Robust Autoencoder-based State Estimation in Power Systems

Abstract

Smart Grids are critical cyber-physical systems where monitoring is crucial, especially the process of state estimation. Since this task strongly depends on the reliability of power grid meters and their communication channels, it is vulnerable to cyber-attacks and, particularly, false data injection attacks (FDIAs), which are modifications on the meter readings that are often hard to detect. In this paper, we propose a method to construct a robust state estimator based on a variational autoencoder trained on the Fisher-Rao distance, which is a measure of dissimilarity between probability distributions. Then, we introduce a novel method to generate FDIAs that exploits knowledge of the state estimator and its learning procedure, for which we show effectiveness. Finally, numerical results and comparison with state-of-the-art methods confirm that our approach can archive similar estimation errors for clean and noisy (attacked) measurements.