Hybrid Model-Aided Learning for 5G-NTN Handover in High-Mobility Platforms

Ubiquitous 5G/6G network access increasingly relies on non-terrestrial Networks (NTN), yet the mobility patterns of Low Earth Orbit (LEO) satellite constellations create significant challenges for seamless connectivity. In existing studies, handover management in NTN environments has been handled using heuristic-based approaches or deep Q-learning (DQN) models, which often lack the foresight needed to anticipate mobility changes, resulting in frequent handovers and connectivity disruptions. To address these limitations, we propose a hybrid model-aided learning framework that combines a transformer-based predictive model with reinforcement learning (RL) to manage handovers in real-time adaptively. By introducing a short prediction horizon from the transformer model before applying an advantage actor-critical (A2C) RL approach, our framework reduces handover frequency and accelerates convergence. Numerical results validate the effectiveness of this approach, showing higher rewards, higher demand satisfaction, greater stability, and enhanced efficiency compared to DQN-based methods and RL without predictive components.