PhyP-Net: Physics-Constrained Deep Learning for Sub-Second EV Power Demand Prediction in Hybrid Energy Storage Systems
IEEE Access · 2026 · 14 · 97698-97717
DOI: 10.1109/access.2026.3707203 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Jaehee Hong und 1 weiteren Autor:innen. Die Arbeit nennt 53 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Institute of Information and Communications Technology Planning and Evaluation -Innovative Human Resource Development for Local Intellectualization Program gra…, Korea Institute of Marine Science and Technology Promotion funded by the Ministry of Oceans and Fisheries.
Abstract (Originalsprache)
Hybrid energy storage systems (HESS) combining batteries and supercapacitors require sub-second power demand forecasts for nonlinear model predictive control (NMPC) in electric vehicles (EVs). However, existing EV power prediction studies operate at temporal resolutions of one second or coarser, leaving the sub-second regime unexplored. This paper proposes PhyP-Net, a physics-constrained power forecasting network that predicts EV power demand at 10 Hz resolution—matching the NMPC control cycle—using 11 input features derived from 7 controller area network (CAN) bus signals. The architecture integrates a residual long short-term memory (LSTM) encoder with attention-based temporal pooling and a multi-head predictor that decomposes the 50-step (5 s) horizon into short-, mid-, and long-term heads. Training combines a time-weighted data-driven loss with a Huber-based physics constraint enforcing the powertrain relationship and multi-task learning of torque and revolutions per minute (RPM) auxiliary targets. Experiments on a Tesla Model 3 dataset show that PhyP-Net attains a short-term coefficient of determination of 88.90% and the strongest 5 s multi-step accuracy. Although a linear baseline achieves a lower one-step peak failure rate (PFR), PhyP-Net provides the most robust transient tracking across the full horizon and substantially outperforms the recurrent and Transformer baselines, achieving the best overall accuracy–reliability balance. Ablation shows that multi-task learning produces the largest impact on transient reliability (PFR), whereas removing the physics constraint increases the internal-consistency violation by 58% at negligible statistical cost. Embedded inference on an NVIDIA Jetson consumes 4.64 ms per sample (4.6% of the real-time budget), confirming real-time inference feasibility on embedded hardware.
Förderung
- Institute of Information and Communications Technology Planning and Evaluation -Innovative Human Resource Development for Local Intellectualization Program gra…
- Korea Institute of Marine Science and Technology Promotion funded by the Ministry of Oceans and Fisheries
Zitieren
APA
Hong, J., & Im, T. (2026). PhyP-Net: Physics-Constrained Deep Learning for Sub-Second EV Power Demand Prediction in Hybrid Energy Storage Systems. IEEE Access, 14, 97698-97717. https://doi.org/10.1109/access.2026.3707203
Vancouver
Hong J, Im T. PhyP-Net: Physics-Constrained Deep Learning for Sub-Second EV Power Demand Prediction in Hybrid Energy Storage Systems. IEEE Access. 2026;14:97698-97717. doi:10.1109/access.2026.3707203
BibTeX
@article{Hong2026,
author = {Hong, Jaehee and Im, Taeho},
title = {PhyP-Net: Physics-Constrained Deep Learning for Sub-Second EV Power Demand Prediction in Hybrid Energy Storage Systems},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {97698--97717},
doi = {10.1109/access.2026.3707203}
}
EUVIMED