Hybrid methodology based on knn regression and boosting classification techniques for locating faults in distribution systems

Abstract

This  paper  presents  a  hybrid  methodology  for  fault  location  in  power  distribution  systems,  by  using  a  regression  technique  based  on  k nearest neighbors and a classification technique which uses multiple simple classifiers in a Boosting strategy. The proposed methodology first subdivides the power system into zones to train the classification technique. Then the parameters of the regression technique are adjusted and the classification technique is trained. Finally, for an unknown case, the regression technique estimates the fault distance, and the classification technique locates the fault in one of the predefined zones, solving the multiple estimation problem. The IEEE34 node test feeder is used to test the proposed fault locator, where a good performance is obtained either in the classification task (minimal precision of 95.7 %) and the regression task (highest absolute error of 8.05%). The proposed method can be easily implemented in a power system, is fast and has low computational effort.