Deep Learning Based Real-Time Detection of False Data Injection Attacks in Power Grids

Abstract

False data injection attack is an advanced class of modern cyber-attacks against the state estimation algorithm of the smart grid. Such attacks can inherently delude the bad data detectors at the control center and develop critical scenarios by corrupting the set of estimated states. This work furnishes an effective detection of such class of attacks with predefined bounds. The detection policy involves a robust, nonlinear deep learning approach that is capable of not only forecasting the operating states of the grid, but also can be effectively deployed by the operator to determine any attacks within the raw measurements. It is seen that such scalable models working in real-time promote a robust performance under measurement noise as well. The proposed model with its set of optimal hyper-parameters showcases a better state forecasting scheme with minimum error margin than most of the state of the art forecasting strategies. A diligent analysis on the IEEE 14 bus test system effectively promotes the aforementioned propositions.