Graph Neural Network-Based Multi-Agent Reinforcement Learning for Active Voltage Control: Performance and Topology Robustness

SSRN Electronic Journal

Abstract

Graph neural networks (GNNs) have emerged as power tools for modeling graph-structured data, motivating their application in constructing policy networks for multi-agent reinforcement learning (MARL) agents in active voltage control. However, existing work has focused primarily on graph convolutional networks (GCN), yet limitations of this architecture have led to alternatives such as graph attention networks (GAT) and gated graph convolutional networks (GGCN). Thus, we develop a general framework for systematic study on these GNN architectures when used as policy networks for MARL on the active voltage control task and implement three GNN-based policies: spectral-based (GCN), attention-based (GAT), and memory-based (GGCN), all of comparable size to conventional vector-based neural network baselines. Through extensive experiments on small, medium, and large-scale power distribution networks, we show that GNN-based policies perform similarly to baselines on small networks but demonstrate remarkable performance advantages on medium and large networks. Further, we conduct a novel evaluation of GNN-based MARL on looped (meshed) power networks and investigate whether incorporation of edge features can significantly impact performance. Our results show that GNN-based approaches maintain near-identical performance on both radial and looped network variants, while achieving competitive power loss efficiency and incorporation of edge features does not significantly impact performance. Additionally, we also compare the performance of the GNN approach against OPF-based method.

Authors 3

1. Chijioke Eze corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

   RWTH Aachen University, Templergraben 55, 52056 Aachen, 52056, Germany

2. Chetan Prakash Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

   RWTH Aachen University, Templergraben 55, 52056 Aachen, 52056, Germany

3. Antonello Monti Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

   RWTH Aachen University, Templergraben 55, 52056 Aachen, 52056, Germany