Convolutional autoencoder anomaly detection and classification based on distribution PMU measurements

Abstract

The huge volume of data that is streamed from distribution phasor measurement units (DPMU) toward distribution management system (DMS) can rarely be used in the raw format. Data-driven analysis to extract information out of massive raw data has revealed promising opportunities to overcome this challenge. This paper utilizes convolutional autoencoders (Conv-AE) for the sake of anomaly detection based on the DPMU measurements in distribution systems. The Conv-AE is unsupervised and independent of the event type. It dispenses preprocessing and feature extraction along with obtaining the essential information directly from the measured data. The anomaly detection is followed by a supervised classifier which is developed to identify the type of anomaly. This classifier is a convolutional neural network that is designed and fine-tuned for the problem at hand. It uses an ensemble learning method to augment the fully-labelled dataset. Performance of the proposed methodology is evaluated on modified IEEE 34-node and IEEE 123-node test feeders in various operating conditions, presence of noise, and different scenarios of DPMU outage. Moreover, a real-world DPMU dataset is utilized to evaluate performance of the Conv-AE model in practical conditions. Results confirmed effectiveness of the proposed technique to be used in future DMS platforms.