Temporal-Frequency Domain Channel Prediction for LEO Satellite Communication System: A Novel TFformer Structure

Low Earth Orbit (LEO) satellite is one of the most promising infrastructures for realizing global high-speed interconnection services. Accurate satellite-to-ground channel state information (CSI) is crucial for meeting these demands, as it plays a vital role in ensuring the reliability and efficiency of communication links. However, the high dynamics and long delays in LEO satellite communications pose challenges to the effectiveness of the obtained CSI, resulting in channel aging issues. To this end, we propose a temporal-frequency transformer (TFformer) based channel prediction scheme to predict CSI from both temporal and frequency perspectives. Specifically, to tackle the challenge of accurately extracting time-domain features of fast time-varying satellite-to-ground channels, we propose the frequency transform block (FTB) and temporal frequency attention (TFA) modules to achieve frequency domain feature extraction and temporal-frequency feature fusion. The FTB module is proposed to extract features of CSI in the frequency domain by establishing the mapping between time and frequency domain. Meanwhile, we propose the TFA module to achieve the effective combination of frequency and temporal features by assigning distinct attention weights to different temporal-frequency components. To further address the dynamic characteristics of satellite-to-ground channels, we propose the mixture of experts (MoE) module to extract the variation trends of key parameters as representative features. Moreover, the proposed TFformer achieves linear computational complexity with negligible information loss by randomly selecting Fourier components. Numerical results show that our proposed TFformer outperforms the conventional and deep learning-based channel prediction schemes in both prediction accuracy and system-level performance.