Multi-scale Time-Frequency Dual-channel Spectrum Prediction

The rapid growth of IoT and 5G communication has intensified spectrum demands for wireless systems. Accurate spectrum prediction enhances sensing and access capabilities in cognitive radio networks. Current research uses short-term historical data for static feature extraction. This approach fails to capture long-term variations and cross-band correlations. We propose MS-SpecFormer, a multi-scale spectrum prediction framework. It uses dual-scale temporal feature extraction. This method models both long-term occupancy trends and short-term interference patterns. Our framework includes an adaptive graph network. It detects energy leakage between bands and modulation sideband effects. We also implement a dual attention mechanism for time-frequency feature integration. Tests on real spectrum datasets show significant improvements over existing benchmarks. Our method reduces RMSE by 6.1% and MAE by 4.9% in long sequence prediction. MS-SpecFormer demonstrates stronger modeling capabilities. It provides robust support for practical spectrum management.