False data injection attacks detection with modified temporal multi-graph convolutional network in smart grids

Abstract

False Data Injection Attack (FDIA) detection can prevent the tampering of important data in the smart grid. This is of great significance to the operation and control of modern power systems. Since the existing FDIA detection methods are limited by the sequential input of data in Euclidean space, they cannot accurately describe the compelling correlation between data components. Therefore, this paper proposes a novel FDIA localization detection method based on graph data modeling and graph deep learning. The proposed approach tries to disaggregate the primary data into graph structured data with graph topological relationships based on graph theory, then designs specialized networks for data with different graph topologies. Moreover, the designed multi-graph mechanism and temporal correlation layer can better fully mine the correlation features between data components, with its attribute characteristics, to construct deep learning on the specific graph topology for FDIA detection. Extensive simulation experiments and visualization show that the proposed scheme is more effective than the conventional detection model, and its overall accuracy in 14-bus, 118-bus and 300-bus systems is 98.3%, 96.4% and 95.8%. It also proves that this scheme has high robustness and generalization ability in different scenarios.