Electricity Theft Detection Machine-Learning Models for Windhoek Informal Settlements

Abstract

Monitoring and managing resources for electricity distribution is a challenging task, as there are several ways in which revenue leakages may happen. As hardware equipment used in the electricity distribution gets worn out, it becomes inefficient, thus affecting the quality of electricity distributed. Such inherent and unavoidable losses are known as technical losses. Conversely, non-technical losses are man-made and result in direct income loss for electricity utilities through human errors or fraudulent actions. Examples include billing fraud, meter tempering, or illegal connections. This paper focuses on mitigating challenges that arose from non-technical challenges by using machine learning techniques to detect potential electricity theft. Historical electricity purchasing data of Goreangab and Havana clients from 2016 to 2022 was used to train K-means and Isolation Forest clustering algorithms. Early results of the trained models showed K-means clustering with a silhouette score of 0.711 and Isolation forest performing well under various contamination factors. Results also showed the potential of detecting electricity theft early at lower costs and shorter times than conventional monitoring means. Besides, most developing countries still need to install additional equipment for smart grids and usually resort to random meter inspections. Apart from such an alternative being far less precise because of the randomness, it is also costly in the long run as it requires resources to transport inspectors to the client's premises. As researchers, we are convinced that our solution can reduce revenue leakages for utility companies at a low cost with high precision if optimized.