Multi Step Temporal Spectrum Occupancy Prediction using Deep Learning

Spectrum occupancy prediction offers many advantages in Dynamic Spectrum Access (DSA) type applications and has been performed using conventional algorithms such as linear predictors, Bayesian prediction, etc. Deep learning (DL) algorithms such as Long Short Term Memory (LSTMs), Convolutional Neural Networks (CNNs), and its variants have been increasingly adopted for these applications over recent years. Moreover, these approaches can provide accurate spectrum occupancy forecasts for multiple predictive or future time steps. However, a performance analysis of different DL techniques for different input model conditions needs to be explored in the literature. To this end, we consider different DL algorithms viz., vanilla LSTM, Encoder-Decoder LSTM, and CNN models that perform multi-time step ahead spectrum occupancy prediction for Markovian input data. We analyse the performance of these models and extract performance characteristics for a range of predictive time steps and input model parameter scenarios. Our findings suggest that vanilla LSTM and 1D CNN models have the ability to perform closely with that of Bayesian techniques. These algorithms perform well in case highly correlated input data. However, there is a performance degradation when the predictive horizon width increases.

RADHAKRISHNAN Niranjana;  SITHAMPARANATHAN Kandeepan;  YU Xinghuo;  BALDINI Gianmarco; 

2025-07-30

IEEE

JRC136845

979-8-3503-8963-0 (online),   

https://ieeexplore.ieee.org/document/10815776,    https://publications.jrc.ec.europa.eu/repository/handle/JRC136845,   

10.1109/ICSPCS63175.2024.10815776 (online),