Abstract
UAV, mounted aerial base stations offer remarkable deployment flexibility for 6G Non Terrestrial Networks (NTNs), yet their hovering energy demands present a critical sustainability bottleneck, particularly as network topologies continuously evolve in practice. As a first step, we develop MARL-ESA (Energy-Saving Adaptation), a Multi Agent Reinforcement Learning (MARL) framework achieving 48% energy savings in static UAV assisted heterogeneous networks (HetNets) compared with always-on operation through centralized training with decentralized execution (CTDE). However, MARL-ESA remains inherently topology specific, requiring over 1500 retraining episodes whenever UAV positions or traffic patterns change. To overcome this limitation, we propose Meta-MARL-ESA, a meta reinforcement learning extension that enables zero shot generalization across unseen network configurations. The problem is formulated as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) over a task distribution τ∼p(T), with a permutation invariant Graph Neural Network (GNN) policy trained via Model Agnostic Meta Learning (MAML). Simulation results demonstrate that Meta-MARL-ESA achieves 12× faster convergence on new topologies, preserves 82% of optimal performance in zero shot deployment, improves energy efficiency (EE) by 21.5% over converged baselines in out of distribution (OOD) scenarios, and recovers to 98% of optimal performance within 8 decision epochs during dynamic UAV repositioning events.