Conformal Anomaly Detection for Data-Driven Dynamic Models in Power Systems

The increasing integration of distributed resources and complicated inverter interfaces necessitate data-driven dynamic modeling in power systems. This paper focuses on the anomaly detection and localization of data-driven components (DDCs), especially AI-based DDCs, to promptly identify inaccurate or untrustworthy components. The novelty includes: (1) A Support Vector Data Description (SVDD)-based conformal detection approach to facilitate anomaly detection and localization of neural networks; and (2) An enhanced SVDD to further incorporate time-series trajectories and DDC training process, thereby holding the same data preference to provide well-calibrated conformal metrics. Extensive case studies show that the developed conformal anomaly detection can effectively alert DDC inaccuracy and locate the malfunctioning DDC in real-time without relying on ground-truth trajectories.