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

Abstract This paper introduces a transformative 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 conventional methods that necessitate location-specific training, the coordinate-based Long Short-Term Memory (LSTM) architecture processes historical sequences of atmospheric and geographic data spanning 60 minutes to simultaneously predict weather conditions and Excess Path Loss (EPL) for 5 hours, incorporating inherent Gaussian uncertainty quantification. The proposed framework effectively addresses the significant limitation of geographic generalization that has been a challenge for previous 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) at trained locations and a notable average error across ten entirely unseen European cities, representing a substantial improvement over existing methods. The framework sustains an average prediction accuracy across all untrained European cities, with leading locations such as Berlin attaining high accuracy. The architecture exhibits a high degree of uncertainty calibration reliability and facilitates real-time deployment without the need for location-specific retraining, thereby establishing a new paradigm for global-scale predictive link adaptation in extensive LEO constellations.