<p>False data injection attacks (FDIAs) pose critical cybersecurity risks to smart grids by evading conventional detection mechanisms, compromising grid operational. Current deep learning approaches to solving the FDIA localization problem for smart grids frequently ignore the global spatial dependencies, failing to model dependencies between nonadjacent nodes and consequently compromising localization precision. Additionally, the computational complexity of deep learning architectures compromises localization interpretability, diminishing the credibility of results. To address the above challenges, we propose an interpretable deep learning-based FDIA localization method that leverages the multi-head graph convolutional attention network. Unlike prior approaches focusing solely on local spatial dependencies within power grids, the proposed method integrates multi-head graph attention into Chebyshev graph convolution, simultaneously capturing both local and global spatial dependencies. Specifically, each attention head learns distinct spatial dependencies within independent subspaces, which are then dynamically aggregated to capture both global and local spatial relationships, thereby improving localization precision. Furthermore, the model offers receivable interpretations for precise localization outcomes by revealing the model’s focus on various spatial nodes during the decision-making process. Comprehensive evaluations are conducted on IEEE 14-bus and 118-bus test systems. Simulation results reveal that our method surpasses state-of-the-art approaches in localization precision, demonstrating precision improvements of 1.06% and 4.31% for the 14-bus and 118-bus systems, respectively, while providing reasonable interpretability for spatial dependencies.</p>