Electricity-Theft Detection in Smart Grids based on Deep Learning

Abstract

Electricity theft is a major concern for the utilities. With the advent of smart meters, the frequency of collecting household energy consumption data has increased, making it available for advanced data analysis, which was not possible earlier. Indeed, using Smart Grid (SG) networks, which are recently upgraded networks of connected objects, can greatly improve the reliability, efficiency, and sustainability of the traditional energy infrastructure. The SG infrastructure produces a massive amount of data, including the power consumption of individual users. Utilizing this data, machine learning, and deep learning techniques can accurately identify electricity theft users. This thesis presents a Convolutional Neural Network (CNN) based model for automatic electricity theft detection that can achieve high performance classification and detection. The work considers experimentation to find the best configuration of the sequential model (SM) for classification, beginning with two layers and ending with four layers. The best performance has been obtained in two layers’ architecture with the first layer consists of 128 nodes and the second layer is 64 nodes, where the accuracy reached up to 0.92. This enables the design of high-performance electricity signals’ classifier that can be applied several applications. Designing electricity signals classifiers has been achieved using CNN and the data extracted from electricity consumption dataset using SM. In addition, the Blue Monkey (BM) algorithm is exploited to reduce the number of features in the dataset, where these values are used to build models with high performance. In this respect, the emphasis of this thesis has been on reducing the required number features in the dataset in order to achieve a high performance electricity signals’ classifier model. The experiments have justified the high performance of the proposed systems, where combining both the CNN and BM algorithms requires only 666 features viii compared to 1035 features using CNN alone. This demonstrates the superiority of the CNN and BM model over the CNN model in terms of reducing the features of the model while the accuracy remaining the same