Abstract: Given the challenges of satellite Internet,including limited computational resources,dynamically time-varying links,and uneven traffic distribution,this paper proposes an self-adaptive load balancing strategy based on reinforcement learning.Leveraging the control-data plane separation architecture of software-defined satellite networks(SDSN),a two-hop regional partitioning algorithm for SDSN is designed.To address the communication quality disparities between intra-orbit and inter-orbit links,link state values(μ) and weight values(w) are introduced to quantify link performance,prioritizing intra-orbit low-latency links.Built upon the Actor-Critic deep reinforcement learning framework,the SALB-RL model employs multi-agent asynchronous training to optimize key flow selection.Traffic redistribution ratios are computed via linear programming to minimize maximum link utilization while reducing end-to-end delay.A low earth orbit(LEO) Walker constellation is constructed using STK(Systems Tool Kit),a leading system engineering software,and traffic datasets derived from dynamic network topologies are used for training and validation.Experimental results demonstrate that SALB-RL achieves over 95% network-wide load balancing performance by redistributing only 10% of critical flows.Compared with state-of-the-art satellite Internet DRL models and traditional terrestrial load balancing algorithms,SALB-RL improves average load balancing performance by about 3% while ensuring more stable delay characteristics.This work highlights that SALB-RL effectively balances load balancing efficiency and routing overhead,offering an optimal solution for intelligent management of dynamic satellite networks.

Key words: Software-defined satellite network, Load balancing strategy, Reinforcement learning, Critical flow