Detection of False Data Injection Attacks in Power Grids Based on Spatiotemporal Feature Fusion

With the development of new power systems, fast and accurate detection of False Data Injection Attacks (FDIA) is crucial for the secure operation of power grids. Existing FDIA detection models based on spatiotemporal correlations have poor feature extraction capabilities and suffer from feature shift, facing the issue of disrupted inherent spatiotemporal correlations in measurement data. To address this, we propose a detection model based on adaptive fusion of spatiotemporal features. First, Graph Convolutional Networks (GCN) is used to extract static spatial features from the power grid topology, while Graph Attention Networks (GAT) captures the dynamic spatial features. Next, Long Short-Term Memory (LSTM) is employed to analyze the temporal variations and extract temporal features. Finally, the extracted spatiotemporal information is projected into feature space for alignment, and feature fusion is performed using a bilinear attention mechanism. The fused features are then used for FDIA detection. Experimental results show that, compared to existing detection models, the proposed model performs better in terms of detection accuracy and robustness.