User Preference-Aware Deep Reinforcement Learning for Multi-Objective Optimal EV Charging Control

초록

This paper presents a deep reinforcement learning (DRL)-based electric vehicle (EV) charging control framework that dynamically optimizes the supplied voltage and current in response to real-time battery states, environmental conditions, and user preferences. Unlike conventional model-based controllers, the proposed method learns optimal charging strategies through interaction with a virtualized electro-thermal environment, enabling adaptability to nonlinear battery dynamics and stochastic user behavior. The charging problem is formulated as a constrained Markov decision process, and Proximal Policy Optimization (PPO) is employed to learn adaptive control strategies. A hybrid two-stage learning structure is designed, combining indirect reinforcement learning in simulation for safe policy acquisition and direct learning for real-world adaptability. A multi-objective reward function is developed to reflect user-specific trade-offs among charging efficiency, charging time, and target SoC achievement, while safety constraints are embedded through penalty terms. Extensive experiments demonstrate that the proposed DRL controller achieves faster convergence, improved charging efficiency, and stronger constraint satisfaction compared to the conventional constant current-constant voltage (CC-CV) and model predictive control (MPC) benchmarks.

키워드