Towards forecasting and prediction of faults in electricity distribution network: a novel data mining & machine learning approach

Silva, HCE 2020, Towards forecasting and prediction of faults in electricity distribution network: a novel data mining & machine learning approach , PhD thesis, University of Salford.

[img] PDF
Restricted to Repository staff only until 7 April 2021.

Download (4MB)

Abstract

The electricity supply system includes a large-scale power generation installation and a convoluted network of electrical circuits that work together to efficiently and reliably supply electricity to consumers. Faults in the electricity distribution network have a direct effect on its stability, availability and maintenance. Consequently, quick elimination, prevention and avoidance of faults and the causes that generated them, is of special interest . The possible opportunity to both analyse the distribution of faults and predict future failures that may arise can significantly help electricity distribution operators who are accountable for the detection and repair of such problems. Such information is also crucial for any future planning and design of electricity distribution networks as it would significantly help to prevent problematic areas or and identify any additional measures necessary for the protection of underground and overground cables and equipment. The derived information would also be very useful to avoid any potential penalties associated with future network faults imposed by the regulators. Any network component faults result in an outage of power not only in the area fed by them but also in the neighbouring area. Fault prediction in distribution systems has always been of immense importance to utilities to ensure reliable power supply. This research aim is to develop data mining, and machine learning models to accurately predict and forecast Electricity Distribution Network Faults. The specific research objectives are to gain a deeper understanding of Electricity Distribution Network faults and to accurately predict network faults using the National Fault and Interruption Reporting Scheme (NAFIRS) database. Furthermore, this research not only proposes solutions but also provides an in-depth discussion of the associated technical, data gathering and data processing challenges. This research employed multiple case research design, as this allows more opportunities for multiple experiments and cross observation . This research has proposed a new method that analyses historical fault data and seeks to understand the impact of faults with other factors such as the Main Equipment Involved, Component and Direct Cause. This proposed data mining model may be used to safeguard the electrical power distribution system’s key equipment which can be severely damaged by some upcoming faults. The author of this thesis has proposed a new fault segmentation framework which distributed network operators can use to perform fault segmentation. This approach gives the option of performing multidimensional segmentation using various fault characteristics such as a number of faults, a number of minutes lost, and a number of customers affected. Multidimensional segmentation is a powerful conceptual model for the analysis of large and complex datasets. This study provides an in-depth discussion of equipment failure related network faults and compares the performance of a range of forecasting methods with a variety of accuracy measures. The study also provides an in-depth analysis of visual data mining concepts and discusses how using 2D and 3D calendar heat map methods can help provide a relatively new perspective in evaluating temporal patterns in electricity distribution network faults. Finally, the research discusses how external factors, such as local population density, affects electricity distribution network faults. Various classification algorithms were used to build prediction models. Those models were validated and compared for accuracy. The author has also sought to accurately understand the behaviour of Customer Minutes Lost (CML) performance indicators and sought to predict the annual CML figure using other annual financial and network performance indicators such as a number of customers affected, Totex, and Network load. It is anticipated that the work presented within this thesis will to lead to several original contributions to the scientific community who are working with data mining, machine learning and electricity distribution networks.

Item Type: Thesis (PhD)
Contributors: Saraee, MH (Supervisor)
Schools: Schools > School of Computing, Science and Engineering
Depositing User: Hethanadura Charith Eshan Silva
Date Deposited: 07 Oct 2020 13:02
Last Modified: 07 Oct 2020 13:03
URI: http://usir.salford.ac.uk/id/eprint/58270

Actions (login required)

Edit record (repository staff only) Edit record (repository staff only)

Downloads

Downloads per month over past year