Deep Learning Empowered Spectrum Prediction with Wavelet Decomposition

As manifold wireless services and applications emerging, it has been a key challenge to utilize limited spectrum resource to support a extremely large number of devices with various resource demands. Currently, most of the existing model-based or deep learning (DL) based spectrum prediction scheme can tackle short-term prediction issues. However, long-term prediction issue remains a technical challenge since the complex variation patterns hidden in the long received signal strength (RSS) sequence cannot be sufficiently captured by a simple feed-forward neural network. To enhance the prediction performance, in this paper, we propose a DL-based model with wavelet decomposition, which decomposes the historical RSS sequence into low-frequency trend component and high-frequency seasonal components. Specifically, a time-delay based exponential smoothing attention long short-term memory (ESA-LSTM) network is adopted to capture the low-frequency trend component, while the correlations between time blocks consisting of multiple time steps in each of multi-scale high-frequency components are captured by the Patch time series transformer (PatchTST) network. We have conducted extensive simulations with a dataset collected on 2.4 GHz frequency band. The simulation results show that our proposed model can achieve average reductions in mean square error performance of 56.1%, 34.8% and 15.3% over the traditional auto-regressive integrated moving average (ARIMA), sequence-to-sequence based LSTM (S2S-LSTM) and PatchTST models, respectively.