Assessing model transferability to predict the state of battery health across independent datasets
Year: 2026
Type: article
Abstract: Data-driven battery prognostic models are usually evaluated within the dataset on which they were trained. This study evaluates the zero-shot transfer of six state of health forecasting models, from regularized linear regression to recurrent and attention-based networks, across three public lithium-ion datasets, together with uncertainty estimates and alarm policies. Dataset separability, measured by classifier AUC and maximum mean discrepancy, did not explain transfer error: the pair with the largest distributional distance included the easiest transfer. Error depended primarily on target domain characteristics rather than source-target separability. The most regularly degrading dataset (ZHU) exhibited the lowest transfer error for every model, supporting target degradation regularity as a leading hypothesis for cross-dataset predictability. At a 10-cycle horizon, however, a persistence forecast that assumes no change had the lowest average cross-dataset error; learned models outperformed it only marginally on a few transfers and were significantly worse on others. Monte Carlo Dropout reached nominal coverage in all but one scenario, but its intervals were often an order of magnitude wider than the forecast error, so a risk-aware alarm policy based on them warned earlier within a dataset yet alarmed almost continuously under transfer. Cross-dataset prognostic benchmarks should therefore report a persistence reference and target-domain predictability alongside model accuracy.
Source: Scientific Reports
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7237348293
FWCI:
Cites:
Cited by:
Related to:
