A Time-Frequency Joint Wideband CSI Prediction Model and Its Validation Using 5 G Commercial Networks

Channel state information (CSI) prediction, which can efficiently avoid the channel aging problem, will be essential for the design of next generation advanced transceiver. To predict wideband CSI, this paper investigates time-frequency joint channel prediction based on deep learning (DL) models. Firstly, single-step and multi-step ahead prediction problems for wideband CSI are constructed, and dataset generation methods are presented. Then, a novel time-frequency joint channel prediction model is proposed to predict the wideband CSI, which merges the convolutional neural network (CNN) and the encoder of Transformer with the probabilistic sparse self-attention mechanism. This model can capture the time-frequency correlation information hidden in the wideband CSI datasets and reduce computational complexity significantly compared with the traditional Transformer model. Simulated wideband CSI datasets are generated, and optimal hyper-parameters are determined by autocorrelation analysis, similarity analysis and cross-validation. The performance of the proposed model is evaluated in terms of prediction accuracy, robustness and computational complexity, and is compared with classical DL models. Finally, realistic wideband CSI datasets generated according to channel data acquisition in fifth-generation (5 G) commercial networks are used to validate the proposed model.