A Novel Multi-Scale Time Fusion Transformer for Long-Range Spectrum Occupancy Prediction

Radio spectrum resource prediction holds a crucial position in the field of spectrum resource management. The challenges of long-range spectrumprediction stem from the long-time dependencies, intricate nonlinear correlations, and multi-scale characteristics of electromagnetic spectrum data. This paper proposes the Multi-Scale Time Fusion Transformer (MSTFformer) architecture, which not only deeply mines the spectrum occupancy characteristics, but also designs a pyramid attention mask to effectively reduce the computational complexity of the multi-head attention mechanism. The MSTFformer consists of three components, namely the Trend Building Net (TRENDNET), Time Series Encoding Net (TENET), and Pyramid Attention Mechanism Decoding Net (PADNET). The TRENDNET is employed to capture the long-range trend features of electromagnetic spectrum data over varying time lengths. The TENET leverages BiLSTM memory units to enhance the nonlinear attributes of local temporal information and utilizes Gating Propagation (GP) and Gating Residual (GR) components to regulate information flow. The PADNET automatically constructs a multi-scale pyramid to extract correlation information within and between scales, thereby promoting the interaction of cross-scale information, and using pyramid masks to reduce model complexity. Experimental results show that compared with the other ten methods, the MSTFformer ranks third in inference speed, while surpassing the second-best method in prediction accuracy with reductions in MSE, RMSE, and MAE by 71.4%, 43.2%, and 26.6%, respectively.