ATS-CP: An Adaptive Framework for Traffic State Collection and Prediction in Software-Defined LEO Satellite Networks

Low Earth Orbit (LEO) satellite networks have emerged as a major advancement in satellite communications, enabling high-speed, globally distributed connectivity. However, due to the limited coverage of individual satellites, achieving global service requires increasingly large-scale constellations. This expansion calls for more agile and stable management schemes. Software-defined architectures, offering centralized control and real-time coordination, provide a promising foundation. Nevertheless, large constellations introduce substantial control overhead, i.e., excessive signaling and data exchange for status reporting. To this end, we propose an adaptive traffic state collection and prediction (ATS-CP) framework to reduce control overhead without compromising traffic forecasting performance. Specifically, ATS-CP is designed as a unified framework that integrates a topology-aware traffic collection mechanism to adaptively select the most informative data and reduce control overhead, a structure-aware matrix completion module tailored for large-scale LEO satellite networks, and a residual spatio-temporal predictor for accurate forecasting. These components are jointly optimized to capture their intrinsic dependencies and effectively adapt to the dynamic topology of large-scale LEO constellations. Extensive experiments show that ATS-CP effectively reduces data collection while maintaining stable prediction performance and robustness, even under stringent data collection constraints. Notably, when reducing the collection rate by 50%, the RMSE increases by less than 0.02.