Learning-Based, Runtime Reachability Analysis of Microgrid Dynamics

Abstract

Reachable dynamics (ReachDyn) is a powerful tool for verifying microgrid dynamics under extensive uncertainties, which, however, faces significant challenges in runtime efficiency and numerical stability. This paper devises Neural-ReachDyn, a learning-based reachable dynamics approach to support the runtime uncertain dynamic analysis of microgrids. Our contributions include: (1) set-based Neural-ReachDyn formulation, which establishes neural network-represented ellipsoids for enclosing possible microgrid dynamics under uncertainties in a data-driven manner; (2) set-based Neural-ReachDyn training, which develops an axial length-based loss function to train the reachable set towards conservativeness and tightness with enhanced robustness. Case studies in a typical droop-based microgrid validate the accuracy, efficiency, and adaptability of the devised method under different uncertainties and operating scenarios.