Graph Neural Network-Based Cybersecurity of Smart Grids
Loading...
Date
2022-08-17
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Power grids are currently evolving into modern smart grids by integrating Information and Communication Technologies (ICT) into large-scale power networks to generate, transmit, and distribute electricity more efficiently. In such networks, physical measurements are acquired by Remote Terminal Units (RTUs) and Phasor Measurement Units (PMUs) and delivered to the Supervisory Control and Data Acquisition Systems (SCADAs). Power system operators then leverage these measurements to take necessary actions. Although the integration of ICT networks into the power systems greatly enhances the power grid operations and control, it may jeopardize the power system by creating new cyberattack surfaces for intruders. False data injection attacks (FDIAs), for instance, represent a major class of cyberattacks in which an intruder misleads the grid operators to take erroneous actions. If the Power System State Estimation (PSSE) module converges to a false operating point due to the injected false data, actions taken by the grid operator based on the false operating point can lead to serious physical consequences including systematic issues, failures, and even blackouts. In addition, the accuracy of power system analysis tools such as energy management, contingency and reliability analysis, load and price forecasting, and economic dispatch depends on these measurements. Therefore, power system operation strongly depends on the accuracy of measurements and the integrity of their flow through the system. Re-phrased differently, metering devices represent highly attractive targets for adversaries who try to obstruct the grid operation by corrupting the measurements. Moreover, if an intruder has sufficient information about the grid and their falsified injected measurements satisfy the power flow equations, they can bypass the residue-based traditional Bad Data Detection (BDD) modules and create an unobservable (stealth) FDIA. Therefore, it is essential to design powerful defense mechanisms to protect the power systems. Due to the superiority of machine learning (ML) methods along with the increasing volume of collected historical data samples, ML-based models have been proposed for the security of the smart grids against FDIAs in recent years. Despite their effectiveness, ML-based methods may overfit and fail to detect FDIAs especially in situations when the ML architecture does not capture the underlying physical system generating the data. As an example, Convolutional Neural Networks (CNNs) are well-suited to image and video processing since locality of pixels is well modeled by the sliding kernels. Conversely, a Recursive Neural Network (RNN) architecture might be more applicable to model recurrent relations such as those encountered in sequence to sequence language modeling and machine translation applications. Undirected graphs can be used to capture the smart grid topology; buses and branches of the grid can be represented by nodes and edges of the undirected graph, respectively. The Graph Neural Network (GNN) architecture, in particular, immensely benefits from this architectural matching promise. Besides, the prediction of the filter weights in GNNs instead of being performed manually can be executed automatically via Graphical Signal Processing (GSP) techniques which makes GNNs more attractive to smart grid applications. Due to GNN’s highly efficient modeling capability of non-Euclidean data structures, they are adopted in numerous areas such as social networks, physical systems, traffic networks, and molecule interaction networks. Despite their potential, to the best of our knowledge, no work has explored GNNs to protect the smart grids against FDIAs. Therefore, this thesis proposes GNN based, scalable, and real-time models to detect, localize, and mitigate the FDIAs. The proposed defense mechanisms efficiently combine model and data-driven approaches by incorporating the inherent physical connections of modern AC power grids and by exploiting the spatial correlations of the measurements.
Designing, analyzing, and simulating smart grid infrastructures as well as predicting the impact of power network failures and cyberattacks strongly depend on the topologies of the underlying power network and communication system. Despite the substantial impact that the communication systems bring to smart grid operation, the topology of communication systems employed in smart grids was less studied. The power community lacks realistic generative communication system models that can be calibrated to match real-world data. To address this issue, this work also proposes a framework to generate the underlying topological graphs for the communication systems deployed in smart grids by mimicking the topology of real-world smart grids. In this regard, the Chung-Lu algorithm is updated to guarantee the communication network connectivity and to match the degree distribution of a real-world smart grid rather than following an expected degree distribution. In addition, key characteristics of communication systems such as diameter, average shortest paths, clustering coefficients, assortativity, and spectral gap were taken into consideration to generate the most similar real-world communication network for smart grid studies. The proposed algorithm to generate realistic cyber graphs for smart grid studies will benefit the power community.
Cascading failures in power systems can be triggered by a small perturbation that leads to a sequence of failures spreading through the system. The interconnection between different components in a power system causes failures to easily propagate across the system. The situation gets worse by considering the interconnection between cyber and physical layers in power systems. A plethora of studies have addressed cascading failures in power systems especially in terms of assessing their impact on the system. Understanding how failures propagate into the system in time and space can help the system operator to take preventive actions and upgrade the system accordingly. Due to the nonlinearity of the power flow equations as well as the engineering constraints in the power system, it is essential to understand the spatio-temporal failure propagation in cyber-physical power systems (CPPS). This work hence also proposes an asynchronous algorithm for investigating failure propagation in CPPS, where the physics of the power system is addressed by the full AC power flow equations. Various practical constraints including load shedding, load-generation balance, and island operation are considered to address practical constraints in power system operation. The propagation of various random initial attacks of different sizes is analyzed and visualized to elaborate on the applicability of the proposed approach. The proposed model can shed light on the cascading failure evolution in CPPS.
Description
Keywords
Smart Grid, Machine Learning, Graph Neural Networks, Cascading Failure, Cybersecurity, False Data Injection Attacks, Cyber-Physical Power Systems, Spatio-Temporal.