Machine-learning-based single-phase-to-ground fault detection in distribution systems

Abstract

It is still a focus of research to detect the faulty feeder timely and accurately in resonant grounding distribution systems. The conventional methods commonly use single faulty feeder detection methods, such as wavelet transform method, transient energy method, and the fifth harmonic current method, etc. However, their reliability is not satisfied due to the partial fault information is considered. A novel approach to identify the faulty feeder based on discrete wavelet packet transform (DWPT) and machine learning is proposed in this paper. The time-frequency matrices are obtained by utilizing the DWPT to the collected transient zero-sequence current signals of the faulty feeder and sound feeders. The feature vectors will be extracted manually by calculating time-frequency matrices with statistical quantities. The two classifiers (Adaboost+CART and SVM) are trained by a large number of feature vectors under various kinds of fault conditions and factors, respectively. The faulty feeder detection can be achieved by the trained two classifiers. A PSCAD/EMTDC simulator is established to simulate a practical 10-kV resonant grounding distribution system. Simulation results of the testing cases validate that the proposed approach of fault detection is able to achieve good identification accuracy.