SEGMTM: A Spectrum Prediction Method Based on Enhanced Graph Convolution and Multi-scale Time Decomposition

The development of wireless communication technology has led to increasing pressure on spectrum resources, making the rational allocation and utilization of these resources a significant challenge both now and in the future. Although spectrum data is a complex nonlinear time series, it exhibits a high degree of temporal and spatial correlation, providing new directions for addressing the issue of spectrum resource scarcity. In response to this situation, this study constructs a multi-scale spatio-temporal spectrum prediction method based on deep learning. First, we analyze the correlations present in different channels of spectrum data and utilize singular spectrum analysis (SSA) to decompose the complex spectrum data into a series of frequency components with underlying structures and patterns. Subsequently, we propose a spectrum prediction model (SEGMTM) that includes an attention-based enhanced graph convolutional network module (A-EGCN) and a multi-scale temporal module (MTM) to model the spatial and temporal correlations of the spectrum data, respectively. Furthermore, to reduce model complexity, we design a D-Regression module for auxiliary predictions. We validate the effectiveness of the proposed method through spectrum quality prediction and spectrum state prediction on two real measured spectrum datasets. Experimental results demonstrate that the proposed method achieves outstanding performance in both prediction tasks, with particularly notable advantages in long-term prediction tasks. In the spectrum quality prediction task, evaluation metrics show an improvement of 1.72% to 21.19%, while in the spectrum state prediction task, the accuracy improves by 1.28% to 3.51%.