Spatial-Temporal Traffic Prediction in LEO Satellite Networks via Hybrid GCN-Transformer

Low Earth Orbit (LEO) satellite networks are emerging as a critical component in future 6G networks, where accurate traffic prediction is essential for proactive scheduling, resource optimization and load balancing. However, such characteristics of satellite networks as dynamically changing topology, non-Euclidean structure, and significant spatial-temporal correlation pose severe challenges to traffic prediction. This paper proposes a GCN-Transformer scheme that integrates Graph Convolutional Network (GCN) with Transformer Encoding architecture. The GCN component captures local and global spatial dependencies, while the Transformer models long-range temporal dynamics via self-attention. This joint design is capable of processing highly coupled spatial-temporal data and maintaining stable performance under highly dynamic network topologies. Experiments on short-term and long-term prediction tasks (15-60 minutes) show that this scheme consistently outperforms other benchmark schemes in all metrics. The results demonstrate that GCN-Transformer achieves robust spatial-temporal modeling and stable prediction, offering substantial value for adaptive scheduling and resource optimization in satellite networks.