Airborne Handover Strategy for LEO Satellites Based on a Candidate-Selection assisted Dueling Deep Recurrent Q-Network

Low Earth orbit (LEO) in-flight connectivity (IFC) suffers from three coupled challenges: 1) partial observability caused by fast-varying blockage and channel dynamics, 2) action-space explosion due to rapidly changing visible-satellite sets, and 3) QoE degradation from rebuffering and outage-reconnect behaviors in video services. To address these issues, we propose a Candidate-Selection assisted Dueling Deep Recurrent Q-Network (CS-DRQN) handover strategy. Specifically, an LSTM-based dueling value network is used to capture temporal channel-buffer dependencies under partial observability; a candidate screening module with multi-factor scoring and azimuth diversity is introduced to compress redundant actions; and a disconnection-aware reward is designed to jointly penalize rebuffering, outage events, and unstable reconnection behavior.We evaluate five policies under shared physical-world stochastic realizations: Proposed DRQN, No-CS DRQN ablation, PPO-MLP No-CS, Baseline DQN, and an Ideal-Link DQN upper-bound reference. Experiments over nine time slices (400 episodes per slice) show that, in non-ideal channels, CS-DRQN consistently achieves the best overall performance, with near-zero outage and negligible tail rebuffering across dates. Compared with Baseline DQN, CS-DRQN yields higher reward, significantly lower P99 rebuffering, and stronger cross-date robustness. Mechanism-level results further indicate that CS-DRQN attains higher smooth-handover effectiveness with lower dependence on outage-reconnect behavior. These findings demonstrate that temporally aware decision-making and candidate-space compression are both essential for robust LEO-IFC handover.These results indicate that the proposed method improves user-level QoE (e.g., outage and tail rebuffering) while maintaining favorable QoS-related handover behavior.