Channel Prediction via Hybrid-Attention Parallel Synchronous Transformer
Shufeng Lu, Yongyan Li, Baoxin Su, Xu Su, Minglei You · IEICE Transactions on Communications · 2025
To address the sum-rate degradation caused by channel aging, accurate prediction of future channels based on pilot signals is critical. Existing model-based or neural network-based approaches relying on sequential prediction suffer from error propagation and significant sum-rate loss in multi-frame channel prediction. While Transformer models have recently enabled parallel processing through encoder self-attention mechanisms, their application in massive MIMO systems remains underexplored. Besides, considering the limitations of conventional self-attention and positional encoding strategies, We propose a Parallel Synchronous Transformer with Hybrid Attention Mechanism incorporating hybrid positional encoding. Firstly, the Hybrid Attention Mechanism with hybrid position encoding ensures high adaptability and efficient attention calculation. Moreover, this method with Hybrid Attention Mechanism synchronously predicts several frames in future channels by historical channel states information (CSI), effectively eliminating error propagation and achieving near-optimal sum-rate performance. The simulation results demonstrate that the proposed model achieves a prediction normalized mean square error (NMSE) below −15 dB in two distinct user mobility scenarios. Notably, the system's achievable sum-rate closely approaches that of the ideal CSI case. These findings highlight the model's exceptional accuracy and its ability to maintain a high sum-rate, even under high-mobility conditions.