Battery Health-Aware Multi-Objective Reinforcement Learning for Electric Vehicle Energy Consumption Optimization
International Journal of Electric Power and Energy Studies · 04.11.2026 · 5 (3) · 11-24
DOI: 10.62051/ijepes.v5n3.02 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Electric Power and Energy Studies, online erschienen am 04.11.2026. Verfasst von Noah Lewis und 1 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Thermal auxiliaries and regenerative braking form the controllable part of electric-vehicle (EV) energy use, but the settings that save energy are not always those that preserve the battery: regenerating into a cold pack promotes lithium plating, and every ampere-hour drawn from a hot pack ages it faster. We cast EV energy management as a three-objective problem-energy drawn from the battery, capacity fade and cabin discomfort-and learn a single preference-conditioned policy for the whole trade-off surface. The proposed preference-conditioned soft actor-critic (PC-SAC) jointly sets the regenerative-braking intensity, the cabin climate power and the battery thermal-management power. It combines decomposed twin critics and preference relabelling with two design elements that proved essential: exogenous re-centring of the learning costs and a physically consistent effort parameterisation of the thermal actions. Driving-scenario probabilities from a contrastive encoder complete the state. Experiments use 3,131 real-world trips (468 h, 27 drivers) from the Vehicle Energy Dataset, a powertrain model calibrated on measured battery power (trip-level R2 = 0.91) and a driver-disjoint test set of 853 trips. On the test drivers, one PC-SAC policy spans a wide trade-off: its health-oriented point changes capacity fade by −28.5 % at +6.2 % energy relative to a production-style controller, and its energy-oriented point changes energy by −21.8 % at the expense of comfort. Measured by its own training objective, it does not improve on the well-tuned production controller at the energy- and comfort-weighted preferences and reaches parity at the health-oriented one (ΔJ = +0.001 ± 0.031), which locates the value of learning in the battery-health region that the rules do not reach. Its front attains a hypervolume of 0.0235 ± 0.0015, against 0.0114 for three mode-specific agents without a battery objective and 0.0204 for three rule-based controllers. In ablations, removing the decomposed critics and preference relabelling lowers the hypervolume.
Zitieren
APA
Lewis, N., & Baker, M. (2026). Battery Health-Aware Multi-Objective Reinforcement Learning for Electric Vehicle Energy Consumption Optimization. International Journal of Electric Power and Energy Studies, 5(3), 11-24. https://doi.org/10.62051/ijepes.v5n3.02
Vancouver
Lewis N, Baker M. Battery Health-Aware Multi-Objective Reinforcement Learning for Electric Vehicle Energy Consumption Optimization. International Journal of Electric Power and Energy Studies. 2026;5(3):11-24. doi:10.62051/ijepes.v5n3.02
BibTeX
@article{Lewis2026,
author = {Lewis, Noah and Baker, Miles},
title = {Battery Health-Aware Multi-Objective Reinforcement Learning for Electric Vehicle Energy Consumption Optimization},
journal = {International Journal of Electric Power and Energy Studies},
year = {2026},
volume = {5},
number = {3},
pages = {11--24},
doi = {10.62051/ijepes.v5n3.02}
}
EUVIMED