Centrally Coordinated Multi-Agent Reinforcement Learning for Power Grid Topology Control

• DOI: 10.48550/arxiv.2502.08681
• Publication Date: 2025-02-12
• Authors: Barbera de Mol; Davide Barbieri; Jan Viebahn; Davide Grossi
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; I.2.11; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; I.2.6; I.2.8; Computer Science - Multiagent Systems; I.2.11; I.2.8; I.2.1; I.2.6; I.2.1; Multiagent Systems (cs.MA); Machine Learning (cs.LG)

Power grid operation is becoming more complex due to the increase in generation of renewable energy. The recent series Learning To Run a Network (L2RPN) competitions have encouraged use artificial agents assist human dispatchers operating power grids. However, combinatorial nature action space poses challenge both conventional optimizers and learned controllers. Action factorization, which breaks down decision-making into smaller sub-tasks, one approach tackle curse dimensionality. In this study, we propose centrally coordinated multi-agent (CCMA) architecture for factorization. approach, regional actions subsequently coordinating agent selects final action. We investigate several implementations CCMA architecture, benchmark different experimental settings against various L2RPN baseline approaches. exhibits higher sample efficiency superior performance than results suggest high potential further application higher-dimensional as well real-world settings.