity estimation.