Abstract:

Electric vehicle fast charging requirements will substantially impact the operation of linked power-transportation networks as the adoption rate of electric vehicles rises. This work suggests a multi-objective system-level fast charging station recommendation approach to dynamically assign electric vehicles to suitable stations to advance the interests of the coupled system, fast charging stations, and electric vehicle users. The recommendation problem is formulated as a sequential decision-making problem, and a deep reinforcement learning method is adopted. Graph attention networks are introduced to deal with the network-structure coupled system states. Considering the heterogeneity between entities, the authors propose a physical connection-based graph formulation method with feature projection to integrate multi-dimensional information from charging stations, traffic nodes, and power grid buses into a graph. The graph convolution of coupled system states can then be realized to promote environment perception. Besides, to address the long time-delay action execution in recommendation problem, a double-prioritized DQN(λ) training mechanism is developed to update the guidance strategy, where an attention-prioritized cache construction method is proposed to improve the training efficiency cooperated with prioritized experience replay. The proposed graph reinforcement learning method is trained and evaluated in a joint power-transportation simulation platform. Simulation results show that the proposed strategy can promote the interest of multiple facets in coupled power-transportation networks by handling the requests in a real-time manner. Its feasibility and robustness in urban transportation systems are also demonstrated.