Performance Analysis of Long Short-Term Memory-based Markovian Spectrum Prediction

Dynamic Spectrum Access (DSA) solutions are widely developed to alleviate the spectrum scarcity issues commonly experienced with traditional radio frequency (RF) spectrum allocation schemes. Spectrum prediction techniques help minimize the inefficiencies inherent in DSA solutions and have been gaining traction in DSA applications. The applicability of machine learning and deep learning-based methods to make these predictions are being studied extensively. Among deep learning models, Long Short-Term Memory (LSTM) is a recurrent neural network that demonstrates high performance in various applications such as handwriting recognition, language modelling, and text prediction. LSTMs are also applied in telecommunication scenarios such as modulation classification, interference classification, and spectrum prediction. This article investigates the prediction performance of an LSTM based system model for temporal spectrum prediction. We utilize simulated Markovmodel-based spectrum data as well as spectrum measurements data to evaluate the prediction performance. Our results suggest that the predicted class scores of the LSTM based spectrum prediction model can be described using mixtures of truncated gaussian distributions. We also estimate the performance metrics using the mixture model and compare the results with the observed prediction performance over the simulated and measured datasets.

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

2021-11-25

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC

JRC125325

2169-3536 (online),   

https://ieeexplore.ieee.org/document/9605607,    https://publications.jrc.ec.europa.eu/repository/handle/JRC125325,   

10.1109/ACCESS.2021.3125725 (online),