Abstract:

Accurately predicting the traffic between satellites and ground users is crucial for stable operation of the satellite in response to changes of user demand and environment， thus maintaining efficient performance of the satellite communication system. Traditional machine learning methods require ample historical sample data to predict the traffic of target satellites. However， during major holidays， large-scale population movements can create tidal phenomena in traffic between different satellite service areas. Traditional methods struggle to address this issue due to the absence of historical tidal data specific to the target satellites. To address this issue， this paper presents a satellite traffic migration prediction method based on spatiotemporal correlation. The method initially identifies spatially adjacent satellites using angular conditions and quantifies the spatiotemporal correlation of traffic between different satellites by calculating the time-lagged Pearson correlation coefficient， thereby constructing a source satellite traffic dataset. Subsequently， a common pattern learner is used to dynamically adjust the transfer weights of traffic samples in the source domain dataset. A weighted transfer long short-term memory network is then employed to optimize the transfer learning process， enhancing the model’s ability to predict traffic for the target satellite. Simulations demonstrate that the proposed method can effectively forecast dynamic changes in target satellite traffic.

Key words: satellite traffic prediction, transfer learning, spatiotemporal correlation, LSTM, traffic tides