Statistical Feature Selection From Massive Data in Distribution Fault Diagnosis

Selecting proper features to identify the root cause is a critical step in distribution fault diagnosis. Power engineers usually select features based on experience. However, engineers cannot be familiar with every local system, especially in fast growing regions. With the advancing information technologies and more powerful sensors, utilities can collect much more data on their systems than before. The phenomenon will be even more substantial for the anticipating Smart Grid environments. To help power engineers select features based on the massive data collected, this paper reviews two popular feature selection methods: 1) hypothesis test, 2) stepwise regression, and introduces another two: 3) stepwise selection by Akaike´s Information Criterion, and 4) LASSO/ALASSO. These four methods are compared in terms of their model requirements, data assumptions, and computational cost. With real-world datasets from Progress Energy Carolinas, this paper also evaluates these methods and compares fault diagnosis performance by accuracy, probability of detection and false alarm ratio. This paper discusses the advantages and limitations of each method for distribution fault diagnosis as well.