DA‐MADRL: Handover Optimization for LEO Satellite Networks via Domain‐Assisted Multi‐Agent Deep Reinforcement Learning

ABSTRACT Frequent handovers in Low Earth Orbit (LEO) satellite networks are inevitable due to fast satellite movement and highly dynamic topology, which poses significant challenges to service continuity, link quality maintenance, load balancing, and handover stability. To address these issues, this paper proposes a domain‐assisted multi‐agent deep reinforcement learning method for handover optimization, namely, DA‐MADRL. DA‐MADRL employs a multi‐agent deep reinforcement learning framework under a domain‐assisted management architecture to enable autonomous handover decisions. It develops a channel model that incorporates propagation loss and satellite attitude effects, establishes a satellite and ground link characterization mechanism through attitude and channel coupling, and designs a multi‐objective reward function to jointly optimize link quality, node utilization, handover reduction, and resource allocation. Simulation results show that DA‐MADRL outperforms baseline methods in key performance indicators. In particular, it reduces the average number of handovers by 46%–57%, while maintaining competitive service quality and load‐balancing performance, and it improves load balancing performance by up to 12% over Min‐Load, especially in high‐density scenarios. Moreover, from a security‐oriented perspective, the significant reduction in handover frequency implies a smaller control‐plane attack surface, fewer signaling interactions, and improved decision stability. These results indicate that DA‐MADRL is an effective and adaptive solution for dynamic handover optimization in LEO satellite networks, with notable advantages in both performance and security‐related resilience.