Unsupervised data analytics for power system modeling and monitoring

Several supervised data analytics methods have been proposed for power system modeling and monitoring applications in recent years. However, these supervised methods require a large number of accurately labeled data to perform well. Due to the scarcity of power system events such as line outages, transformer outages, generation trips, etc., it is difficult to collect enough labeled data and the accuracy of the labels is not guaranteed. In this work, we present unsupervised approaches to address two key problems in power system modeling and monitoring. In the first work, we considered the power system transmission line parameter correction problem. We formulate the parameter correction problem as a sparse unsupervised regression problem by exploiting the sparsity of the parameter errors. We demonstrate that our sparsity-aware approach outperforms benchmarks using the IEEE 57-bus and 118-bus test systems, especially when the lines with parameter errors are closely located. In the second work, we consider the unsupervised detection of power system events based on available PMU measurements. We develop a convolutive dictionary model to capture the unique spatio- temporal correlation of short-time Fourier transform (STFTs) of PMU data streams from multiple PMUs in the presence of a power system event. We leverage this convolutive dictionary model to formulate a binary hypothesis testing for event detection and develop a generalized likelihood ratio test to perform unsupervised event detection. The efficacy of the proposed detector is demonstrated using experiments based on real-life PMU data collected from a US Interconnection.

MAJOR ADVISOR: Jinsub Kim
COMMITTEE: Eduardo Cotilla-Sanchez
COMMITTEE: Xiao Fu
MINOR ADVISOR: Raviv Raich
GCR: Leonard Coop