Graph-Attentional Multiagent Reinforcement Learning for V2G Emergency Dispatch in Cold-Weather Coupled Networks

Urban power-transportation systems are prone to cascading failures in cold weather, rooted in the risk of charging station outages caused by faults of distribution grid and mileage reduction of electric vehicles (EVs). Motivated by this, this paper proposes an online graph-attentional multi-agent deep reinforcement learning (GAMD) method to enhance the power system resilience and EV charging convenience in cold weather. To address the high-dimensional and non-Euclidean characteristics of the state of power-transportation network, the method integrates a graph attention network for global state perception with a decentralized multi-agent system for collaborative decision-making. The proposed method optimizes conflicting objectives, including the resilience of EV charging service, traffic efficiency, and the economic efficiency of emergency power supply. For the first time, the proposed method incorporates the traffic impedance variations induced by V2G dispatch to take into account the dynamic traffic condition. The proposed method is validated with a high-fidelity microscopic simulation platform (SUMO) driven by real-world data. Results suggest the superiority of the proposed method over state-of-the-art benchmarks, with noted improvements across multiple performance metrics. The aforementioned endeavor establishes a robust and scalable paradigm for managing a dispatch-intensified coupled urban power- transportation system.