Adaptive-A-GCRNN: Enhancing Real-Time Multi-Band Spectrum Prediction Through Attention-Based Spatial-Temporal Modeling

Multi-band spectrum prediction is a crucial task for spectrum management and improving spectrum utilization. Despite the complexity and spatiotemporal variability of spectrum data, which make accurate prediction challenging, leveraging spatial and temporal features of spectra can significantly enhance prediction accuracy. In this paper, we propose a hybrid deep learning model for multi-band spectrum prediction. Our model incorporates an Adaptive-GCN approach to learn spatial dependencies among spectra, as well as a GRU to extract temporal features. Additionally, we employ an attention mechanism at the output of the GRU to enhance the model's ability to capture long-term temporal dependencies. The proposed model's adjacency matrix is learnable, enabling an adaptive graph model without requiring the entire training dataset. This not only improves the model's adaptability but also allows for near real-time applications. We compared our model with the A-GCRNN on a real-world spectrum dataset, and the results demonstrated improved accuracy and flexibility of the proposed model.