Zhangyao Song

Accurate channel prediction is critical for mitigating the impact of channel aging in high-mobility environments. One of the primary challenges in channel prediction lies in the rapid variations of wireless channels, which lead to significant temporal fluctuations in the data distribution and result in pronounced non-stationary characteristics. Existing methods are limited in their ability to explicitly model or address the non-stationary nature of time-varying channels. To overcome this limitation, we introduce the Diffusion-based Spatio-Temporal Channel Prediction method via Non-stationarity Decoupling (DST-ND). Specifically, the DST-ND approach incorporates a non-stationarity decoupling module based on moving average. This module effectively separates time-varying channels into their non-stationary and stationary components. Additionally, a spatio-temporal diffusion model is employed to capture both temporal and spatial dependencies in channels, utilizing the global receptive field enabled by the attention mechanism. Experimental results on both real-world measurement data and simulation data demonstrate that the DST-ND method achieves state-of-the-art prediction performance across a variety of high-mobility scenarios.