A Lightweight Channel Prediction Network for UAV-LEO Satellite Communications

Abstract

A lightweight channel prediction network is proposed for millimeter wave low Earth orbit (LEO) satellite communications supported by unmanned aerial vehicles (UAVs). To address the unique challenges posed by the high mobility and significant delays characteristic of UAV-LEO channels, we investigate the temporal-spatial prediction capacity of linear models composed of low-complexity multilayer perceptrons. Specifically, in the proposed channel prediction network, the channel mixing operations are carried out across the temporal and spatial dimensions of historical UAV-LEO channels, adeptly extracting global temporal-spatial correlations. Numerical results demonstrate that compared to the state-of-art channel prediction approaches, the proposed network not only provides precise channel predictions but also maintains low training overhead.