Real-Time Continuous Neural Correction of Reduced- Dynamic LEO Orbit Predictions for LEO-PNT

Abstract Low Earth orbit positioning, navigation, and timing (LEO-PNT) services broadcast predicted orbits whose errors grow since the last orbit fit, mainly due to hard-to-model atmospheric drag, so a physics-only prediction soon becomes too inaccurate for broadcast. We present continuous neural correction of reduced-dynamic orbit prediction (CNC-RDOP), a deep neural network added on top of a geodetic-standard reduced-dynamic prediction, leaving the orbit determination unchanged. A gated recurrent unit network reads the last 30 minutes of prediction error, the predicted orbit minus a real-time kinematic position solution in the along-track, cross-track, and radial frame, plus the predicted position and velocity, and forecasts the error over the next 60 minutes. The forecast is refreshed at every measurement epoch, so the broadcast orbit is updated continuously rather than once per fit. The network is trained on GRACE-FO C orbits (2022-2023), validated on 2024, and tested on an independent 2025 set; the kinematic solution serves only as input, while a precise post-processed orbit provides the reference, so the reported accuracy reflects real-time operation. Under a hybrid broadcast rule, the 5-minute error stays at about 2-31 m for up to 6 hours after a fit, against a physics-only growth from about 2 to 463 m. Within the first 10 minutes the physics prediction is already about 2 m and broadcast directly; the neural correction applies beyond it. Applied to the companion satellite GRACE-FO D without retraining, the same network reproduces these results to within about 1%, so one network can serve co-orbital LEO-PNT satellites in real time.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-10119188/v1
Primary Topic
GNSS positioning and interference
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Real-Time Continuous Neural Correction of Reduced- Dynamic LEO Orbit Predictions for LEO-PNT

Amir Allahvirdi-Zadeh, Ahmed El‐Mowafy, Sheik Fattah
Research Square
GNSS positioning and interference
preprint

Real-Time Continuous Neural Correction of Reduced- Dynamic LEO Orbit Predictions for LEO-PNT

Amir Allahvirdi-Zadeh, Ahmed El‐Mowafy, Sheik Fattah
preprint en

Abstract

Abstract Low Earth orbit positioning, navigation, and timing (LEO-PNT) services broadcast predicted orbits whose errors grow since the last orbit fit, mainly due to hard-to-model atmospheric drag, so a physics-only prediction soon becomes too inaccurate for broadcast. We present continuous neural correction of reduced-dynamic orbit prediction (CNC-RDOP), a deep neural network added on top of a geodetic-standard reduced-dynamic prediction, leaving the orbit determination unchanged. A gated recurrent unit network reads the last 30 minutes of prediction error, the predicted orbit minus a real-time kinematic position solution in the along-track, cross-track, and radial frame, plus the predicted position and velocity, and forecasts the error over the next 60 minutes. The forecast is refreshed at every measurement epoch, so the broadcast orbit is updated continuously rather than once per fit. The network is trained on GRACE-FO C orbits (2022-2023), validated on 2024, and tested on an independent 2025 set; the kinematic solution serves only as input, while a precise post-processed orbit provides the reference, so the reported accuracy reflects real-time operation. Under a hybrid broadcast rule, the 5-minute error stays at about 2-31 m for up to 6 hours after a fit, against a physics-only growth from about 2 to 463 m. Within the first 10 minutes the physics prediction is already about 2 m and broadcast directly; the neural correction applies beyond it. Applied to the companion satellite GRACE-FO D without retraining, the same network reproduces these results to within about 1%, so one network can serve co-orbital LEO-PNT satellites in real time.

Research Square
Curtin University (AU)
Curtin University of Technology
GNSS positioning and interference
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon