An Efficient and Scalable Algorithm for the Creation of Representative Synthetic AC-OPF Datasets

The research on data-driven optimal power flow applications mostly depends on synthetic datasets due to the lack of accessibility to real-world datasets. These datasets are typically generated in place in the studies present in the literature, with an inadequate focus on the quality of the data used for training the models and on replicability. Most of the studies sample uniformly at random in the neighborhood of the nominal loading, obtaining a homogeneous dataset representative of a small portion of the feasible operating states for a grid, with a poor generalizability for the trained model. This paper proposes an efficient two-step algorithm for the creation of high-quality synthetic datasets for the AC-OPF problem, testing it on grids with up to a thousands of buses. The results confirm that the proposed methodology is consistently better than uniform random sampling, with datasets covering a wider range of operating points.