Application of Machine Learning Algorithms for Electricity Theft Detection in Smart Grids

Abstract

Electricity theft is a global problem that negatively affects both electricity generation and distribution companies and ordinary consumers. More specifically, it destabilizes the economic development of utility companies, causes risks and negatively affects the cost of energy. The development of smart grids (SGs) plays an important role in detecting electricity theft. Intelligent electric grids are expressed by a two-way flow of information, producing massive data regarding customer consumption, which through machine learning and deep learning techniques, can be utilized to detect electricity theft.In the context of this masters thesis, automated systems based on machine learning and deep learning were developed, with the aim of detecting cases of rheumatic electricity theft in SGs. To address dataset weaknesses, such as the problems of missing measurements and class imbalances, linear interpolation and synthetic minority oversampling (SMOTE) techniques were applied to generate synthetic data, respectively. The proposed dataset was evaluated using four different machine learning techniques: logistic regression (LR), decision trees (DT), random forests (RF), and convolutional neural networks (CNN), to determine the optimal settings that achieve the highest accuracy. In all algorithms, their maximum performance was examined, with the aim of improving the results from existing works, showcasing that the use of the SMOTE method in combination with the proposed Convolutional Neural Network, yields the maximum accuracy results. Overall, the competitiveness of the method compared to other methods evaluated on the same dataset was demonstrated. Finally, though the validation procedures of the CNN model, we showcased 98.3% AUC and 99.4% accuracy, thus significantly outperforming other machine learning models in existing works.