Gomes, Douglas Pinto Sampaio (2020) Vegetation High-Impedance Fault Detection and Characterization using Machine Learning. PhD thesis, Victoria University.

Abstract

Vegetation High-Impedance Faults (VHIFs) are relevant and under-addressed power dis- tribution system disturbances. They are low-energy events, represented by the contact between power lines and nearby vegetation, that are not detected by traditional protection devices. Despite not harmful to power equipment, they can ignite fires in vegetation with great potential to life and property damage. After devastating HIF-related fires in 2009, the Victorian Government found the lack of technical solutions to prevent similar disasters and funded a vegetation ignition testing program to foster further research. It staged hundreds of VHIFs that generated the data pertained to this thesis. In the related literature, High-Impedance Faults (HIFs) comprise an extensive research field, but few works are solely dedicated to studying VHIFs. Although generally treated as a single problem, different high-impedance conducting surfaces introduce significant variance in faults’ characteristics and behaviours. For these reasons, the staged VHIFs recordings represent a niche type of faults having specific behaviours with significant potential for insights regarding phenomenon characterization. The main contributions from this thesis result from using the staged VHIF data to address the knowledge gaps related to its characterization and detection method. Initial investigations presented the likely presence of discriminative features in the signals’ high- frequency (HF) spectrum. The results gave confidence for the production of a machine learning-based VHIF classifier, conceptualized and discussed as part of a potential detection method. Subsequently, the existence of discriminative information and invariance in the HF signals was proved with the application of renowned signal representation techniques and machine learning algorithms. A study regarding the importance of using HF signals was also performed to support the chosen approach when conceptualizing the classifier. It led to the finding that although the accessibility of such signals might be not optimal, they may be imperative for an effective VHIF detection method. To deflate some of the potential implementation concerns, a low-cost, proof-of-concept prototype was produced, attesting the capabilities of real-time classification. Lastly, an unsupervised learning technique was used to capture some of the convoluted and complex fault signatures in the time domain. The found patterns led to insights about VHIF behaviour and signatures signals that resulted in more detailed phenomena characterization.

Search Google Scholar

Repository staff login