All AI models are wrong, but some are useful… for power systems
Date:
Artificial intelligence (AI) holds great promise for the future of power system operation, yet widespread technological breakthroughs remain elusive. This talk addresses the question of why AI has not yet transformed power systems and explores the pathway to achieving this transformation. While modern AI systems benefit from vast data and immense computational resources, their application in power systems is constrained by physical laws, safety-critical decisions, and limited real-world data.
The talk presents a structured taxonomy of machine learning (ML) paradigms relevant for power systems: supervised learning, physics-informed learning, weakly-supervised and self-supervised learning, reinforcement learning, and graph neural networks. Each paradigm is discussed through the lens of power system applications including dynamic security assessment, optimal power flow (OPF), inverter control, and state estimation. Special attention is given to the challenges of data scarcity, generalization to unseen scenarios, and model interpretability. Several recent research contributions by the speaker and collaborators are highlighted, such as constraint-driven deep learning for N-k security and foundation models for grid learning.
Concluding, the talk calls for collaborative, data-efficient approaches to AI for power grids, emphasizing the need for representations and models that generalize well across changing topologies, grid conditions, and operators. This guest talk was presented at the CRESYM cresROAD event on 08 April 2025.