Graphical Convolution Network Based Semi-Supervised Methods for Detecting PMU Data Manipulation Attacks

Abstract

With the integration of information and communications technologies (ICTs) into the power grid, electricity infrastructures are gradually transformed towards smart grid and power systems become more open to and accessible from outside networks. With ubiquitous sensors, computers and communication networks, modern power systems have become complicated cyber-physical systems. The cyber security issues and the impact of potential attacks on the smart grid have become an important issue. Among these attacks, false data injection attack (FDIA) becomes a growing concern because of its varied types and impacts. Several detection algorithms have been developed in the last few years, which were model-based, trajectory prediction-based or learning-based methods. Phasor measurement units (PMUs) and supervisory control and data acquisition (SCADA) system work together to monitor the power system operation. The unsecured devices could offer opportunities to adversaries to compromise the system. In the literature review part of this thesis, the main methods are compared considering computing accuracy and complexity. Most work about PMUs ignored the reality that the number of PMUs installed in a power system is limited to realize observability because of high installing cost. Therefore, based on observable truth of PMU and the topology structure of power system, the graph convolution network (GCN) is proposed in this thesis. The main idea is using selected features to define violated PMU, and GCN is used to classify susceptible violated nodes and normal nodes. The basic detection method is introduced at first. And then the calculation process of neural network and Fourier transform are described with more details about graph convolution network. Later, the proposed detection mechanism and algorithm are introduced. Finally, the simulation results are given and analyzed.