Does Partial-Charge Battery State-of-Health Estimation Transfer Across Cells and Datasets? A Benchmark on NASA, Oxford and CALCE Data
Year: 2026
Type: preprint
Abstract: Data-driven estimation of the state of health (SOH) of lithium-ion batteries from charging curves often reports errors below 1%, but such figures are frequently obtained on cycles of the same cells that were used for training. We build a common benchmark from three public aging datasets (NASA, Oxford and CALCE; 18 cells and 4,936 cycles down to 70% SOH) and estimate SOH from the top-of-charge voltage window of 3.92-4.19 V, which the constant-current charging curves of all three datasets share. Five regressors are evaluated with absolute features, with features normalized to each cell’s own first charge, against a cycle-count baseline and a training-free capacity-ratio estimator. With a random split of cycles the mean error is 0.8 percentage points (pp) of SOH; when a whole cell is held out it is 2.0 pp; when a whole dataset is held out it rises to 7.9 pp and the coefficient of determination becomes negative for most configurations. Counting cycles alone is a strong baseline within the Oxford and CALCE datasets and is the best predictor on held-out CALCE cells (2.3 pp). Reference normalization lowers the cross-dataset error of tree ensembles on two of the three datasets (e.g., from 13.5 to 5.9 pp on CALCE) but not on Oxford. Our analysis links the failure to a dataset-dependent relation between window charge and capacity (slopes 0.66-1.56) and to degradation features that lie outside the common window. Code and protocol are released for reproducible evaluation.
Author: Shahi, Haider Md. Faiyaz Hussain
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225648013