Detection of electricity theft in developing countries-A machine learning approach

Abstract

In developing countries, energy theft negatively affects the growth of utilities through loss of revenue and damage to the grid. The size and variety of the utility data set require extracting meaningful features to counter theft, which is difficult and computationally expensive. Recent developments have made machine learning more accessible to researchers, enabling its application in big data analysis for power utilities. Through greater access to training resources, as well as commercial and open-source machine learning tools, it has become easier to test large sets of data against various algorithms and automate many of the processes such as data cleaning and feature extraction, a procedure known as Automated Machine Learning (AutoML). These tools, along with frequent data collection by utilities, lend themselves to the use of machine learning to solve power grid issues such as anomaly detection. This paper focuses on feature extraction from monthly consumption records, previous investigations, and other customer information to detect power anomalies critical in the detection of theft. Using AutoML, features were extracted, and models were then trained and tested on data gathered from investigations. The results show that by using machine-learning algorithms, anomaly detection can be 4 times more effective than present manual detection techniques, increasing from 10% to 40% while reducing the number of unnecessary audit investigations by 61%.