Development of electricity theft detection and mitigation in smart grid

Abstract

Electricity theft (ET) is a ubiquitous problem ravaging all electric utilities worldwide. Theft of electricity is caused by so many factors, but developing a formidable anti-theft solution is one of the major problems facing electric utilities globally. Like a virus, ET is slowly wreaking havoc on power utilities worldwide and its dreaded curves need to be flattened. Since ET cannot be totally eradicated in power grids, the motivation for this research is to profoundly detect and mitigate ET in electric networks. ET must be utterly detected and mitigated to uncover the power pilferers, promote healthier electricity grids, generate more income for the utilities, improve the reliability and sustainability of power systems, and consequently help in salvaging the economies of nations worldwide. Power losses occasioned by ET could be redressed by either generating more power to compensate for the theft-inflicted power shortfalls or by mitigating the theft, but mitigating the theft is more significant and more cost effective. Artificial intelligence-based (AI-based) machine learning (ML) methods are the state-of-the-art and superior approach for the detection of ET or non-technical losses (NTL) in power grids when compared with the conventional methods of electricity-theft detection (ETD). The experimental work in this thesis centres on the detection of ET using the real-world energy consumption dataset provided by the State Grid Corporation of China (SGCC), a state-owned SG electric system, and the largest electric utility company in the world. The case-study dataset which has thus been obtained from the smart meters of electricity consumers is formidable because it has been used extensively in the existing literature by many researchers to develop various ETD models. This gives room for comparison of results among several ETD models developed using same SGCC dataset. In the experiments, ETD is performed with the infusion of the features from convolutional neural network (CNN) model into random forest (RF) model to form a hybrid model termed CNN-RF. The hybridization of the models is done in a quest to achieve better NTL prediction results, as the combined strengths of CNN and RF achieves complete elimination of undesirable false positives in the composite model. RF is noted to be highly effective and efficient in resolving classification problems, hence it is a choice candidate for the hybrid solution. Meanwhile, before finally adopting the proposed CNN-RF model, the performances of CNN and RF models were individually checked. Simulations were performed using Python, in a Google Colaboratory (Colab) Integrated Development Environment (IDE). The performance metrics employed to evaluate the developed models are precision, recall, F1 score, accuracy, Matthews correlation coefficient (MCC), area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (PR-AUC), true negative rate (TNR), false positive rate (FPR), and false negative rate (FNR). The proposed model show very interesting and reliable performance results, achieving 100.00% precision, 98.36% recall, 99.17% F1 score, 99.20% accuracy, 98.40% MCC, 99.13% AUC, 99.55% PR-AUC, 100.00% TNR, 0.00% FPR, and 0.02% FNR. Overall, the proposed model outperformed other SGCC dataset-based ETD model results presented in previous research. The proposed model achieves unprecedented high hit ratio, making it more-effective and more-efficient in detecting NTL. Higher performance scores from ETD models are proportional to greater mitigation of NTL attainable by utility inspectors or technicians during onsite inspections. The feat achieved in this research by profoundly detecting ET in SG, with its anticipated increased onsite mitigation prospects, is a fulfilment of the aim and objectives of the research. Besides, the higher detection capability achieved by the proposed model has also simultaneously proffered answers to the research questions. The proposed model is therefore recommended as a suitable ETD solution for deployment by electric utilities of various economies of the world.