An Unsupervised Learning Framework for Event Detection, Type Identification and Localization Using PMUs Without Any Historical Labels

The power system requires new monitoring and controls due to changes both at the generation side as well as the load side. Synchrophasor technology with synchronized and high-resolution measurements provided by Phasor Measurement Units (PMUs) has been recognized as a key contributing technology for advanced situational awareness, including event identification, where the application of machine learning techniques is a hot topic recently. However, recent methods focus on supervised learning techniques that require event records, which may be unavailable due to labeling cost. Even if labels exist, the uneven labeled data may cause biased learning models. To address these challenges, an unsupervised learning approach is proposed for conducting fast event identification. Specifically, a highly sensitive and accurate change-point detection method is firstly introduced for finding events via data distribution changes. After detection, event type identification is achieved via a two-stage information filtering. In stage 1, we use cluster number in principal component analysis (PCA) to split the event types. In stage 2, we narrow down the type by evaluating cluster compactness for measuring event severity. Finally, we solve the event localization problem based on a hierarchical clustering to group PMUs with significant changes across change points. Numerical results show fast and robust performances of the proposed methods for different events at different locations.