Three-stages concatenated Machine Learning model for SFN prediction

Abstract

The single frequency network (SFN) has been assumed worldwide by telecommunication operators to save radio frequency resources and homogenize the network. Its applications have transcended the digital terrestrial television and digital radio to become part of the key techniques of the broadband and broadcast convergence for LTE-A, 5G and beyond. However, the transition from a multi frequency network (MFN) to an SFN involves multiple measurement campaigns and tuning of the network to achieve the expected up-performance and quality of service. This paper aims to propose a machine learning model to predict the SFN performance from the legacy MFN parameters. The model is based on regression and classification machine learning algorithms concatenated in three consecutive stages to predict SFN electric-field strength, modulation error ratio and gain. The training and test processes are performed with a dataset of 389 samples from an SFN/MFN trial in Ghent, Belgium. The best performance is obtained with concatenating gradient boosting, random forest, and linear regression, which allows predicting the SFN electric-field strength with an R2 of 92%, the modulation error ratio with 95%, and SFN gain with 87% from only MFN and position data. Besides, the model allows classifying the data points according to positive or negative SFN gain with an accuracy of 93%.