Transfer Learning for Data-Driven Power Flow and Optimal Power Flow Applications

Abstract

Ensuring real-time grid operations is essential for maintaining both stability and efficiency in today’s dynamic power systems. While machine learning (ML)-based approaches enable fast inference, these models are often trained using datasets derived from static grid configurations, such as fixed topologies. Adapting these models to evolving grid conditions introduces additional complexity and necessitates acquiring supplementary datasets, which require computationally intensive solvers.

This paper presents a method that improves computational efficiency compared to conventional transfer learning techniques for adapting ML models used in power flow (PF) and optimal power flow (OPF) analysis to changing grid conditions. Our findings indicate that only 6% of the original dataset requires recalibration, and the entire process of data point regeneration and model fine-tuning is completed in under 3 seconds in the benchmark IEEE 14-bus, IEEE 118-bus, and PEGASE 1352-bus systems.