Neural Network Algorithm for Predicting Spectrum Occupancy in Cognitive Radio Systems

The active development of modern communication systems, as well as the growth of provided communication services and radio access, have led to a deficit of the radio frequency spectrum. The existence of this problem has led to an increase in demand for the development of cognitive radio systems (CRS) with dynamic spectrum access technology (DSA). A particular task in the development of such systems is the implementation of the terminal of the secondary user of the control of the beginning and end of the work of the primary user in the operating frequency range. To effectively solve this problem, it is proposed to use spectrum monitoring with an algorithm for long-term forecasting of the state of spectrum occupancy. To implement this approach, it is proposed to use neural networks, namely long short-term memory (LSTM). The purpose of this study is to evaluate the performance of a neural model with the following values for the number of hidden neurons in LSTM layers: 50, 100, and 300. A comparative evaluation of the performance of various LSTM architectures was also made. The results of the experiment allow us to choose the final neural network model, which effectively solves the problem of forecasting the state of occupancy when implementing DSA technology in CRS.