Data-Driven Event Assessment in Power Systems using Gaussian Mixture Models

An event detection and assessment methodology is proposed to identify the instant of occurrence of an event in a power system using measurement data. This is achieved using univariate Gaussian mixture models (GMM) on measured frequency data obtained from all generator buses, wherein a window of the signal is decomposed into a mixture of Gaussians. Evolution of the GMM parameters due to the sliding window is used to pinpoint the exact time when an event is detected on that bus along with its severity. Multiple measurements are further analyzed simultaneously by multi-variate GMM, and complexity of the mixture helps rank severity of the power system events according to their system-wide impact. The proposed techniques have been validated using data obtained by simulating various events on a 10-machine test system, a 16-machine test system, and real wide area measurement data as well.