Classification of Faults in Power System with Probabilistic Neural Networks: An Imbalanced Learning Approach

Modern day power grids with its inherent operating characteristics are susceptible to faults. Grid operators must detect as well as classify the current system operating conditions like normal or faulty from the current raw sets of measurement data available at supervisory control and data acquisition (SCADA) system. With the rapid deployment of micro PMUs, faults are detected from the raw measurements in real time, but their classification still possess a challenging task. This paper focuses on a diligent comparison between several deep and machine learning techniques for classifying faults in real time. In real life scenarios, line to ground (L-G) faults being the most frequent one while three phase to ground (LLL-G) faults being rare, an imbalanced dataset is generally developed for supervised learning approach leading to biased classification of faults. In order to alleviate this current concern, data oversampling policy over the imbalanced dataset based on synthetic minority oversampling technique (SMOTE) is proposed. The dataset used in this work is derived from the Drexel University's Reconfigurable Distribution Automation and Control (RDAC) software/hardware laboratory.