Evaluation of Transformer Empowered Channel Prediction for 5G and Beyond Communication

Correctly predicting the channel is crucial for tackling the issue of channel aging in mobile communications, where channels change quickly over time. Current methods for predicting channels mostly rely on processing signals in sequence, meaning the channel for the upcoming frame is predicted one at a time. Consequently, as frames progress, the precision of these predictions tends to decline because errors tend to accumulate during the sequential prediction process. For preliminary research aimed at optimizing unmanned aerial vehicle (UAV) communication channels by integrating intelligent channel predictor to the UAV communication system, this paper evaluates the performance of a channel predictor based on a Transformer mechanism in comparison with traditional serial data processing methods for Rayleigh fading channels in non-line-of-sight cellular communication scenarios. The study explores the effectiveness of machine learning models in predicting channel state Information. The Transformer model, recognized for its ability to handle sequential data and mitigate prediction error propagation, is analyzed for its computational efficiency and predictive accuracy against gated recurrent unit (GRU)-based and Kalman filter-based models. The study highlights the Transformer’s superior performance in long-term prediction due to its attention mechanism, offering a promising approach for real-time communication applications.