A Spectrum Prediction Framework Based on Graph Convolutional Recurrent Neural Network Integrated with Variational Mode Decomposition

Wireless spectrum prediction is crucial for dynamic spectrum management and congestion mitigation. However, existing methods often emphasize spatio-temporal feature modeling while neglecting the frequency-domain structure of the spectrum. Moreover, spectrum data inherently exhibits non-stationarity, multi-scale characteristics, and strong spatio-temporal coupling, making it difficult for traditional prediction models to fully capture its complex dynamics, thereby limiting prediction accuracy. To address these issues, this paper proposes a spectrum prediction model that integrates Variational Mode Decomposition (VMD) with a graph convolutional recurrent neural network. First, VMD is employed to adaptively decompose the raw spectrum into multi-scale intrinsic modes, achieving frequency-domain decoupling and noise suppression, thus enhancing the model's sensitivity to spectral details. Second, the Spearman correlation coefficient is utilized to construct a graph topology among frequency bands, and a Graph Convolutional Network (GCN) is applied to extract spatial dependency features, combined with Gated Recurrent Units (GRU) to model temporal evolution patterns. Furthermore, an attention mechanism is introduced to dynamically weight the hidden states, focusing on critical information and improving training efficiency. Experiments on real-world spectrum datasets demonstrate the superior performance of the proposed model, achieving an accuracy of 96.79%, MAE of 0.3342, RMSE of 0.4632, and R² of 0.9969. Keywords: Spectrum prediction, Variational Mode Decomposition (VMD), Graph Convolutional Network (GCN), Gated Recurrent Unit (GRU), attention mechanism