Learning for Power System Dynamics: The Generalization Challenge

Jochen Cremer – TU Delft

Fast, accurate evaluation of power system dynamics can enable new use cases for managing the variability of modern grids. However, currently these use cases are not realized as simulations are slow. ML models were studied in the past as fast surrogates for such dynamic simulations. However, to this date, ML models often fail when disturbances fall outside their training data. This talk explores why extrapolation, especially for discrete events, is so challenging, and how transfer learning helps (and where it falls short), pointing toward new paths for robust, generalizable ML in power system dynamics.