An unsupervised false data detection method based on graph autoencoder and attention network in power grid

False Data Injection Attacks (FDIAs) can bypass traditional state estimation detection, which threatens the security of power systems. Data-driven detection is an effective method for detecting FDIAs. However, for supervised detection methods, it is difficult to obtain a large number of anomaly labels in the power system. To address the issue of insufficient anomaly labels, this paper proposes an unsupervised FDIAs detection method based on graph autoencoder and graph attention neural network (GAE–GAT). In the method, the GAE aggregates the characteristics of topology and operation data in power system. To improve node representation, attention mechanism is adopted to aggregate adjacent node weights. This method is tested on the IEEE-14 and 118 bus systems for three scenarios: static topology, dynamic topology, and renewable energy integration. The results demonstrate that compared with previous unsupervised detection methods, the proposed method can greatly improve detection accuracy.