Consumption Pattern Detection Through the Use of Machine Learning : Clustering Techniques for Non-Technical Losses Detection RERORT

Abstract

Machine learning algorithms applied towards detection of non-technical losses are increasingly becoming a go-to solution and tools that help detection losses in such a way are being developed by power utilities of all sizes across the globe. Non-technical losses represent the biggest fraction of distribution losses in electric power systems of all countries making utilities experience significant revenue fatalities. These losses are largely represented by fraudulent activities performed by the customers such, which lead to additional losses, damaging the infrastructure of networks and deteriorating the safety of the grid. This thesis proposes an extensive review of different methods and case studies of non-technical losses in grids in different parts of the world, their effect on the economies of scale and other aspects of social well-being such as links between the rate of fraud being committed in the national grid to human development index (HDI) and other metrics. Modern technologies of power consumption estimation and regulation such as smart meters are given a close look as well as their infrastructure, role and possible applications are also reviewed in this thesis. As for the machine learning part overview, over 80 research papers were carefully studied in order to quantify the progression of interest towards applied artificial intelligence algorithms from the industry, key performance indicators and state-of-the-art fraud detection models were analysed and findings were extracted. The findings suggest several optimal application thresholds of learning techniques under examination with respect to specifics of the problem, available resources and budgets as well as volume and quality of the data. In the experimental section of the thesis, a specific threat model was developed and effectively employed through the generation of synthetic fraudulent profiles according to a methodology presented in the literature. The underlying data that was used to satisfy the question and the scope of this thesis was extrapolated from a larger source of consumption record that was in open access for all in such a way stimulating the further development of the field by a wide range of researchers and scientists around the world. The algorithm of choice, namely k-means clustering, was profoundly reviewed, studied and applied and at a later instance executed as three parametristic approaches to tackle the same problem from different perspectives and get a deep look-out of the unsupervised clustering and the patterns in the data. Results were insightful and superiority of performance was shifting between an approach that used artificial feature extraction and the approach with kWh reading samples being examined over a dynamic time warping temporal sequencing algorithm. Finally, the conclusions were drawn suggesting the scope of future work followed by a brief cost-benefit analysis designed for utilities considering investing in the fraud detection tool development as well as an environmental sustainability outlook of the thesis concept’s implementation is proposed.