Learning for Power System Dynamics: The Generalization Challenge
Date:
This talk focuses on the challenge of generalization in learning-based approaches for power system dynamics, with a focus on transient simulation and security assessment.
While classical RMS and EMT simulations remain central to dynamic studies, their computational bottlenecks limit large-scale scenario analysis. Recent advances in machine learning offer fast surrogate models, yet face critical hurdles in extrapolation, data sparsity, and model accuracy, particularly in low-inertia and rapidly evolving grids. The presentation explores physics-informed and self-supervised learning to enhance robustness, discusses opportunities for grid foundation models to unify EMT and RMS simulations, and highlights the potential of graph-based formulations for transient studies. Key application areas include real-time security assessment, integration of offshore wind, HVDC/FACTS planning, and oscillation analysis. The talk concludes by outlining research needs in data generation, representation learning, and validation to move towards foundation models for transient power system simulations.
Slides: