6 - A graph mapping based supervised machine learning strategy for PMU voltage anomalies' detection and classification in distribution networks

Abstract

Timely anomaly detection and classification in voltage signals for distribution systems allows the design of preventive and corrective control actions to avoid damage or loss of equipment, as well as partial or total power outages. Recently, PMUs are being used to monitor distribution systems, where these capture the dynamic information of the system through the variables of voltage, current, and phase angle. The PMU sampling rate (10–60 samples/s) allows voltage anomalies to be captured within and outside the normal operating conditions (±2% of nominal voltage value). Hence, the PMU records can be processed using algorithms for the detection and classification of anomalies or intermittent small-magnitude events. Currently, there are different algorithms that properly perform one or both stages (detection and classification); however, they require a filtering stage and a decomposition of the data register into mono-components for their correct operation, which translates into information loss and an increase in computational burden. Furthermore, the filtering stage limits the detection and classification of small-magnitude anomalies in voltage, opening an opportunity for the development of new algorithms. In this chapter, a supervised machine learning strategy is presented; the strategy combines graph theory with recurrence quantification analysis for features extraction that allows anomaly detection and classification. The proposed methodology is robust to noise with low computational burden and easy interpretation, therefore its application for online monitoring of distribution systems is feasible. To validate the proposal, case studies are presented to analyze scenarios with voltage anomalies related to power quality events and small-magnitude intermittent voltage anomalies that occur during normal operation conditions. Thus synthetic records generated through a Monte Carlo model and PMU records obtained from a distribution system are processed.