Minimum Round-Trip Time Prediction for Low Earth Orbit Satellite Networks

Abstract

Low Earth orbit (LEO) satellite networks, characterized by wide coverage, low latency, and high bandwidth, will be a key component of the future 6G mobile communication networks. However, the periodic handover of satellites in LEO networks will lead to dynamic path changes and fluctuations in propagation delay, affecting the end-to-end performance. Establishing a minimum round-trip time (minRTT) prediction model for LEO satellite networks is crucial for optimizing the design of related mechanisms such as routing, congestion control, and loss recovery. To this end, this paper first conducts a Starlink measurement to collect real-life data and then proposes a novel model combining hybrid attention (HA) mechanisms with multiscale convolutional neural networks (MCNN) and long short-term memory networks (LSTM) to explore minRTT prediction. The method utilizes HA to highlight important positional features in the minRTT sequence, while the MCNN and LSTM are exploited to capture the variation patterns of minRTTs, thereby enhancing the prediction accuracy. Experimental results show that the proposed HA-MCNN-LSTM model outperforms existing methods.