Dual-attention-based spatio-temporal detection model against false data injection attacks in smart grids

• A dual-attention spatial-temporal detection model is proposed to leverage topology-aware node features for FDIA detection. • GAT can adaptively adjust inter-node connection weights via attention mechanisms with capturing of spatial dependencies. • TCN can mitigate gradient explosion issue inherent in recurrent architectures, offering long-sequence modeling capabilities. • Comprehensive simulations demonstrate the proposed method’s superiority over state-of-the-art approaches. As the next-generation power grid, the Smart Grid integrates digitization and intelligence into the system. However, its advanced intelligence also introduces potential cybersecurity risks. In particular, False Data Injection Attacks (FDIAs) exploit the Smart Grid’s heavy reliance on real-time data. By spoofing sensor measurements or manipulating network communications, these attacks inject malicious information into the grid to disrupt the topological structure of the power system. To address this challenge, this paper proposes a Dual-Attention Spatial-Temporal Fusion Network (DASTFN) for FDIA detection in Smart Grids, leveraging Graph Attention Networks (GAT) and Temporal Convolutional Networks (TCN). Given the complexity of network topology, DASTFN introduces a dual-attention mechanism that assigns varying importance to each node and dynamically aggregates information from neighboring nodes by learning adaptive weights. This enables more precise capture of mutual influences and relationships between nodes. Meanwhile, TCN automatically extracts temporal features from grid data through its causal dilated convolution property, effectively detecting attacks that leave temporal traces in sequential data. The proposed DASTFN model comprehensively captures spatial–temporal characteristics of grid data, enhancing both detection accuracy and computational efficiency. Simulations on IEEE 14-bus and 118-bus systems demonstrate its superiority over existing methods in terms of ablation analysis, detection performance, and robustness under adversarial conditions.