Federated learning-based spectrum and energy efficiency enhancement in HAP-assisted LEO satellite communication

초록

The integration of low Earth orbit (LEO) satellites and high-altitude platforms (HAPs) into next-generation satellite networks enables high-speed connectivity and broad coverage. Maximizing spectrum efficiency (SE) and energy efficiency (EE) is essential for ensuring sustainable and high-throughput communication in LEO-HAP systems. Yet, the inherent dynamics of satellite movement, adjustable HAP positions, and geographically dispersed ground users (GUs) introduce substantial complexity into resource management. Addressing these challenges requires solving a complex, large-scale optimization problem involving joint control of transmit power, beamforming centers, user associations, and HAP placement. To solve this problem, we propose Fed-DAC, a federated learning-based Deep Actor-Critic algorithm that enables decentralized parallel training, thereby improving scalability and reducing computational overhead, to maximize the minimum SE and EE across all GUs. Simulation results show that Fed-DAC achieves approximately 50 % faster convergence than single-agent DAC and exhibits 20 % greater convergence stability compared to FedProx-DAC, while preserving optimality. Evaluations conducted across nine major cities in South Korea confirm Fed-DAC's significant performance over conventional iterative and static baseline methods. © 2025 Elsevier Ltd

키워드