A graph attention network-based framework for dynamic topology change-aware FDIA detection in smart grids

False data injection attacks (FDIAs) pose persistent threats to the security of modern power systems by compromising data integrity through the manipulation of measurements. While data-driven detection models can effectively identify these attacks under stable grid topologies, ensuring accurate results, their performance significantly deteriorates when the grid topology changes due to events like fault restoration, routine maintenance, or power flow redistribution. This lack of adaptability in traditional data-driven methods leads to a substantial decline in detection accuracy in such dynamic environments. Addressing this challenge, this study proposes a dynamic topology change-aware FDIA detection method based on graph attention networks, named the AST-TGT model. The model incorporates two key sub-modules: an attention mechanism module and a temporal convolutional module, to address the spatial interference effects arising from dynamic topological changes. First, a spatio-temporal attention mechanism was introduced to enhance the representational capacity of network nodes by dynamically assigning attention weights, thereby improving the model's understanding of both the topological structure and measurement data. Subsequently, a spatio-temporal convolutional block, composed of temporal convolutional layers and graph attention layers, was proposed to capture the joint temporal and spatial dependencies of the power grid topology and dynamic operational data. Extensive experiments conducted on the IEEE 14-bus and IEEE 118-bus standard test systems demonstrate that, compared to other data-driven detection models, the proposed model effectively improves the accuracy and stability of FDIA detection in the face of grid topology changes and exhibits high robustness in complex interference environments.