Multi-band spectrum prediction algorithm based on HGCN and simplified ReLU-GRU

The increasing scarcity of spectrum resources, coupled with rising demand, has made effective spectrum management crucial. However, the complexity and spatio-temporal variability of spectral data present significant challenges for accurate spectrum prediction. This paper proposes a novel multiband spectrum prediction model that integrates a hypergraph convolutional neural network (HGCN) with a simplified rectified linear unit-gated recurrent unit (ReLU-GRU) network which eliminate the reset gate. In this framework, the HGCN employs hypergraphs to represent spectral data, where nodes correspond to individual frequency bands and hyperedges capture multivariate relationships among them. The simplified ReLU-GRU is used to model the temporal dependencies between frequency bands, effectively fusing the extracted features for enhanced prediction performance. By replacing the traditional hyperbolic tangent (tanh) activation function with a linear rectification function (ReLU) in the state update process, the model mitigates the issue of gradient vanishing and accelerates the training process. To further improve convergence, an attention mechanism is incorporated to weight the output of hidden states. Experimental evaluation on a real-world spectral dataset from sensors in St. Gallen demonstrates that the proposed model achieves a 4.43% improvement in prediction accuracy compared to the traditional LSTM model and a 0.56% improvement over the GCN-GRU model, exhibiting superior stability. The results also show that the simplified ReLU-GRU is particularly effective in predicting highly variable data, outperforming the traditional tanh-GRU, especially in scenarios with significant fluctuations.