Synergistic Power Optimization in Satellite-Terrestrial Coexistence Networks Via Attention-Driven Deep Reinforcement Learning

Satellite communication with low Earth orbit (LEO) satellites is a key technology for non-terrestrial networks (NTN) in 6G. However, LEO satellites face challenges such as extensive coverage and interference with terrestrial networks (TN), particularly in NTN-TN coexistence scenarios with ambiguous geographical boundaries. This paper proposes an attentiondriven multi-agent deep Q-network (ATT-MADQN) for dynamic LEO cell transmission power control. By incorporating an attention mechanism optimized for extracting key information, the proposed method enhances learning efficiency and interference management within the MADQN framework. Simulation results demonstrate that ATT-MADQN improves system throughput and coverage probability (CP) while outperforming conventional MADQN in terms of robustness and efficiency.