Dual-mode resilient dispatch for V2G systems with PBT-enhanced multi-agent deep reinforcement learning

Vehicle-to-grid system holds dual potential for enhancing grid economic efficiency and resilience against extreme disasters. However, existing works typically rely on deterministic assumptions lacking proactive anticipation, and the balancing of multi-dimensional objectives depend on experience-based weighting. To address these limitations, this paper proposes a dual-mode resilient dispatch framework integrating a day-ahead situational awareness module with a Population-Based Training (PBT)-enhanced multi-agent deep reinforcement learning strategy. To achieve accurate system mode identification under data scarcity, a semi-supervised perception module is constructed, utilizing density-based clustering to label unknown data and a deep generative model to synthesize samples. On this basis, a differentiated multi-objective mechanism is established to optimize synergistically the economic efficiency in normal modes and the system survivability in resilient modes. Furthermore, the framework incorporates a PBT evolutionary strategy to search for Pareto-optimal weights, ensuring adaptive multi-agent collaboration. Results suggest the superiority of the proposed method over existing benchmarks, demonstrating effective coordination in both routine operation and emergency response scenarios.