Expanding the Transient Stability Region of Attraction of Networked Grid-Interactive Inverters: A Probabilistic Active Learning Framework

What the math gives to ML

The paper presents a concrete uncertainty-guided frontier-search mechanism for enlarging a conservatively certified region of attraction. It combines an inner region certified by a neural Lyapunov function with a Gaussian-process surrogate whose posterior uncertainty selects expensive trajectory simulations near the unknown stable-unstable boundary. The transferable asset is an active stability-envelope learner for recurrent networks, state-space models, neural ODEs, and learned controllers. It predicts a sharp transition in the learned stability margin and can reduce the number of long-horizon rollouts required to map that transition.