A hierarchical decision framework for dynamic operation of mobile charging stations

The growing adoption of electric vehicles (EVs) has led to increasing demand for flexible and responsive charging services in urban areas. To address spatiotemporal imbalance between charging supply and EV demand, this study investigates the real-time operation of mobile charging stations (MCSs), a promising solution that leverages portable chargers to meet spatiotemporally distributed charging demand. We propose a two-layer online decision framework that integrates reinforcement learning (RL) with exact optimization to jointly address upper-level deployment of the MCS and lower-level charging service scheduling decisions. In the upper layer, a deep Q-network (DQN) learns an adaptive dispatching policy to determine the next deployment location of the MCS based on real-time system states. In the lower layer, an exact optimization solver computes the optimal charging schedule by jointly selecting which EVs to charge, assigning them to ports, and determining the service duration at each stop. The proposed framework follows an event-driven paradigm and enables closed-loop coordination between hierarchical decision layers. Numerical experiments on both synthetic and realistic settings validate the proposed framework. In small-scale tests, the RL policy achieves 24.5 % higher cumulative rewards than the strongest greedy baseline, with greater improvements over queue-based, static, and random heuristics. In a realistic case study with 20 candidate facilities and spatiotemporally heterogeneous demand, it improves cumulative rewards by 11.2 % over the strongest baseline under practical operational constraints. The findings are robust to variations in the number of ports, dwell time, and demand scale; moreover, in a multi-MCS setting the RL policy also outperforms the strongest baseline. These results highlight the effectiveness, scalability, and practical value of the proposed hierarchical framework for dynamic EV charging operations.