Frequency Spectrum Density Prediction Based on Sequence-to-Sequence Deep Learning Model

Frequency bandwidth becomes more and more precious owing to the explosion of wireless communications. Hence, frequency spectrum prediction technologies become important for better exploiting the idle frequency resources. However, most of them are subject to single-step prediction. In this paper, we propose a frequency spectrum density prediction method based on a Sequence-to-Sequence (S2S) neural network, which can predict the wireless frequency spectrum density in forthcoming time slots relying on historical records. In the application of monitoring a single frequency band, we focus on leveraging the time-domain correlation between subsequently received signals and accordingly, Long-Short Term Memory (LSTM) is employed as a basic unit of our S2S model. In the more complicate Application of monitoring multi-bands, we attempt to jointly leverage both the frequency and time-domain correlations among signals received at different time-slots and different bands. Accordingly, convolution product based LSTM (ConvLSTM) is employed as the new building block of our S2S learning model. Furthermore, We discuss design guidelines for the hyperparameters of our S2S model. The simulation results demonstrate that our S2S learning model outperforms the classical auto-regressive moving average (ARIMA) model. Particularly, in order to verify the generality of our model, both the GSM-900 and WiFi scenarios are engaged in our comparison.