ST-GAGCN-LEO: A Spatiotemporal Graph Attention and Gated Convolutional Network for LEO Satellite Traffic Prediction

Low Earth orbit (LEO) satellite traffic prediction is critical for ensuring high communication quality, enhancing network efficiency, and optimizing early fault warning systems. To better plan LEO satellite traffic and capture spatio–temporal dependencies, we propose a prediction model based on spatio–temporal feature extraction, termed spatial–temporal graph attention and gated convolutional network (ST-GAGCN). This model integrates the graph attention network (GAT), gated temporal convolutional network (Gated TCN), and multihead attention mechanism. The GAT component processes structured data and node features within the satellite network, identifying spatial connections between nodes and generating high-level spatial representations. The Gated TCN analyzes time-series data, capturing temporal variations in traffic and extracting advanced temporal features. The multihead attention mechanism further enhances the model's ability to capture complex spatial and temporal relationships, providing a comprehensive representation. By effectively allocating attention, the model focuses on the most critical features in LEO satellite traffic prediction, thereby optimizing its performance. Simulation results demonstrate that ST-GAGCN significantly improves mean absolute error, root-mean-squared error, and mean absolute percentage error metrics, achieving superior prediction accuracy compared to existing methods.