Liquid Neural Network–Enabled Starlink Satellite Ephemeris Accuracy Upgrade

This paper proposes a method to improve the ephemeris accuracy of Starlink satellites using liquid neural networks (LNN). Focusing on the orbit prediction data of 100 Starlink low Earth orbit (LEO) satellites, the study constructs an LNN-based ephemeris optimization framework, which includes data acquisition, preprocessing (time/coordinate conversion, sequence construction), model initialization, training, prediction, and performance evaluation. The LNN, inspired by biological neuronal dynamics, features a liquid layer with leaky integration, enabling it to model long-term dependencies in time-series data. Its dynamic state updates and sparse connectivity enhance the capture of temporal patterns in orbital data. Comparative experiments with random forest (RF), support vector machine (SVM), and long short-term memory (LSTM) networks show that the LNN achieves superior performance: the average prediction error is reduced by 17.9% compared to the original model (to 339 m), the median error by 35.1% (to 162 m), and 81 satellites exhibit improved accuracy, outperforming the comparison methods in overall stability and individual adaptability. Sensitivity analysis of key parameters (e.g., hidden dimension, learning rate, leak coefficient) reveals that the LNN’s performance is highly dependent on parameter tuning, with optimal ranges identified to balance expressiveness and efficiency. This research demonstrates that the LNN effectively addresses the challenges of temporal phasing, uneven error distribution, and individual variability in Starlink ephemeris data, providing a new approach for enhancing satellite orbit prediction precision.