A Machine Learning Approach to Event Analysis in Distribution Feeders Using Distribution Synchrophasors

This paper proposes a machine learning (ML) approach to detect, identify, and analyze the events that occur on distribution networks using data streams from real-world distribution-level phasor measurement units (PMUs). First, we develop two statistical event detection methods. One is based on testing absolute values around median and the other one is based on testing residuals on a non-linear estimation. Both methods use moving windows as well as dynamic window size. This allows us to detect events of different types and durations. Next, we use field expert knowledge to assign labels to the detected events and subsequently develop a multi-class support vector machine classifier to classify power quality events. Finally, we apply the above developed techniques to detect, identify, and analyze the events in a micro-PMU data stream from a real-world test site in Riverside, CA. We particularly study the oscillation events that occur somewhere across the distribution feeder itself, where their impacts are observed remotely by the available micro-PMUs.