.png){kind=icon}
.png){kind=icon}
Return
Graph World Model for Energy-Efficient Hybrid Beamforming in LEO Satellite Networks
.png){kind=icon}
PRE
AI
R
X
B
S
DOI：10.1109/tgcn.2026.3716365.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Rapid channel variations in Low Earth Orbit (LEO) satellite networks require frequent updates of the digital precoder. Conventional iterative hybrid beamforming (HBF) schemes may therefore incur substantial online computational cost, whereas learning-based alternatives may require extensive environment interactions. This paper proposes a graph world model for resource allocation (GWM-RA), a model-based learning framework for energy-efficient, surrogate-assisted non-iterative digital precoding in a two-timescale HBF architecture. Specifically, analog beam steering and scheduler-provided user association are configured on a long timescale using slowly varying statistical channel information, while the short-timescale digital precoder is generated through a single forward pass conditioned on instantaneous channel state information (CSI). Under the considered independent and identically distributed (i.i.d.) block-fading setting, the GWM-RA world model is trained using environment-generated rate and power labels to approximate the immediate action–performance relation conditioned on the current graph state and a candidate digital-precoding action. This differentiable performance surrogate provides one-step estimates of the reward and constraints for policy optimization, thereby reducing the number of true-environment evaluations required under the adopted training protocol. Numerical results show that GWM-RA achieves higher energy efficiency than the considered conventional optimization algorithms and standard model-free reinforcement learning baselines. Under the same training protocol, GWM-RA reaches the target energy-efficiency threshold with over 60% fewer environment interactions than the matched GNN-MFRL counterpart.
Keywords:
Energy efficiency
heterogeneous graph neural networks
hybrid beamforming
satellite-terrestrial integrated networks
world model
Journal
I
IF:
6.7
Papers:
1.3K
Citations:
4.3K
Organization
Cited Papers