Accurate Channel Prediction Based on Transformer: Making Mobility Negligible

Abstract

Accurate channel prediction is vital to address the channel aging issue in mobile communications with fast time-varying channels. Existing channel prediction schemes are generally based on the sequential signal processing, i.e., the channel in the next frame can only be sequentially predicted. Thus, the accuracy of channel prediction rapidly degrades with the evolution of frame due to the error propagation problem in the sequential operation. To overcome this challenging problem, we propose a transformer-based parallel channel prediction scheme to predict future channels in parallel. Specifically, we first formulate the channel prediction problem as a parallel channel mapping problem, which predicts the channels in next several frames in parallel. Then, inspired by the recently proposed parallel vector mapping model named transformer, a transformer-based parallel channel prediction scheme is proposed to solve this formulated problem. Relying on the attention mechanism in machine learning, the transformer-based scheme naturally enables parallel signal processing to avoid the error propagation problem. The transformer can also adaptively assign more weights and resources to the more relevant historical channels to facilitate accurate prediction for future channels. Moreover, we propose a pilot-to-precoder (P2P) prediction scheme that incorporates the transformer-based parallel channel prediction as well as pilot-based channel estimation and precoding. In this way, the dedicated channel estimation and precoding can be avoided to reduce the signal processing complexity. Finally, simulation results verify that the proposed schemes are able to achieve a negligible sum-rate performance loss for practical 5G systems in mobile scenarios.