Estimating Lithium-Ion Battery Capacity from Charging-Curve Health Indicators: How Far Does It Generalise to an Unseen Cell?

Accurate battery capacity estimation is important for evaluating the health of lithium-ion batteries. However, a complete discharge test is difficult to perform when a battery is already installed in a vehicle. This study investigates whether battery capacity can be estimated from health indicators extracted from a single charging curve and whether the prediction can generalise to an unseen cell. The NASA ageing data from four 18650 lithium-ion battery cells were used. After charge-discharge pairing and data cleaning, 628 charging samples were retained. Ten health indicators were extracted and divided into time, voltage and temperature groups. Linear Regression, Random Forest and Gradient Boosting were evaluated using a pooled random split and leave-one-cell-out validation. Under the pooled random split, Random Forest with all features achieved the lowest mean absolute error (MAE) of 0.0119 Ah. However, the tree models showed clear performance reductions under leave-one-cell-out validation. Linear Regression with time and voltage features achieved the best cross-cell result, with an MAE of 0.0329 Ah and an R² of 0.9446. Adding voltage features reduced the cross-cell MAE by 38.6% compared with using time features alone, while temperature features did not provide a further improvement. B0006 was the most difficult cell to predict. The results show that charging-curve indicators can estimate battery capacity effectively, but random splitting may overestimate the generalisation ability of complex models. Testing on unseen cells is therefore necessary before practical application.