LLM-Enhanced Transformer Self-Learning Framework for Dynamic Power Allocation in High-Throughput Satellite Communications

Abstract

In high-throughput satellite (HTS) communication systems, dynamic power allocation is a core part of radio resource management, especially for supporting data-intensive applications that require reliable connectivity across wide geographic areas. However, existing solutions face notable limitations: traditional heuristic algorithms (e.g., PSO, GA) are constrained by high computational complexity and inability to leverage historical experience, while deep reinforcement learning (DRL) struggles with poor generalization to unseen request patterns and insufficient adaptability to dynamic traffic. To address these challenges, we propose an enhanced end-to-end self-learning framework for HTS dynamic power allocation. The framework first adopts a Large Language Model (LLM)-based module to parse user natural language requests, automatically extract and standardize key parameters, and intelligently schedule downstream processes. It then employs a hybrid optimization architecture: a Transformer model captures spatiotemporal correlations of traffic demands to generate high-quality initial power solutions, which are further refined by PSO-GA heuristic algorithms under power constraints to minimize the mismatch rate (MR). Additionally, an online self-learning mechanism based on Low-Rank Adaptation (LoRA) uses high-quality heuristic optimization results to incrementally fine-tune the Transformer, enhancing generalization without excessive computational overhead. Extensive simulations on a GEO HTS platform show that the proposed framework outperforms baseline methods (PSO, GA, DRL) in terms of allocation precision, cross-scenario generalization, and operational efficiency, effectively adapting to the dynamic demands of HTS power allocation.