Selective Traffic State Collection and Prediction Scheme for Software-Defined LEO Satellite Networks

Software-defined Low Earth Orbit (LEO) satellite networks enable efficient and centralized inter-satellite communication management by leveraging real-time network data, supporting high-speed, global communication services. However, due to the necessity of deploying extensive constellations in LEO systems to achieve global geographical coverage, massive data transmission and processing may lead to a shortage in communication throughput and computational resources. This paper proposes a Selective Traffic State Collection and Prediction (STS-CP) scheme to reduce ISL data collection in LEO satellite systems. We first introduce a metric to quantify ISL importance, allowing data collection to focus on the most essential links. To address data incompleteness, we propose a matrix completion module based on a GAT-based autoencoder, which summarizes local dependencies and global patterns to effectively infer uncollected traffic states. A spatio-temporal residual network is proposed for traffic prediction, capable of comprehensively capturing both spatial and temporal correlations. Experimental results validate that the proposed scheme can reduce data collection while preserving the accuracy and robustness of traffic prediction.