Enhanced Multi-Band Spectrum Prediction Using Singular Spectrum Analysis and Attention-Based BiLSTM

The mismatch between growing service demands and scarce spectrum resources in wireless communications has led to spectrum shortages and deteriorating electromagnetic quality. The complexity and variability of spectrum data pose challenges for accurately predicting spectrum usage. A prediction method combining singular spectrum analysis (SSA) with bidirectional long and short time memory (BiLSTM) network and attention mechanism is proposed to improve the prediction performance. The method first constructs the original time series using the SSA locus matrix. Then the subsequences representing different time series components are extracted, and the correlation analysis of the decomposed subsequences is carried out. Finally, an attention-based BiLSTM (A-BiLSTM) prediction model is used. The model predicts these subsequences and assigns weights based on their correlation coefficients to refine the predictions. Experimental results validate the effectiveness of the proposed method and show that A-BiLSTM significantly improves the prediction accuracy and overall model performance.