Physics- and Learning-based Detection and Localization of False Data Injections in Automatic Generation Control

Abstract

In this paper, we propose two complementary methods of detecting False Data Injection (FDI) attacks in Automatic Generation Control (AGC). The first is a physics-based method which relies on Interaction Variables, and is derived by using a more detailed spatial dynamic model of the control area than the currently used Area Control Error (ACE). The second method of detecting FDI attacks in AGC is based on Deep Learning. This method mainly depends on historical data (tie-line flow, and frequency) and ACE data, and employs a Long Short Term Memory (LSTM) neural network to build a model using the available historical data to learn the data patterns, and then predict ACEs through the learned patterns. The performance of both methods is verified through simulations on a 5-bus power system. Our results show that both methods yield high detection accuracy. The physics-based method performs better than the learning-based method, although, at the cost of requiring significantly more noise-free measurements.