Channel State Information Prediction Using Transformer Models for High-Mobility Wireless Networks

Channel state information (CSI) prediction is pivotal for ensuring robust and efficient operation in high-mobility wireless networks, where rapidly changing channel conditions demand accurate and timely forecasts. This paper presents a novel non-autoregressive Transformer-based framework that predicts future downlink CSI from previous uplink measurements for frequency division duplex (FDD) systems. Unlike conventional autoregressive or recurrent approaches that predict sequentially, our method leverages a purely encoder-based Transformer to produce multi-step downlink predictions in parallel, reducing error propagation problem and latency. Experimental results show that our Transformer framework significantly lowers the mean squared error compared to benchmarks, highlighting the robustness and suitability of our model for next-generation networks in high-mobility scenarios.