Using Critical Slowing Down Features to Enhance Performance of Artificial Neural Networks for Time-Domain Power System Data

Abstract

This paper explores deep learning approaches to event classification on real world time-domain power system data. We use a statistical method to measure a physical phenomenon known as critical slowing down (CSD) and use this as a feature engineering preprocessing framework to localize events from large intervals of data. Several previous works have discussed power system event detection, including statistical methods like correlation, Principal Component Analysis (PCA) reconstruction, and local outlier factor search. This work aims to improve upon the statistical methods that have been linked to high-sample rate time-domain event detection and then will be evaluated using artificial neural networks. To evaluate how well CSD localizes events from non-events in high sample rate time-series data, we used a Z-score function to predict the time of an event and extract a six second interval centered around the prediction. The performance of CSD-applied data against the raw data was then compared using two ANN architectures: the Fully Convolutional Network (FCN) and the Residual Neural Network (ResNet). The results of both architectures demonstrate that applying CSD to the data significantly improves event localization for larger data intervals, thus signifying an improvement in event detectability.