3.5 GHz Outdoor Radio Signal Strength Prediction With Machine Learning Based on Low-Cost Geographic Features

Abstract

In the fifth-generation (5G) and beyond 5G (B5G) networks, received signal strength (RSS) prediction has always been a fundamental task in network planning and optimization phases. The existing machine learning (ML)-based RSS prediction models seldom testify to the model validity with reliable drive test data in 5G networks and tend to neglect the low-cost extraction in feature contriving. This article presents an ML scheme based on the geographic feature to predict RSS. We elaborately select four features closely related to RSS from the easily acquired geographic dataset and design low-cost methods for computing them. Experiments are executed in the large-scale outdoor scenario at 3.5 GHz for a 5G network where the real RSS data are collected by the field measurement in the urban city. We adopt four state-of-the-art ML algorithms for the proposed scheme and compare the algorithm accuracy with a Stanford University interim (SUI) model, an ECC-33 model, and a ray tracing method. Experiment results show the validity of the extracted features. Besides, the proposed scheme performs well in its model accuracy and computational efficiency compared with the existing methods. Furthermore, it is promising to build a universal, accurate, and efficient ML-based RSS prediction scheme based on massive field data in 5G or B5G networks.