Deep Learning–Based Channel Prediction for LEO Satellite Communications under Dynamic Propagation Conditions
Journal of High-Frequency Communication Technologies · 2026
DOI: 10.58399/kuac4786 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of High-Frequency Communication Technologies, 2026. Verfasst von Solomon Malcolm Ekolama und 1 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The rapid expansion of Low Earth Orbit (LEO) satellite constellations for 6G Non-Terrestrial Networks demands reliable short-horizon channel prediction under highly dynamic propagation conditions. LEO channels evolve under accelerated orbital motion, severe Doppler curvature, atmospheric attenuation, and ionospheric scintillation, rendering quasi-static and linear prediction models inadequate for latency-sensitive broadband services. Existing learning-based approaches improve statistical forecasting but rarely enforce orbital-kinematic consistency, incorporate synchronised environmental metadata, or validate gains at the protocol level. This paper proposes a physics-constrained, protocol-integrated deep learning framework for short-horizon forecasting of Ku-band channels. The study contributes by formalising LEO channel evolution as a bounded-acceleration stochastic process, embedding Doppler-gradient regularisation to enforce second-order temporal consistency, constructing a synchronised multimodal benchmark dataset, and demonstrating direct translation of prediction accuracy into adaptive network performance gains. The proposed hybrid architecture integrates dilated causal convolution, bidirectional Long Short-Term Memory, attention-based feature fusion, and uncertainty-aware inference. Training and validation utilise a unified dataset comprising 420 hours of measured Ku-band ground-station data and 50,000 simulated satellite passes, enriched with orbital telemetry and environmental metadata at 10 Hz. At a 200 ms prediction horizon, the model achieves a Normalised Mean Squared Error of 0.112 and an R² of 0.89, representing an 18–32% improvement over Kalman, ARIMA, GRU, and Transformer baselines. Integration with adaptive modulation and handover mechanisms yields a 12.1% increase in spectral efficiency and a 70.3% reduction in handover failures. These findings establish physics-informed, environment-aware learning as a deployable and operationally impactful solution for resilient LEO broadband communications.
Zitieren
APA
Ekolama, S. M., & Ebregbe, D. (2026). Deep Learning–Based Channel Prediction for LEO Satellite Communications under Dynamic Propagation Conditions. Journal of High-Frequency Communication Technologies. https://doi.org/10.58399/kuac4786
Vancouver
Ekolama SM, Ebregbe D. Deep Learning–Based Channel Prediction for LEO Satellite Communications under Dynamic Propagation Conditions. Journal of High-Frequency Communication Technologies. 2026. doi:10.58399/kuac4786
BibTeX
@article{Ekolama2026,
author = {Ekolama, Solomon Malcolm and Ebregbe, David},
title = {Deep Learning–Based Channel Prediction for LEO Satellite Communications under Dynamic Propagation Conditions},
journal = {Journal of High-Frequency Communication Technologies},
year = {2026},
doi = {10.58399/kuac4786}
}
EUVIMED