Early Anomaly Detection and Classification with Streaming Synchrophasor Data in Electric Energy Systems

The large-scale streaming data collected from the increasing deployed phasor measurement unit (PMU) devices poses significant difficulties for real-time data-driven analytics in power systems. This dissertation presents a dimensionality-reduction-based monitoring framework to make better use of the streaming PMU data for early anomaly detection and classification in power systems. The first part of this dissertation studies the fundamental dimensionality of large-scale PMU data, and proposes an online application for early anomaly detection using the reduced dimensionality. First, PMU data under both normal and abnormal conditions are analyzed by principal component analysis (PCA), and the results suggest an extremely low underlying dimensionality despite the large number of raw measurements. In comparison with prior work of utilizing multi-channel high-dimensional PMU data for power system anomaly detection, the proposed early anomaly detection algorithm employs the reduced-dimensional data from PCA, and detects the occurrence of an anomaly based on the change of core subspaces of the low-dimensional PMU data. Theoretical justification for the algorithm is provided using linear dynamical system theory. It is demonstrated that the proposed algorithm is capable to detect general power system anomalies at an earlier stage than would be possible by monitoring the raw PMU data. The second part of this dissertation investigates the classification of a special anomaly in power systems, low-frequency oscillation, which may cause severe impacts on power systems while at the same time is difficult to be accurately classified. We present a robust classification framework with online detection and mode estimation of low-frequency oscillations by using synchrophasor data. Based on persistent homology, a cyclicity response function is proposed to detect an oscillation, through the use of the low-dimensional features (pre-PCA features) extracted from PCA. Whenever the cyclicity response exceeds a numerically robust threshold, an oscillation can be detected. After the detection, PCA is applied again to extract the low-dimensional features (post-PCA features) from the multi-channel transient PMU data. It is shown that the post-PCA features preserve the underlying modal information in a more robust way in comparison to raw synchrophasor measurements. Based on the post-PCA features, fast Fourier transform (FFT) and Prony analysis can be subsequently applied to extract modal information of the oscillation. The proposed classification framework offers system operators a data-driven analytical tool for fast detection of low-frequency oscillation and robust mode estimation against high measurement noise