CMTF-Net: A Complex-Valued Multi-Scale Time–Frequency Cross-Domain Attention Network for MIMO CSI Prediction

With the widespread adoption of multiple-input–multiple-output (MIMO) technology, channel state information (CSI) prediction has become a crucial technique for enhancing the performance of wireless communication systems. Traditional channel prediction methods face performance bottlenecks under high-speed mobility and complex channel conditions, making it difficult to meet the requirements of modern communication systems. To address this issue, this paper proposes a fully complex-valued cross-domain modeling framework, termed a complex-valued multi-scale transformer with time–frequency cross-attention network (CMTF-Net), for MIMO CSI prediction. CMTF-Net integrates a learnable multi-scale short-time Fourier transform (LMS-STFT), complex-valued multi-head self-attention (C-MHSA), and bidirectional cross-domain attention for complex-valued sequences (BCDA-CVS). These modules are designed to preserve amplitude–phase consistency, adapt time–frequency representations to CSI evolution, and enable information interaction between temporal and spectral features. On the simulated Overall Test set, CMTF-Net achieves the lowest MAE of 0.000032 and the highest Corr. (ρ) of 0.8230 among the compared methods, while maintaining competitive SE and BER values of 0.4240 and 0.2411 at SNR = 10 dB. On the DICHASUS measured datasets, CMTF-Net also shows favorable Test-ID and Test-OOD performance. For example, on DICHASUS-2186, it obtains Corr. (ρ)/SE/BER values of 0.8367/0.4935/0.2243 on Test-ID and 0.8061/0.4697/0.2351 on Test-OOD. These results indicate that CMTF-Net provides a balanced performance profile across prediction accuracy, spatial alignment, and communication-oriented evaluation.