Spatial-temporal Traffic Prediction of Space-Integrated-Ground Networks based on Continuous Knowledge Distillation

The accurate traffic prediction serves an important role in the operation and management of Space-Integrated-Ground networks (SIGN). However, due to the limited satellite resources, coupled spatial-temporal relationship of traffic, and high dynamic changes, this problem is non-trival to solve. Therefore, this paper investigates this problem and makes the following contributions. First, to overcome satellite resource limitations, an architecture for traffic prediction based on Software-Defined Network (SDN) is proposed. The architecture substitutes models for data to reduce the interaction frequency and avoids large volumes of training data transmission to reduce bandwidth waste. Second, by utilizing the coupled spatial-temporal satellite network traffic relationship, an attention-based spatial-temporal prediction network (ASTPN) is proposed to extract the temporal/spatial correlations of the satellite traffic sequences to improve prediction performance. Third, to tackle with the variable traffic characteristics caused by the high dynamic of satellites, an adaptive knowledge distillation (AKD) training scheme is developed which is composed by continuous knowledge distillation (CKD) and learning during teaching (LDT). The former extracts common traffic historical knowledge without extensive retraining with new data, while the latter will continuously update the teacher model based on the student model. AKD continuously updates model to maintain prediction accuracy under dynamic satellite traffic by iteratively using CKD and LDT during the training process. Numerical results demonstrate the prediction accuracy can be enhanced by using the proposed ASTPN and adaptive KD training scheme. Furthermore, the amount of transmission data within SGIN during training and learning can be largely reduced compared with traditional traffic sampling transmission architecture.