A Diffusion-based Generative Machine Learning Paradigm for Dynamic Contingency Screening

Abstract

Dynamic contingency screening is a challenging task in dynamic security assessment, when traditional numerical approaches are computationally intensive and often not able to repeatedly solve full AC power flow for all possible contingencies in real time, especially for large-scale power grids. Moreover, the severity caused by a contingency is not identical for all operating points, which does not necessitate solving all possible contingencies computationally inefficient and time-consuming. This paper introduces a novel, diffusion-based generative machine learning paradigm that transforms contingency analysis from conventional scenario selection to a proactive, likely-unsupervised scenario generation. The margin to the steady-state voltage stability limit determines the ranking of contingencies corresponding to each operating point. By leveraging physical information from each operating point, the proposed approach anticipates the contingencies most likely to be critical, without relying on static assumptions or exhaustive simulations. This data-prompted generative approach enables the identification of high-risk scenarios under varying load and generator conditions, providing dynamic security assessment in real time. The correctness, effectiveness, and scalability of the methodology are demonstrated through methodological derivations and comprehensive experiments on multiple IEEE benchmark systems, including IEEE-6, IEEE-14, IEEE-30, and IEEE-118, highlighting its potential to incorporate contingency screening in complex, evolving smart grids.