Spectrum Usage Analysis and Prediction using LSTM Networks
Masters Thesis, University of Utah. 2020.
The tremendous growth of wireless services has created an ever-increasing demand for the radio frequency spectrum. However, most of the spectrum, especially in the sub-6 GHz frequency ranges have been allocated. Given the observation that a large part of the allocated spectrum remains unused in various locations and at different times, dynamic spectrum access technologies that allow for opportunistic use of the allocated bands when they are idle, are being developed. In this thesis, we study the spectrum usage in the frequency range of 700 MHz to 2.8 GHz at Salt Lake City, Utah. Our study indicates that several portions of these frequencies are under-utilized, with an average of only 19% usage. Furthermore, we observe that certain frequency bands demonstrate clear usage patterns, e.g., show higher utilization during the daytime compared to night-time; that can be exploited for opportunistic secondary usage of the spectrum.
We propose a spectrum prediction system using Long Short-Term Memory (LSTM) neural networks to predict the occupancy of a channel in future time slots. We further introduce an LSTM based Window Selector to find the optimal window of future forecasts that increase the utilization of the network while minimizing the interference caused by the opportunistic user. Our experiments show that the Multivariate LSTM model can be reliably used to guide the choice of the channel for the opportunistic user. The multi-step LSTM models can be used to forecast spectrum usage with approximately 96% accuracy on the frequency bands exhibiting discernible usage patterns.