Battery prognostics using single-cycle statistical inference and parametric discharge curves
Journal of Energy Storage, cilt.181, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 181
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.est.2026.124158
- Dergi Adı: Journal of Energy Storage
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC
- Anahtar Kelimeler: Curve fitting, Li-ion battery, Machine learning, Prognosis, State-of-health
- Gazi Üniversitesi Adresli: Evet
Özet
Battery management systems require prognostic methods that generalize across diverse operating conditions while remaining computationally efficient. This paper presents a unified prognostic framework whose regression models, once trained offline on multi-cycle historical data from a fleet of batteries, support at deployment: (i) prediction of state of health (SOH) trajectories from a single discharge cycle, and (ii) forecasting of full discharge voltage curves from operating conditions for scenario analysis. The approach is trained and validated using a dataset of 26 Lithium-ion (Li-ion) packs cycled under various discharge regimes. For SOH forecasting, the framework operates in two stages. In the offline training stage, each pack’s full or partial capacity trajectory is represented by a modified exponential decay law, and the governing decay parameter b is extracted via curve-fitting as a supervised regression target. In the subsequent online inference stage, this already-trained regression model predicts b for a new, previously unseen discharge cycle without requiring future cycle observations from that battery, enabling RUL estimation from a single discharge cycle at deployment. This formulation is data-efficient because it recasts SOH prognostics as a low-dimensional regression problem through parametric curve fitting, rather than a time-series forecasting task that typically requires multiple cycles. For discharge-curve forecasting, a constrained tri-regional cubic Hermite parameterization is employed to capture open-circuit potential and loading conditions, with a fixed early-time knot, a cycle-specific knot predicted as a regression target, and an end-time determined by a cutoff-voltage event. Regression models are assessed using cross-validation and Pareto-based selection, and a stacking ensemble is used to enhance robustness and generalization. The resulting models generate consistent discharge curves under the evaluated constant-current profiles and enable what-if studies under candidate operational profiles. Integrating discharge-curve forecasting with SOH prediction yields a prognostics workflow that, once its regression models have been trained and calibrated on experimental cycling data, can reduce the need for additional physical discharge experiments when evaluating new candidate operating conditions.