Detecting Power Grid Frequency Events from μPMU Voltage Phasor Data Using Machine Learning

The increase in renewable power generation on the electrical distribution grid is leading to new and unprecedented challenges that system operators have not previously experienced. The increase in asynchronous generation and decrease in rotating electrical loads on the grid are causing a commensurate reduction in system wide inertia. Declining inertia leads to an increased number of frequency and Rate of Change of Frequency (RoCoF) excursions, which if not successfully managed, may lead to unwanted loss of load, generator trips, and damage to assets on the grid. This paper, therefore, investigates the use of high-accuracy voltage phasor data from micro-synchrophasor measurement unit (μPMU) to detect anomalous power system frequency events using unsupervised machine learning techniques. A combination of a feature window selection, Clustering LARge Applications (CLARA) and an adaptive thresholding method has been employed to detect frequency events. This paper has employed real data from two grid-connected solar farms in Norfolk, England, to assess effectiveness in detecting anomalous frequency events, which have also been compared with conventional power quality recordings to validate and assess the method outlined in this paper.