Developing new approaches for detecting and solving security vulnerabilities in internet of things based smart grids

Thanks to the Internet of Things applications, smart grid systems have been developed for more efficient use of traditional electricity grids. Despite its many advantages, smart grid applications can be exposed to cyber-attacks due to the inherent vulnerabilities of communication infrastructures. These cyber-attacks can negatively affect both individuals and societies. Therefore, it is crucial to understand the nature of these cyber-attacks and develop artificial intelligence-based methods to solve them. Detecting cyber-attacks that aim to steal energy for low billing is an important problem to be solved for energy providers in the deployment of smart grid applications. The hybrid usage of data-driven machine learning methods provides satisfactory results for solving the problem. In this thesis, the types of cyber-attacks on the Internet of Things based smart grids are analyzed in detail according to network layers, possible threats and potential solutions are presented. Furthermore, two artificial intelligence-based approaches are proposed for the detection and solution of security vulnerabilities in smart grids. Firstly, a deep learning-based approach has been developed for energy theft detection using balanced and imbalanced datasets derived from user consumption patterns in smart grids. The proposed Deep Neural Network (DNN)-based approach achieves a success rate of 97.46% in detecting one of the two different attack vectors. Secondly, Convolutional Neural Network (CNN) based hybrid methods were developed for the detection of energy theft from balanced datasets in smart grids, and a CNN+LR approach with a success rate of 95.34% was proposed. With this proposed approach, malicious users who want to underestimate the consumption data generated by smart meters through cyber-attacks are successfully detected. In this study, six different attack vectors are applied to the honest user data to synthetically generate falsified consumption data. Five different deep learning based hybrid models are applied for the detection of each attack vector. In order to solve the data imbalance problem, the Generative Adversarial Network (GAN) method, which has almost never been used in this field, is used.