Spectrum Prediction for Satellite based Spectrum-Sensing Systems Using Deep Learning

In this paper, we explore a satellite based spectrum-sensing system, where the spectrum -sensing data are used to assist spectrum sharing. However, these data may be outdated due to the long propagation delay of the satellite links. Such outdated data may incur wrong spectrum-sharing decisions, resulting in co-frequency interference. To avoid this negative effect caused by the long delay, we propose a joint long short-term memory and autoregressive moving average (LSTM-ARMA) aided spectrum-prediction scheme, where a LSTM-ARMA model is constructed and trained relying on a specially designed loss function, which can decrease prediction error by combining the LSTM and the ARMA. Furthermore, using the historical spectrum-sensing data, the well-trained LSTM-ARMA is used to predict the future spectrum occupancy in advance. The prediction performance of the proposed LSTM-ARMA is evaluated relying on an actually measured dataset captured from the Tiantong-1 satellite. Performance evaluations show that the proposed LSTM-ARMA scheme outperforms the conventional LSTM, the ARMA and the convolutional neural network and bidirectional long short-term memory schemes in terms of a lower mean absolute error (MAE). Moreover, the proposed LSTM-ARMA can simultaneously predict the spectrum situation of multiple transponders, whilst maintaining a low MAE.