GMAT: A Graph-Enhanced Multi-Agent Transformer for Coordinated Voltage Regulation in Heterogeneous Distribution Networks

Abstract

Multi-agent reinforcement learning faces challenges in capturing the spatial interdependencies and physical constraints of electrical networks. Existing paradigms often treat agents as isolated entities, leading to fragmented situational awareness and suboptimal coordination. To overcome these limitations, we introduce the Graph-Enhanced MultiAgent Transformer (GMAT) for real-time coordinated voltage regulation. GMAT constructs a topology-aware graph representation from the feeder topology and current nodal voltagepower features to extract spatial dependencies in the distribution network. These embeddings are subsequently processed by a Multi-Agent Transformer framework to model high-order inter-agent coordination dependencies among heterogeneous agents, including photovoltaics, energy storage systems, electric vehicle charging stations, and smart buildings. Simulations in the PowerGridworld environment demonstrate that GMAT exhibits superior sample efficiency and reward convergence, maintaining voltages within prescribed thresholds while enhancing system-wide energy efficiency. This method provides a robust, topology-aware solution for the intelligent and coordinated operation of power distribution networks.