Unsupervised Physics-Informed Transfer Learning for Power Flow

Maintaining stability and operational efficiency in today’s dynamic power grids demands rapid, real-time decision-making. While machine learning (ML) methods offer fast inference, they are typically trained on data from static grid configurations, such as fixed topologies, which limits their adaptability to evolving conditions. Adapting these models to new grid states often requires generating additional datasets using computationally intensive solvers, adding further complexity. In this work, we propose an unsupervised transfer learning approach that embeds the power balance equation directly into the loss function, providing a more efficient alternative to conventional transfer learning methods. This approach enables the adaptation of ML models for power flow (PF) analysis under changing grid conditions without the need for newly labeled training data. Our results demonstrate a 110× improvement in fine-tuning efficiency across standard test systems, including the IEEE 14-bus, IEEE 118-bus, and PEGASE 1352-bus networks. We validate our method through practical implementation and comparative evaluation.