Reinforcement learning-based robust routing strategies against cascading failures in LEO satellite networks

Low Earth orbit (LEO) satellite networks are increasingly critical for global communications but remain vulnerable to cascading failures triggered by uneven traffic distribution and limited onboard capacity. Traditional static routing strategies are difficult to adapt to the highly dynamic topology and spatiotemporal load fluctuations inherent in LEO constellations. To address this challenge, this paper proposes a robust routing strategy based on reinforcement learning. First, we establish a realistic time-varying spatiotemporal load model that incorporates tidal traffic effects and spatial phase differences, alongside a virtual node topology mapping mechanism. Second, the routing decision process is formulated as a Markov decision process. We design a distributed Q-learning algorithm, where satellite nodes autonomously perceive local congestion states and adaptively reconfigure link weights using residual capacity and load-reciprocal priority strategies. Extensive numerical simulations demonstrate that the proposed method significantly outperforms those static baseline approaches. Specifically, it maintains the higher connectivity even under severe node failures, achieving a substantial robustness gain at the critical breakdown threshold. Furthermore, the algorithm exhibits the rapid convergence and superior cost-effectiveness, achieving high reliability at a lower normalized construction cost. Current results are conducive to understanding the robustness of LEO satellite networks, which will help to devise effective routing strategies.