Mon Feb 16, F. Flechtner
• Sorted by Date • Sorted by Last Name of First Author •
Li, Ziqiang, Liu, Wanke, Tang, Chengpan, and Zhang, Xiaozhong, 2026. Short–term high–precision prediction of LEO navigation satellite clock offset based on a hybrid FFT─LSTM model. Advances in Space Research, 77(1):981–994, doi:10.1016/j.asr.2025.09.056.
• from the NASA Astrophysics Data System • by the DOI System •
@ARTICLE{2026AdSpR..77..981L,
author = {{Li}, Ziqiang and {Liu}, Wanke and {Tang}, Chengpan and {Zhang}, Xiaozhong},
title = "{Short-term high-precision prediction of LEO navigation satellite clock offset based on a hybrid FFT─LSTM model}",
journal = {Advances in Space Research},
keywords = {LEO satellite, Clock offset prediction, Fast Fourier Transform (FFT), Long Short-Term Memory (LSTM), FFT─LSTM},
year = 2026,
month = jan,
volume = {77},
number = {1},
pages = {981-994},
abstract = "{Accurate prediction of Low-Earth Orbit (LEO) satellite clock offset is
important for enhancing real-time positioning performance in
LEO-augmented Global Navigation Satellite Systems (LeGNSS).
However, traditional models may fail to capture the complex
periodic and nonlinear characteristics of LEO navigation
satellite clock offset, which are influenced by rapid orbital
dynamics and environmental factors. To address the issues, this
study proposes a hybrid FFT─LSTM model that synergizes fast
Fourier transform (FFT) and Long Short-Term Memory (LSTM)
networks. First, FFT is used to extract dominant periodic
components of the satellite clock offset. Then, a Spectral
Analysis (SA) model that combines a quadratic polynomial and the
identified periodic terms is applied to fit and predict the
trend and main periodic terms of the clock offset. Finally, LSTM
networks is employed to model and predict the fitting residuals.
The proposed model was validated using the clock offset data
from the GRACE-C and Sentinel-6A satellites. The experimental
results show that the FFT─LSTM model achieves remarkable
prediction accuracy, with 1-hour root mean square errors (RMSE)
of 0.102 ns and 0.236 ns for GRACE-C and Sentinel-6A,
respectively. Compared with those of the linear polynomial (LP)
model, quadratic polynomial (QP) model, autoregressive
integrated moving average (ARIMA) model, gray model (GM), and
LSTM model, the prediction accuracies of the hybrid FFT─LSTM
model for 1 min, 2 min, 5 min, 10 min, 30 min, and 1 h are
improved by approximately (90.11, 89.57, 87.09, 89.70, 32.15)\%,
(78.24, 71.03, 70.85, 80.08, 11.05)\%, (73.07, 73.65, 60.09,
84.97, 8.98)\%, (85.38, 87.41, 74.97, 91.45, 42.99)\%, (94.12,
91.13, 85.72, 92.77, 75.80)\%, and (94.80, 92.45, 88.55, 94.61,
75.10)\%, respectively.}",
doi = {10.1016/j.asr.2025.09.056},
adsurl = {https://ui.adsabs.harvard.edu/abs/2026AdSpR..77..981L},
adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}
Generated by
bib2html_grace.pl
(written by Patrick Riley
modified for this page by Volker Klemann) on
Mon Feb 16, 2026 23:51:58
GRACE-FO
Mon Feb 16, F. Flechtner