Experimental Results on the Impact of Memory in Neural Networks for Spectrum Prediction in Land Mobile Radio Bands

Abstract

Aim: We set out to investigate the benefit of the “memory” of long short term memory (LSTM) networks in predicting spectrum occupancy in multiple time horizons in Land Mobile Radio (LMR) bands. Background: ANNs are a popular choice for spectrum prediction. Traditionally, ARIMA models have been at the forefront of forecasting. However, recurrent ANNs have demonstrated good prediction performance. Methodology: We train and evaluate four prediction models: a baseline which simply delays the time series, a seasonal ARIMA model, a TDNN and an LSTM. We test their performance on an hourly dataset in LMR bands collected in Ottawa, Canada between the dates of October 2016 and April 2017. Results: We find that LSTMs provide an improvement in prediction performance compared to the other models. We also compare the computational complexity of our models. Conclusions: The LSTM networks that remember long term dependencies and designed to work with time series provide an improvement accurately predicting spectrum occupancy in LMR bands.