Self-similar traffic prediction for LEO satellite networks based on LSTM

Abstract

Traffic prediction serves as a critical foundation for traffic balancing and resource management in Low Earth Orbit (LEO) satellite networks, ultimately enhancing the efficiency of data transmission. The self-similarity of traffic sequences stands as a key indicator for accurate traffic prediction. In this article, the self-similarity of satellite traffic data was first analyzed, followed by the construction of a satellite traffic prediction model based on an improved Long Short-Term Memory (LSTM). An early stopping mechanism was incorporated to prevent overfitting during the model training process. Subsequently, the Diebold-Mariano (DM) test method was applied to assess the significance of the prediction effect between the proposed model and the comparison model. The experimental results demonstrated that the improved LSTM satellite traffic prediction model achieved the best prediction performance, with Root Mean Squared Error values of 18.351 and 8.828 on the two traffic datasets, respectively. Furthermore, a significant difference was observed in the DM test compared to the other models, providing a solid basis for subsequent satellite traffic planning.