Weakly-Supervised, Strongly Reliable: Machine Learning Challenges for Secure Energy Operations
Date:
In this talk, I shared my reflections on the role of machine learning in secure energy system operations. Power systems are becoming increasingly complex due to higher renewable penetration, greater uncertainty, and faster dynamics from inverter-based resources. Traditional, human-centered decision-making is reaching its limits, as operators cannot respond quickly enough to emerging disturbances.
At the Delft AI Energy Lab, we explore how different machine learning paradigms — including supervised and surrogate models, physics-informed learning, weakly-supervised approaches, reinforcement learning, graph neural networks, and self-supervised “foundation models” — can support operators in maintaining reliability. While these approaches offer promising advances, they also raise fundamental challenges: data scarcity, generalization to unseen scenarios, sim-to-real gaps, and the need for interpretable and trustworthy models.
My central message is that AI for power systems must not only be innovative but also reliable and explainable. Progress will require careful integration of physics with data-driven methods, and close collaboration between researchers and system operators. Ultimately, the goal is to develop models that we can trust — models that help secure, sustainable, and effective energy system operations.
Slides: