Graph neural network-based multi-agent reinforcement learning for active voltage control: Performance and topology robustness

Energy and AI, vol. 25, pp. 100835

Abstract

Active voltage control in power distribution networks with high photovoltaic penetration requires scalable, topology-aware strategies that can exploit the graph-structured nature of electrical grids. We present a multi-agent reinforcement learning framework that integrates graph neural network architectures directly into the distributed actor networks, enabling each agent to reason about local grid topology at inference time. Three representative graph neural network families are compared against conventional topology-unaware baselines under matched parameter budgets: a spectral-based graph convolutional network, an attention-based graph attention network, and a memory-based gated graph convolutional network. Experiments on IEEE 33-, 141-, and 322-bus distribution systems show that topology-aware policies match baseline performance on small networks but deliver substantial improvements on medium and large networks. Specifically, they reduce voltage violations by approximately 47% on the 141-bus system and achieve the lowest voltage-out-of-control rate of 3.20% on the 322-bus system. Beyond radial networks, we report the first multi-agent reinforcement learning evaluation of voltage control on looped (meshed) distribution networks, obtained by reconfiguring the radial feeders, and find that graph neural network-based policies maintain comparable performance under this topology modification. Furthermore, our edge-feature ablation reveals that explicit encoding of line impedances is not essential for strong control performance, and our approach consistently yields lower power losses than both traditional optimization-based and topology-unaware methods across all network sizes. These results indicate that graph neural network-based multi-agent reinforcement learning is a promising and energy-efficient approach for distributed voltage control in modern power grids.

Authors 3

1. Chijioke Eze corresponding Aachen Institute for Automation of Complex Power Systems

   RWTH Aachen University

   Affiliation as printed

   Institute for Automation of Complex Power Systems, RWTH Aachen University, Mathieu Strasse 10, Aachen, 52074, North Rhine-Westphalia, Germany

2. Chetan Prakash Aachen Faculty of Computer Science

   RWTH Aachen University

   Affiliation as printed

   Faculty of Computer Science, RWTH Aachen University, Ahornstr. 55, Aachen, 52074, North Rhine-Westphalia, Germany

3. Antonello Monti Aachen Institute for Automation of Complex Power Systems

   RWTH Aachen University · Fraunhofer Institute for Applied Information Technology

   Affiliation as printed

   Fraunhofer FIT, Schloss Birlinghoven, Konrad-Adenauer-Strasse, Sankt Augustin, 53757, North Rhine-Westphalia,

   Institute for Automation of Complex Power Systems, RWTH Aachen University, Mathieu Strasse 10, Aachen, 52074, North Rhine-Westphalia, Germany