Multi-Objective Engineering Optimization for Electric Vehicle Charging Using Safe Reinforcement Learning

Abstract

Lithium-ion battery fast charging is constrained by lithium plating, solid-electrolyte interphase growth, and thermal runaway, which conventional constant-current constant-voltage protocols cannot mitigate in real time. Existing reinforcement learning approaches address these degradation mechanisms individually but lack a unified architecture coupling domain randomization, predictive constraint enforcement, and multi-proxy degradation awareness. This study develops a risk-aware Soft Actor–Critic controller with a predictive safety filter, trained on a validated single-particle electrochemical-thermal model of a 5 Ah NMC622/graphite cell. Training comprised 500,000 episodes with domain randomization across 5–45 °C, ±15% resistance scaling, and 10–40% initial state of charge; the filter enforced hard constraints on voltage (4.2 V) and temperature (55 °C). At 25 °C, the learned policy reached 80% state of charge in 28.6 ± 0.9 min, 15.9% faster than the tuned baseline (34.0 ± 0.7 min, p < 0.001), while reducing the plating-exposure proxy by 40.0%. At 5 °C, plating exposure decreased by 55.8% with a 7.9% reduction in charge time, and peak temperatures were 2.2–6.2% lower across all conditions. Physics-guided safe reinforcement learning is thus established as a viable framework for simultaneous charging-speed and degradation optimization. Future validation of instrumented cells via differential-voltage analysis and long-term cycling remains essential.