Generative AI for Open-Set Event Detection and Classification in Power Systems Using Synchrophasor Data

Reliable event analytics are essential for the intelligent operation of cyber–physical power systems. This article presents a generative artificial intelligence framework for synchrophasor-based event detection and classification that is data-efficient and robust to previously unseen disturbances. A variational autoencoder-generative adversarial network is trained only on normal operating data to learn the manifold of healthy behavior; reconstruction and discriminator errors serve as complementary anomaly indicators. Two decision mechanisms are developed: a weighted threshold rule for computational efficiency and a convex-hull boundary that adapts to nonlinear error distributions. Streaming phasor measurement unit measurements are processed with a sliding window, producing spatiotemporal detection and classification matrices that are integrated through decision fusion to yield system-level labels. The proposed framework achieves 97.3% (threshold) and 99.3% (convex-hull) detection accuracy, outperforming conventional machine-learning and envelope-based baselines. For classification, the fusion of sliding-window evidence improves overall accuracy to 93.9% and, critically, enables open-set recognition by assigning previously unseen events to a dedicated category rather than forcing them into known classes. These results demonstrate an informatics architecture that couples generative modeling with geometric decision fusion to enhance real-time situational awareness and resilience in industrial power systems.