Abstract: LEO satellite networks are highly dynamic and time-varying, requiring beam scheduling to jointly account for network-state awareness, physical feasibility constraints, and online decision efficiency. Conventional heuristics rely on instantaneous measurements and provide limited look-ahead; multi-objective combinatorial optimization incurs considerable computational cost in large-scale online decision-making; and end-to-end deep-learning models are commonly trained for specific scenarios, making them prone to performance degradation when scenarios change and offering limited interpretability. To address these issues, we propose a two-stage prediction–decision beam scheduling method for LEO satellite networks. In the prediction stage, a large language model with frozen parameters serves as the temporal representation backbone. Multidimensional satellite–ground link observations and textual scene descriptions are encoded into an input sequence to jointly predict future candidate link gains and grid-level traffic demands. In the decision stage, the large language model parses natural-language control requirements into structured constraints. A lightweight decision network is trained using the assignment obtained from an offline 0–1 integer linear program as the teacher label. During online operation, the decision network scores candidate beams, and an online 0–1 integer linear program uses these scores as objective coefficients to coordinate beam assignments across grids. Simulation results show that the proposed method achieves the best performance on the main prediction metrics, including link-gain error and optimal-beam prediction accuracy, as well as on overall decision utility. The prediction stage achieves a mean absolute link-gain error of 1.816 dB, approximately 45% lower than that of the Transformer-based time-series baselines, with an optimal-beam prediction accuracy of 90.2%. Ablation results show that the frozen LLM backbone contributes most to prediction performance: compared with the variant without the LLM backbone, it reduces the link-gain error by 45.0% and improves optimal-beam prediction accuracy by 34.3 percentage points. In the decision stage, the overall utility reaches 0.7094, representing improvements of 16.3% and 7.5% over the genetic algorithm and Double DQN, respectively. These results indicate that the proposed method improves state prediction and beam scheduling in LEO satellite networks by exploiting the temporal representation and semantic parsing capabilities of an LLM.

Key words: Low Earth orbit satellite networks, Beam scheduling, Large language models, Integrated prediction–decision-making