Evidence-gated evaluation of cross-dataset battery state-of-health transfer with a diagnostic state-of-charge proxy
Lower prediction error on an unseen battery dataset does not necessarily demonstrate transferable degradation learning. A model may better align its prediction level without improving its ability to follow changes in individual cells. This study presents an evidence-gated framework for evaluating cross-dataset battery state-of-health (SOH) transfer. It distinguishes absolute-error reduction from centered-error improvement and within-cell rank association, with uncertainty assessed at the cell level. Each dataset is held out in turn, and access to target information follows explicit rules. Few-shot adaptation uses early-cycle observations and a fixed future evaluation horizon. Calibration baselines help interpret apparent gains, while a trace-derived state-of-charge proxy serves only offline diagnostic purposes. Results across four public datasets do not establish consistent transfer behavior. Selected settings combine lower error with positive within-cell associations, but their small magnitudes provide only limited tracking evidence. Other settings yield inconclusive or inverse associations. A target with narrow, high-SOH-dominated support permits diagnostic interpretation only. Greater model capacity or auxiliary supervision does not uniformly resolve this heterogeneity. Changing the reference-capacity definition also alters some evidence classifications, making the corresponding conclusions reference-conditional. These findings show that lower prediction error alone does not establish transferable degradation tracking. Transfer claims must reflect the magnitude, uncertainty, and stability of the supporting evidence, without generalizing success on one target to overall method superiority.
Authors
- Mahdi Jalili (ORCID: https://orcid.org/0000-0002-0517-9420)
- Amir Sheikhahmadi (ORCID: https://orcid.org/0000-0002-0696-2900)
- Laleh Tafakori (ORCID: https://orcid.org/0000-0002-0815-8094)
Institutions
- La Trobe University (AU)
- RMIT University (AU)
Publication Details
- Journal
- Journal of Energy Storage
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1016/j.est.2026.125781
- Primary Topic
- Battery Modeling and Management
- Type
- article
- Field-Weighted Citation Impact
- 0.00