Learning-Based Uncertain Dynamic Verification of MMC-HVDC Offshore Wind Systems

This paper devises a learning-based reachable dynamics (Neural-ReachDyn) method to realize real-time dynamic verification of MMC-HVDC-based offshore wind (OSW) systems under infinite uncertain scenarios. Our contributions are threefold: (1) we establish the formulation of Neural-ReachDyn as learning a series of time-varying ellipsoids, enabling a direct generation of reachable sets under any uncertainty levels without the need for repeated reachability analysis in each scenario; (2) we devise shape matrix decomposition techniques for learning non-degenerate, less-time-dependent ellipsoids to achieve enhanced convergence and accuracy during training; (3) we validate Neural-ReachDyn in a representative OSW system with MMC-HVDC connections and verify the dynamics of both the OSW grid and the MMC controllers using the devised method. Extensive case studies demonstrate that Neural-ReachDyn offers an efficacious tool for verifying the multi-time-scale and converter-dominated dynamics of OSW systems in a real-time manner.