Through the looking glass: Seeing events in power systems dynamics

This paper presents a new method to identify classes of events, by processing phasor measurement units (PMU) frequency data through deep neural networks. Deep tapered Multi-layer Perceptrons of the half-autoencoder type, Deep Belief Networks and Convolutional Neural Networks (CNN) are compared, using real data from Brazil. A sound success is obtained by a transformation of time-domain signals, from dynamic events recorded, into 2D images; these images wee processed with a CNN, taking advantage of the strong dependency existing among neighboring pixels in images. The training, computing and processing was achieved with a GPU (Graphics Processing Unit), allowing speeding-up of the process up to 30 times and rendering the process suitable to increase the online situational awareness of network operators.