Data Driven Control and Management of Power Networks

Power grids globally play a crucial societal and economic role by providing uninterrupted, reliable, and transient-free power to industries, businesses, and households. The advent of renewable power resources and electric vehicles has introduced uncertain generation and highly dynamic load demands, making robust operation of power networks essential to manage transient stability issues and localize blackout events. In light of the increasing stress on modern grid infrastructure and operators, this talk will cover HybridAgent, a reinforcement learning (RL) framework designed to mitigate the effects of unexpected network events and reliably maintain electricity across the network. HybridAgent leverages a heuristic-guided RL framework for optimal topology selection, ensuring safe and reliable grid operation without overloads. It has been extensively tested in the Learning to Run a Power Network (L2RPN) 2023 Challenge hosted by Réseau de Transport d’Électricité (RTE) and Delft University of Technology (TU Delft). With its state-of-the-art AI-based framework, HybridAgent outperformed other existing approaches on multiple datasets featuring forced contingencies in the IEEE-118 network. Despite its reduced action space, HybridAgent topped the leaderboard in the L2RPN NeurIPS 2020 challenge (Robustness track) and was the top-performing agent in the L2RPN WCCI 2020 challenge. Additionally, HybridAgent achieved third position in the L2RPN 2023 challenge hosted by RTE and TU Delft. Detailed analysis demonstrates HybridAgent’s state-of-the-art performance in several test scenarios.

Sep 1, 2024 1:00 PM — 1:20 PM