Channel Prediction Based on Transformer Architecture in High Dynamic MIMO: Mitigating Noise Effect and Eliminating Error Accumulation

In mobile millimeter-wave channels, accurate channel prediction improves the performance of wireless communication systems. However, channel state information (CSI) obtained using pilot signals is interfered by noise, and the sequential prediction of CSI leads to error accumulation, which limits the performance of the channel prediction methods. To solve the noise interference problem and the error propagation problem in channel prediction, we propose a channel predictor based on Transformer network. By using learned positional encoding instead of the traditional sinusoidal positional encoding and introducing a convolutional module to enhance the multi-head self-attention layer, the proposed method reduces the impact of noise. And the parallel prediction capability of Transformer is utilized to address the issue of error propagation. By simulation, within the acceptable computational complexity, the proposed method can achieve better prediction performance in noisy mobile scenarios compared with other neural network-based channel prediction methods. Numerical results show that on average, the normalized mean square error (NMSE) of the proposed method is 2.46 dB lower than that of the canonical Transformer-based channel prediction method under different prediction lengths.