On the robustness of machine-learnt proxies for security constrained optimal power flow solvers

Abstract

In this paper, we focus on the robustness of machine learning based proxies used to speed up, alone or jointly with state-of-the-art mathematical optimization methods, optimal power flow and security-constrained optimal power flow calculations. On data sets for the Nordic32 alternative current security-constrained optimal power flow benchmark, we evaluate the robustness of proxies with respect to load distribution, power factors, on-line generators and network topology, and generator costs. We show that simplified random load sampling procedures that are used in most published academic studies, are insufficient to yield robust machine learnt proxies, and consequently limit their usefulness in the real world. Based on these results, we formulate recommendations for future research.