Attention-Based Spatiotemporal Model for RTT Prediction in LEO Satellite Networks

Low Earth Orbit satellite network (LSN) is considered a key component of next-generation communication, offering wide coverage. However, their high-speed orbital motion leads to frequent handovers and delay variations, impacting real-time application performance. Round-Trip Time (RTT) serves as a crucial network performance indicator and RTT information is important for network optimization. This paper proposes an attention-based spatiotempora model (ASTM) for RTT prediction in LSNs. Leveraging the periodic handover, we designed a twolayer Transformer module that focuses on temporal patterns within individual handover cycles and long-term trends across multiple cycles, embedding handover information into position encoding, enabling the model to capture the impact of handover. In addition, as satellite trajectories significantly influence delay variations, we apply the Dynamic Graph Convolutional Network (DGCN) to analyze the satellite trajectory graph sequence, integrating features using multi-head attention mechanism for final RTT prediction. Extensive experiments show that ASTM outperforms common models, particularly in long-term predictions, with around 65 % improvement over Transformer. Furthermore, a congestion control case study showed that ASTM improved Multipath Mobility-aware QUIC (MM-QUIC) throughput by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 7. 3 8 \%}$</tex>, demonstrating its benefits in practical LSN applications.