Hybrid Near/Far-Field Channel Prediction for RIS-Aided LEO Satellite Networks

Abstract

A hybrid near- and far- field cascaded channel prediction scheme is proposed for reconfigurable intelligent surface (RIS) assisted low earth orbit (LEO) satellite networks. In particular, an efficient neural network architecture, inspired by the intrinsic characteristics of wireless signals and termed the signal-informed network (SIN), is exploited to learn the precise mapping between historical uplink channels and future downlink channels. Specifically, in the proposed SIN, the time-domain autocorrelation modeling required by the channel prediction algorithm is converted into frequency-domain representation modeling, which aims to represent high-dimensional channels in terms of major frequency components. Furthermore, considering the specific non-linear phase information of hybrid-field channels, a multi-branch phase-aware module in SIN is developed to exhibit a physics-compliant channel semantic representation. Finally, a deep supervision-based encoder-decoder architecture with the auxiliary loss function is constructed as the network backbone. Simulation results demonstrate that compared to the state-of-art channel prediction models, the proposed SIN model exhibits superior channel prediction accuracy and convergence speed.