Uncertainty Characterization Framework for Machine Learning Based ACM in Q‐Band Satellite Links

Adaptive coding and modulation (ACM) is a widely utilized technique in satellite communication networks to mitigate the adverse effects of rain fades, particularly in extremely high frequency (EHF) bands. As a core component of fade mitigation techniques (FMTs), ACM improves signal resilience. However, its effectiveness is constrained by reliance on non‐real‐time channel state information. To overcome this limitation, this study explores the use of deep learning (DL) algorithms to predict rain attenuation, enabling more efficient and optimal selection of modulation and coding (MODCOD) schemes within the ACM framework. We present a comparative analysis of nine DL models for rain attenuation prediction, using Q‐band measurements from the Alphasat satellite, recorded at Chilbolton Observatory, UK. Furthermore, we incorporate Monte Carlo (MC) dropout‐based uncertainty quantification (UQ) into MODCOD selection, demonstrating its advantage over conventional approaches lacking UQ. Results show that integrating UQ enhances throughput, particularly in models such as gated recurrent units (GRUs). This work proposes a practical framework for UQ‐driven adaptive MODCOD selection in DVB‐RCS2 systems.