Convolutional LSTM-based Long-Term Spectrum Prediction for Dynamic Spectrum Access

The concept of Dynamic Spectrum Access (DSA) with Cognitive Radio (CR) as a key enabler is considered as a promising solution to alleviate the inefficient use of the radio spectrum. Relying on the presumed knowledge of the spectrum occupancy from sensing, geo-location databases or prediction, DSA allows opportunistic users to share spectrum bands in a non-interfering manner when the bands are not in use by their respective incumbent owners. Several literatures have presented prediction algorithms in order to get meaningful data about future spectrum usage; however, most of them only exploit the spectrum data in time, space and/or frequency dimension(s) to provide a short term, i.e., single next step, prediction. In this work, we propose a novel approach with Convolutional Long Short-Term Memory (ConvLSTM) Deep Learning Neural Network for a long-term temporal prediction that is trained to learn joint spatial-spectral-temporal dependencies observed in spectrum usage. Real environment measurement data from Electrosense are used to evaluate the prediction accuracy of the proposed network for increasing future time steps and different spectrum channels. Prediction result for the next 180 minutes for UHF bands of 450-520 MHz is presented for a 4 km ² area in Spain indicating the prominent and stable prediction performance of ConvLSTM network.