Deep neural network-based stealthy false data injection attack detection on der integrated systems

Abstract

The world's rapidly growing population has led to an increase in demand for consumption, which in turn requires an energy production system that is both economically feasible and environmentally sustainable. To meet these requirements, renewable energy resources (RES) such as wind, solar, and thermal energy have been utilized to improve energy efficiency, while also adhering to stringent carbon emission regulations. By using these clean and renewable energy sources, we can achieve a more sustainable and environmentally friendly energy production system for the future. The shift towards more sustainable grid structures has resulted in a greater adoption of distributed energy resources (DERs) which allows for the generation and distribution of energy from multiple small-scale sources, rather than relying on a few large power plants. By deploying distributed energy resources (DERs) in close to consumers, we can strategically leverage on-site generation and reduce utility costs, as it eliminates the need for significant investments in expanding the power system network. This approach allows energy to be generated and consumed locally, reducing transmission losses and enabling greater control and flexibility over energy production and consumption. Consequently, DERs can offer a more cost-effective and efficient solution for meeting energy demands while also helping to reduce greenhouse gas emissions. Because the demand for energy to growing and power system topologies and strategies rapidly evolving, traditional power systems have become inadequate in meeting the modern society's energy requirements from multiple perspectives. Conventional power networks are designed for unidirectional power flow. Conventional power networks are designed for unidirectional power flow. Additionally, traditional power networks lack the flexibility, resilience, and monitoring and control capabilities needed to effectively manage the modern energy demands. Consequently, to meet the needs of society, smart grids have replaced conventional grids. The smart grid is a complex cyber-physical system that relies on modern information and communication technologies (ICT), advanced control systems, and the electrical grid. This system is composed of two fundamental layers: the cyber layer and the physical layer. The cyber layer includes various communication, information, and control systems that enable the smart grid to collect and analyze data, monitor performance, and facilitate decision-making. The physical layer of the smart grid consists of the electrical infrastructure that provides power to homes, businesses, and industries. This layer includes transmission and distribution lines, transformers, generators, and other equipment that are crucial for the distribution and management of electricity. Smart Grids are equipped with Remote Terminal Units (RTUs), that collects and transmits field data such as smart meters and sensors to monitor the system and retrieve data, for instance active and reactive powers flows on branches and voltages and voltage angles of buses, using ICT. Along with the benefits of smart grid structure and ICT, it may also cause issues on the grid such as cyber security and system security. Smart electrical power systems are encountered with new challenges: cyber security of the smart grids. State estimation is a critical process that ensures the secure and reliable operation of power systems by determining the system's operating state based on available measurements. However, recent research has shown that this process can be susceptible to False Data Injection Attacks (FDIAs), where attack vectors are injected into compromised measurements to bypass bad data detection methods. With the increasing penetration of distributed energy resources (DERs), the traditional state estimation process has become more vulnerable to cyber-attacks, exacerbating the risk of successful FDIAs. In this thesis, we first identify the available measurements and perform state estimation based on the identified measurements. We assume that an attacker targeting the grid compromises the system and gains access to the measurements and data used in state estimation. We also assume that the compromised data is used to launch a False Data Injection Attack (FDIA) on the measurements of the power system. In this thesis, Deep Learning-based method for detecting cyber-attacks in power systems with a high penetration rate of DERs is examined. The proposed method aims to detect anomalies in measurements, that are used in state estimation, with high detection rate. The proposed approach is evaluated implementing historical hourly load data from the New York Independent System Operator (NYISO) to three IEEE systems; 14 Bus, 30 Bus and 57 Bus. To test the effectiveness of the proposed method, four different system configurations with varying levels of DER penetration were used. To reflect the real-life conditions to the work, proposed method's performance also examined under different noise levels. Proposed Deep Learning-based method's performance is compared with widely-used classification algorithms, k-Nearest Neighbor (KNN) and Logistic Regression (LR). The results of the study indicates that LR had a higher attack detection rate than k-NN at low noise levels in the IEEE 14 Bus system. However, it was observed that the Deep Learning-based Deep Neural Networks (DNNs) was more accurate than both algorithms at both high and low noise levels. In the 30 Bus system, which has medium complexity among the introduced systems, it was observed that the k-NN algorithm detected more attacks than the LR algorithm at both low and high noise levels. Similar to the 14 Bus system, it was also observed that the DNN algorithm had a higher ability to detect attacks than both classification algorithms. DNN algorithm performed the highest attack detection ability at different noise levels in 57 Bus system. However, it was observed that the attack detection rate dropped to as low as 90% at high noise levels. The studies have shown that DNNs perform well even in the presence of noisy measurements against False Data Injection Attacks. Although the results are satisfactory, it is possible to achieve higher attack detection rates and performances by using configurations that include different hidden layers, optimizers, loss functions, or completely different algorithms.