AI-Driven Dynamic Relay Selection for Resilient Ground-to-GEO Optical Links: A Deep Reinforcement Learning Approach to Joint Turbulence and Zenith Angle Mitigation
Year: 2026
Type: conference-paper
Abstract: Free-space optical (FSO) communication links between ground stations and geostationary satellites (GEO) are central to Next-Generation(6G) high-capacity space networks. However FSO links are highly vulnerable to turbulence, zenithangle dependent attenuation, and pointing errors, leading to outages and inefficient spectrum use. Traditional relay schemes lack adaptability, while reactive diversity techniques causes excessive switching and instability. We propose an artificial intelligence-based cognitive architecture using Deep Reinforcement Learning (DRL) to dynamically optimize relay satellite selection and forwarding strategies for low Earth orbit (LEO) satellites via Amplify-and-Forward (AF) and Decode-and-Forward (DF). Relay selection is modeled as a Markov Decision Process (MDP), with training based on Proximal Policy Optimization (PPO) and a unified physicsbased channel definition incorporating Gamma-Gamma turbulence, ITU-R P. 1814 zenith attenuation, and Beckmann pointing errors. Simulation results show over 80% reduction in outage probability at$\theta = \mathbf{7 5}^{\circ}$, with a 15dB SNR gain compared to static relays at$\mathbf{B E R}=\mathbf{1 0}^{\mathbf{- 9}}$. Tested against 100 NOAA atmospheric profiles, our approach retains 93% of peak performance. With inference delay < 0.15 ms and 75% fewer channel switches, the DRL solution enables autonomous, resilient satellite networks for 6G Space-Air-Ground Integrated Networks (SAGIN).
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204181443