Carbon-Aware V2G Coordination in PV-Rich Distribution Systems via Weather-Incorporated Two-Stage Graph Reinforcement Learning and Data Distillation

Abstract

The increasing penetration of photovoltaic (PV) generation introduces significant volatility and PV-load temporal mismatch into power distribution systems, necessitating advanced strategies to maintain efficient and sustainable operation. While vehicle-to-grid (V2G) technology presents a promising solution by leveraging electric vehicle (EV) fleets as distributed energy storage, coordinating a large fleet of EVs for PV-rich distribution systems remains a complex, high-dimensional problem. To address this challenge, this paper proposes a novel weather-incorporated two-stage graph reinforcement learning approach based on data distillation for carbon-aware V2G coordination. In the first stage, stochastic operational data, including PV generation and load profiles, are distilled according to weather and day types to simplify data patterns. The resulting distilled profiles are then used to pre-train a guidance policy. At the second stage, the pre-trained guidance policy facilitates more effective policy learning for the refined policy by providing guidance actions, enabling it to efficiently handle stochasticity in the original operational data and thereby improving overall learning performance. To further enhance the adaptability of V2G coordination, we design a hierarchical graph neural network (HGNN) that learns spatiotemporal correlations involved in the V2G coordination problem and effectively adapts to changes in system configuration, such as the addition or removal of charging stations, without requiring retraining. Evaluations are conducted on a modified PV-rich IEEE 33-bus system incorporating multiple charging stations, demonstrating its suitability for practical deployment in PV-rich distribution systems.