Uncertainty Quantification and Failure Prediction for AI-based Congestion Management
Year: 2026
Type: conference-paper
Abstract: The increasing integration of renewable energy sources and demand electrification is increasing the operational complexity of power grids. Reinforcement learning (RL) agents have shown promising performance in real-time congestion management with topology optimization, but quantifying their reliability under unseen operating conditions remains difficult. This paper proposes a framework to quantify and predict the reliability of pre-trained RL agents for real-time congestion management. The approach integrates epistemic and aleatoric uncertainties into a machine learning–based failure prediction model that estimates the probability of RL agent failure one-hour-ahead, accounting for potential contingency scenarios. Results in the Grid2Op environment demonstrate the proposed framework’s ability to anticipate unreliable AI recommendations, achieving an accuracy of 89.5% while maintaining an oversight rate below 7% and a false alarm rate below 11%.
Institutions: INESC TEC, Institute for Systems Engineering and Computers
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7220693449