A Study on Simple Geometries for Modeling User Equipment Geospatial Attachment to Mobile Cells

• Published in: IEEE access Vol. 11; pp. 100097 - 100113
• Main Authors: Qiu, Danny, Samba, Alassane, Afifi, Hossam, Gourhant, Yvon
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial Intelligence; Base stations; budget link; cell coverage; Computational modeling; Computer Science; Data models; Frequencies; geometric modeling; Geometry; Machine learning; Mobile communication; Mobile computing; Mobile network; Polygons; Recall; Scaling factors; Training; Voronoi; geometric modeling; cell coverage; budget link; Mobile network cell coverage geometric modeling Voronoi budget link; Mobile network; Voronoi
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3315129

As data-driven solutions will become an important component of next-generation networks, we were faced with the difficulty of developing cross-domain datasets for training machine learning models. In order to understand how external sources of data should be associated with mobile network data at cell level, we have derived a method for splitting simple geometric coverage models for base stations to obtain coverage models for sectorized cells. Then, we developed a method to compare the coverage models with a ground truth of real measurements. We also proposed to use the notions of convex and concave hulls to quantify when a geometry is undersized or oversized. The two types of geometry that were evaluated were the Voronoi polygons and circular shapes, with sector split. All frequency bands considered, the results have shown that Voronoi sector split model combined with upscaling the coverage shapes was the closest to the ground truth, with an average recall of 0.74. Since upscaling Voronoi polygons should be a good practice to improve coverage modelling, we have also proposed an approach using more affordable data (i.e. cell level aggregates) instead of user locations to find the best scaling factor. All frequencies combined, we have observed an average increase of 0.13 points in the recall between the diagram with the default scale and the scale-tuned diagram.