Nested cell-level voltage-window selection for partial-charge state-of-health estimation of lithium-ion batteries
Year: 2026
Type: preprint
Abstract: Estimating the state of health (SOH) of lithium-ion batteries from a short part of the charging curve is attractive for battery management systems, because vehicles and devices are rarely charged from empty to full. Published errors are often very low, but the voltage window and the data split are usually chosen by the authors, and the sensitivity to measurement conditions is seldom reported. We study a simple pipeline: five features from a voltage window of the constant-current charging phase feed a ridge regression, and the window is selected using only the training cells (nested leave-one-cell-out). On three public datasets of LCO/NMC-type cells (lithium cobalt oxide or nickel–manganese–cobalt oxide cathodes; NASA, 4 cells; CALCE, 6 cells; Oxford, 8 cells), the mean root-mean-square error (RMSE) of SOH is 3.01, 2.28 and 0.72 %, against 5.9, 2.8 and 2.3 % for a linear model that knows only the cycle index. Ridge regression is statistically indistinguishable from random forest, gradient boosting and support-vector regression (paired Wilcoxon \(\:p\approx\:0.42\), 18 cells) and is much cheaper. The best window differs between datasets; a window shift of 0.02 V increases the error by a factor of 1.6 to 3.3 on NASA and Oxford; 10 mV of voltage noise doubles the NASA error unless the voltage is filtered; and models do not transfer between datasets. On 119 lithium iron phosphate (LFP) cells (Supplementary Material), ridge does not beat the cycle-index baseline, and tree ensembles are only moderately better.
Source: Research Square
Authors: Huy Hoang Le, Kim-Anh Nguyen
Institution: University of Da Nang
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236309758
