Deep Learning Aided Spectrum Prediction for Satellite Communication Systems

• Published in: IEEE transactions on vehicular technology Vol. 69; no. 12; pp. 16314 - 16319
• Main Authors: Ding, Xiaojin, Feng, Lijie, Zou, Yulong, Zhang, Gengxin
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.12.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Communications systems; Deep learning; Logic gates; Low earth orbit satellites; Low earth orbits; multi-channel; Neural networks; Occupancy; Performance evaluation; Prediction models; Predictive models; Satellite communications; Satellite constellations; Satellites; Short term; spectrum fusing; Spectrum prediction; Training; Wireless communication
• ISSN: 0018-9545; 1939-9359
• DOI: 10.1109/TVT.2020.3043837

In this paper, we study a spectrum-sharing satellite system, consisting of a pair of geosynchronous earth orbit (GEO) and low earth orbit (LEO) constellation, where the LEO satellites perform wireless transmissions relying on the shared spectrum of the GEO. In order to characterize spectrum-utilization law of the GEO and guide the access of the LEO users, we propose a deep learning based prediction model to generate the spectrum situation, termed as the convolutional neural network and bidirectional long short-term memory (CNN-BiLSTM). Specifically, the historical spectrum occupancy data of the GEO are first preprocessed and then sent into the proposed CNN-BiLSTM model for prediction purposes. Moreover, due to the fact that a user of the GEO always occupies multi-channel, the proposed CNN-BiLSTM model may generate multiple predicted results concurrently. Thus, a fusing network is designed to effectively combine the multiple predicted results. Additionally, the conventional long short-term memory (LSTM), the bidirectional long short-term memory (BiLSTM), the convolutional neural network and long short-term memory (CNN-LSTM) models are also considered for comparison purposes. Performance evaluations show that the proposed CNN-BiLSTM outperforms the LSTM, the BiLSTM and the CNN-LSTM models in terms of both the accuracy and the mean absolute error of the spectrum prediction.