Analysis of non-technical losses in distribution networks using machine learning methods

The human population is constantly increasing on earth. This increase also creates some needs in human life. The need for electricity increases at the right rate with the development level of the world. It is accepted that electricity consumption and usage are higher in societies with a high development level. However, despite the high electricity consumption and usage, the high non-technical electricity loss is thought-provoking. One of the important problems faced by electricity network systems is the losses caused by energy theft; this situation is known as non-technical loss. These unexpected losses seriously threaten the sustainability and reliability of the network infrastructure. The healthy and more efficient operation of energy infrastructure systems is directly proportional to non-technical losses. Detecting and preventing energy theft is of critical importance to ensure a stable energy supply. Reducing non-technical losses provides a healthier energy flow and is also of serious importance in terms of cost. This thesis aims to reduce non-technical losses encountered in the energy sector and to ensure that distribution networks operate more efficiently, sustainably and healthily. In line with the purpose of the thesis, the classification of different types of leakage was carried out using deep learning methods, one of the popular and effective approaches of today. The study offers an innovative approach that aims to both correctly identify existing leakage types and contribute to the energy sector by minimizing energy losses. In the thesis study, the evaluation of leakage data on deep learning architectures using the method we propose adds a significant innovation to the study and is the first study in this context. This innovative approach contributes to the literature on the detection and reduction of non-technical losses in the energy sector and stands out as a pioneering study in terms of showing that deep learning methods can be used effectively in electrical grid analysis. In the study, the data set was evaluated in two dimensions on deep learning architectures. Two-dimensional data was classified with Convolutional Neural Network (CNN) as 97.50% and with Long Short-Term Memory (LSTM) model as 64.17%. In addition, the results obtained from deep learning architectures were compared with conventional (K-NN, SVM) methods. In the study, it was seen that the best result was obtained with the CNN model. Keywords: Non-Technical Losses, Distribution Systems, Illegal Electricity Detection, Classification, Deep Learning