Learning-based Path Loss Prediction for Manhattan Streets using 28 GHz Measurement

Large bandwidth and worldwide spectrum availability at cm/mm-wave bands have a great potential to enable high data-rate wireless transmission for 5G and beyond, and highly accurate coverage predictions are fundamental for network dimensioning and planning to serve the end-users with high quality of service. Although, many statistical, deterministic or learning-based models have been proposed for path loss (PL) prediction, developing a generalizable model remains challenging because mm/cm bands encounter complex propagation environment and significant excess loss due to the scattering and absorption from various street clutter such as foliage, trees, vehicles, and lampposts, challenging to model for PL prediction. In this work, we propose a machine learning-based PL prediction model, which is generalizable and interpretable, based on our 28 GHz measurement data from Manhattan urban streets. We utilize the point cloud and 3d mesh grid building datasets to model the street clutter and building, respectively. We first extract expert features from the point cloud information and then, we propose two feature selection (FS) policies - $l_1$-norm regression-based FS and, statistical information based FS by employing f-value, p-value, and MI, to focus on the most influential features from the designed point cloud features. We also compress the high-dimensional 3d building dataset to minimal features with autoencoders' help. Finally, these features are collectively utilized to create a linear PL prediction model. Furthermore, we introduce a new training and testing procedure to improve generalizability, such that it could help us predict PL on a never-measured street and accurately capture the street-by-street variation. We show that our proposed model gives root mean square error of 4.8 dB with a 1.0 dB standard variation that reflects street-by-street variation, as compared to the conventional slope-intercept model with 6.6 dB mean and 2.1 dB standard variation, in PL prediction.