Federated Learning for Fault and False Data Injection Attack Detection in the IEEE 9-bus System
Abstract
With the advent of the digitization of power grid systems, fault and cyber-attack detection have been a challenging problem in the field. Due to stealthy cyber-attacks and their similar effects on the grid’s voltage and frequency, it is increasingly difficult for statistical methods to distinguish between faults and cyber-attacks. While deep learning (DL)–based approaches have shown promise, their high computational requirements and scalability limitations pose significant challenges for large-scale grid systems. To this end, we propose a novel federated learning (FL) approach that decentralizes the detection procedure by enabling clients (or grid zones) to perform fault and cyber-attack detection. We consider the IEEE 9-bus system modeled in SIMULINK. Fault and false data injection (FDI) attacks are injected into the vulnerable bus for the system based on PQ-sensitivity. The results show that the proposed FL algorithm successfully distinguishes faults and cyber-attacks, achieving competitive performance compared to baseline centralized DL approaches while reducing data centralization requirements and improving scalability.