Machine-learning-based prediction methods for path loss and delay spread in air-to-ground millimetre-wave channels

The unmanned aerial vehicles (UAVs) have been widely applied in various fields due to their advantages like high mobility and low cost. Reliable communication is the premise to ensure the connectivity between UAV nodes. To provide reasonable references for the design, deployment, and operation of UAV communication systems, the precise prediction of radio channel parameters are required. In this study, the authors propose prediction methods for path loss and delay spread in air-to-ground millimetre-wave channels based on machine learning. Random forest and K-nearest-neighbours are the algorithms employed in the methods. Then, a feature selection scheme is proposed to further improve the prediction accuracy and generalisation performance of the machine-learning-based methods. Generally, machine learning algorithms require massive data for training purpose. However, measuring data is time-consuming and costly, especially when the scenario or frequency changes. Therefore, transfer learning methods are introduced to predict path loss with limited data. The proposed methods for path loss prediction are compared to Okumura-Hata and COST-231 Hata models. The lognormal distribution is the contrast model in delay spread prediction. Based on the data generated by ray-tracing software, the new methods have a smaller root mean square errors than contrast models.