Automatic Identification of Fault Types in the Distribution Network using Supervised Learning
A. Ranganathan (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Simon H. Tindemans – Mentor (TU Delft - Intelligent Electrical Power Grids)
Peter Palensky – Coach (TU Delft - Intelligent Electrical Power Grids)
J. Dong – Coach (TU Delft - DC systems, Energy conversion & Storage)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Electrical faults in the distribution system can lead to interruptions in customer power supply resulting in penalties that are borne by the distribution system operator. Accurate fault classification is an important step in locating the fault to achieve faster network restoration times. This reduces the operational costs of the system operator by reducing the duration of interruptions in the power supply. Furthermore, it has been found from practical experience that distortions or instabilities in fault waveforms can result in their classification and subsequent localisation being delayed, causing the loss of valuable customer minutes. The objective of this thesis is to investigate the potential of modern supervised learning techniques for the classification of faults in the distribution network. The problem is split into two parts: one aspect is to study the ability of a supervised learning classifier to differentiate between types of stable faults, and in this pursuit, to also identify better criteria for stability.
First, after a review of pertinent literature, it is found that discrete wavelet transforms and support vector machines (SVM) are suitable for fault signal processing and classification, respectively. Next, identifying characteristics, or features, from the three-phase fault current and voltage waveforms are engineered with the help of the db4 wavelet. The features are used as inputs to an SVM classifier model, and the model is tuned and validated to ensure optimal classification performance. Results showed that the developed SVM can differentiate between real-world instances of single-phase, two-phase and three-phase stable faults with a classification accuracy of 95%. A set of business rules are also developed to characterise instability by performing a windowed Fourier analysis and studying the strength of the fundamental frequency component of fault waveforms. The rules are tested on a set of fault data whose stability is uncertain, and it is found that the developed rules are able to improve upon the older method of stability analysis by increasing the rate of stable fault identification.