Spatial-Temporal Transformer for FDD Downlink Channel Prediction in High-Mobility Scenarios

Frequency division duplex (FDD) systems are widely adopted in high-mobility communications due to the advantages of extensive coverage and interference resilience. However, accurate downlink channel state information (CSI) acquisition in FDD systems faces two key challenges: non-reciprocal uplink-downlink frequency responses and time-variant channel aging. To address these coupled challenges, we propose a spatial-temporal Transformer (STFormer) framework for joint time-frequency domain channel prediction. First, we establish a cross-time-frequency mapping relationship that mathematically connects historical uplink channels to future downlink channels. Building upon this foundation, we design a decoupled spatial-temporal Transformer architecture with dual attention mechanisms that independently capture temporal channel evolution and spatial dependency patterns. To address the dispersive nature of high-mobility channels, we further integrate adaptive feature fusion that considers channel characteristics in the delay- Doppler domain. Furthermore, in order to bridge the gap between time-frequency channel data and Transformer-compatible representations, we introduce a dimension-aware embedding module and a linear decoder for input and output processing, respectively. Experimental validation utilizes a ray-tracing-derived channel dataset constructed for high-speed railways in a real-world geographic environment. Comparative analysis demonstrates that the proposed STFormer surpasses conventional deep learning approaches, achieving state-of-the-art prediction accuracy and approaching the theoretical upper bounds of communication performance metrics.