Cross-Domain Feature-Enhanced Channel Prediction for Satellite Communications Considering Handover Effects

Abstract

Low Earth orbit (LEO) satellite networks are rapidly deployed to provide broadband coverage, but high satellite velocities and frequent handovers create highly dynamic channels that challenge reliable communications. Accurate channel state information (CSI) is essential, yet conventional estimation methods suffer from outdated measurements due to short coherence times and transmission delays. To address these challenges, channel prediction requires urgent attention. In this paper, a satellite channel prediction model (SCPM) specifically designed for LEO satellite communications is proposed. For the first time in the open literature, period-invariant information (PII) associated with satellite motion is proposed and extracted by disentangled representation learning (DRL), improving prediction accuracy during handovers. To address LEO-related Doppler and dispersion effects, SCPM captures cross-domain features and integrates such information to enable accurate prediction of future CSI. Motivated by the remarkable modeling capabilities of large language models (LLMs), a pre-trained LLM is then fine-tuned to capture long-range temporal dependencies in CSI evolution. Simulation results demonstrate that SCPM consistently achieves superior prediction accuracy compared with state-of-the-art (SOTA) models across both TDD and FDD systems, and remains robust to imperfect historical CSI. The proposed framework supports proactive beamforming, resource allocation, and scheduling, advancing the design of robust non-terrestrial networks (NTN). Besides, the model paradigm can be extended to other prediction tasks in LEO satellite communications, where frequent handovers induce abrupt variations in the data to be predicted, and PII mitigates distributional inconsistency between pre-handover and post-handover targets.