Multi-Granularity Traffic Prediction for Satellite Networks Based on Dynamic Adaptive Graph Modeling

Traffic prediction plays a crucial role in the efficient operation of satellite networks. However, due to resource consumption arising from redundant training of multiple individual prediction models, the dynamic and coupled spatial-temporal relationship of traffic, and maintenance of accurate traffic proportions, this problem is non-trivial to solve. Therefore, we consider this problem and makes the following contributions. First, a multi-granularity traffic prediction framework based on a shared feature extraction is designed to jointly predict total network traffic and service-specific traffic of satellite networks. This design ensures that both global and per service predictions benefit from common representations, reduces redundant computations and lowers overall model complexity. Second, a dynamic adaptive graph with Graph Diffusion Convolution (GDC) and Gated Recurrent Units (GRUs) is proposed to extract the spatial-temporal dependency of network traffic by fusing the features of population coverage, satellite distances and historical traffic data. Third, to preserve the proportional relationship of the network traffic, the angle-based loss is employed to minimize the angle deviation between the predicted and truth traffic vectors. Meanwhile, a multi-task loss function is proposed that jointly optimizes the total traffic prediction loss, the service-level losses, and the consistency regularization term to achieve accurate multi-granularity prediction. Numerical results demonstrate that the proposed framework can reduce prediction error and improve correlation for both global and service-level predictions.