Prediction of channel propagation impairments in LEO satellite networks using machine learning

Abstract

This paper presents a multi-task spatiotemporal deep learning framework designed for zero-shot geographic generalization in predictive channel modeling for low Earth orbit (LEO) satellite communications at Q-band (39 GHz). Unlike many existing methods that rely on location-specific training, the coordinate-based long short-term memory (LSTM) architecture processes historical sequences of atmospheric and geographic data spanning 60 min to simultaneously predict weather conditions and excess path loss (EPL) for 5 h, incorporating inherent Gaussian uncertainty quantification. The proposed framework is designed to improve geographic generalization, which remains a challenge for existing machine learning approaches in satellite communications. A comprehensive evaluation conducted across various European climates demonstrates exceptional dual performance, achieving a root mean square error (RMSE) of 0.026 dB at the training locations and an average error of 0.213 dB across ten entirely unseen European cities, demonstrating strong generalization and representing a substantial improvement over existing methods. The framework maintains consistent prediction performance across all untrained European cities, while the uncertainty estimation mechanism provides calibrated confidence measures that may support operational decision-making. Furthermore, the coordinate-based inference strategy enables prediction at previously unseen locations without location-specific retraining, making the proposed framework a promising approach for large-scale predictive link adaptation in future LEO satellite networks. Although the proposed framework achieved promising results on ITU-R model-generated datasets, further validation using independent real-world satellite link measurements is needed to confirm its practical applicability under operational conditions.