Multi-step Traffic Prediction Based on Causal Structure Learning and Improved Seq2Seq for LEO Satellite Networks

The rapid movement of low Earth orbit (LEO) satellites leads to dynamic changes in their ground access traffic. Multi-step traffic prediction can sense long-term changes in traffic, and accurate traffic prediction is essential for analyzing satellite traffic loads. In this paper, we propose a multi-step traffic prediction (MTP-CS) model based on causal structure learning and improved sequence-to-sequence (Seq2Seq) to improve traffic prediction accuracy by fully exploiting the spatio-temporal relationship of satellite traffic. Specifically, we introduce causal structure learning to model the spatial relationship between satellite traffic as a causal graph and utilize graph attention network (GAT) to aggregate spatial features. Then, we propose an improved Seq2Seq (Seq2Seq-DCCN) by constructing its encoder using a dilated causal convolutional network (DCCN), which can capture the long-term temporal relationships of traffic sequences. Experiments on real traffic datasets show that the proposed MTPCS model achieves 16.59% to 21.9% error reduction compared to the baseline model regarding root mean square error (RMSE) metric at different prediction steps.