Adaptive Scheduling of Public Electric Vehicle Fast-Charging Stations Based on State-Aware Multi-Agent Reinforcement Learning

This paper proposes a Priority–Urgency Index (PUI) mechanism to address the multi-objective conflict problem in electric vehicle charging scheduling at public fast-charging stations. The mechanism dynamically quantifies the charging urgency of each vehicle based on remaining dwell time, current state of charge (SoC), and target SoC, enabling differentiated service prioritization under resource scarcity. The PUI is systematically integrated into four multi-agent reinforcement learning (MARL) algorithms. To handle the time-varying number of vehicle entities caused by random arrivals and departures, the chargers are modeled as fixed agents; a partially observable Markov decision process (POMDP) is formulated, and a centralized training with decentralized execution (CTDE) architecture is adopted. On this basis, a state-aware dynamic threshold mechanism is introduced to distinguish urgency levels of charging tasks, and an adaptive reward function is designed to accommodate complex operating conditions. Empirical comparisons show that PUI-MAPPO (multi-agent proximal policy optimization) achieves the best performance among all PUI-enhanced variants. Under extreme supply–demand conditions—such as resource-scarce and heavy-traffic scenarios—PUI-MAPPO improves the target-SoC fulfillment rate and net revenue by up to 42.7% and 23.2%, respectively, and reduces the cumulative grid-limit exceedance by 22.8% to 47.3%, relative to the first-come, first-served (FCFS) baseline. Ablation studies further validate the individual effectiveness of the PUI urgency mechanism, the dynamic threshold framework, and the adaptive reward function.