Carbon-Aware Traffic Steering for TN-NTN Scenario via Graph Attention Reinforcement Learning

Abstract

Sixth-generation (6G) wireless systems envision seamless coexistence between terrestrial networks (TNs) and non-terrestrial networks (NTNs), while the carbon footprint of dense radio access infrastructure has become a critical concern. This paper proposes carbon-aware traffic steering (CATS), a Near-RT RIC xApp for the open RAN (O-RAN) architecture. CATS introduces a network-wide virtual carbon queue into the global observation of a learning-based steering policy. The policy is implemented using a graph attention network (GAT) with type-aware attention to capture heterogeneous user equipment (UE) services and is trained via proximal policy optimization (PPO). Simulation results show that CATS significantly reduces net carbon emissions while preserving quality-of-service (QoS) satisfaction.