Have a personal or library account? Click to login
ESMGFZ EAM Products for EOP Prediction: Toward the Quantification of Time Variable EAM Forecast Errors Cover

ESMGFZ EAM Products for EOP Prediction: Toward the Quantification of Time Variable EAM Forecast Errors

Artificial Satellites
Volume 58 (2023): Issue 4 (December 2023)
By: Robert Dill,  Henryk Dobslaw and  Maik Thomas  
Open Access
|Jan 2024

Figures & Tables

Figure 1.

Root mean squared error (RMS) between AAM forecast and subsequent AAM analysis representing the forecast error for daily 6-day forecasts sampled every 3 hours from January 2016 until December 2022: AAM mass terms (left), AAM motion terms (right) for X1 (top), X2 (middle), and X3 (bottom).
Root mean squared error (RMS) between AAM forecast and subsequent AAM analysis representing the forecast error for daily 6-day forecasts sampled every 3 hours from January 2016 until December 2022: AAM mass terms (left), AAM motion terms (right) for X1 (top), X2 (middle), and X3 (bottom).

Figure 2.

Root mean squared error (RMS) between OAM forecast and subsequent OAM analysis representing the forecast error for daily 6-day forecasts sampled every 3 hours from January 2016 until December 2022: OAM mass terms (left), OAM motion terms (right) for X1 (top), X2 (middle), and X3 (bottom).
Root mean squared error (RMS) between OAM forecast and subsequent OAM analysis representing the forecast error for daily 6-day forecasts sampled every 3 hours from January 2016 until December 2022: OAM mass terms (left), OAM motion terms (right) for X1 (top), X2 (middle), and X3 (bottom).

Figure 3.

Root mean squared error (RMS) between HAM forecast and subsequent HAM analysis representing the forecast error for daily 6-day forecasts sampled every 24 hours from January 2016 until December 2022: HAM mass terms (left), HAM motion terms (right) for X1 (top), X2 (middle), and X3 (bottom).
Root mean squared error (RMS) between HAM forecast and subsequent HAM analysis representing the forecast error for daily 6-day forecasts sampled every 24 hours from January 2016 until December 2022: HAM mass terms (left), HAM motion terms (right) for X1 (top), X2 (middle), and X3 (bottom).

Figure 4.

Neural network to predict level of uncertainty of EAM forecast. Training and input data are the absolute values of the differences between EAM forecast and EAM analysis. The target and output is the level of uncertainty of the EAM forecast. AAM and OAM forecasts are three-hourly sampled with 48 epochs for days. The input series also include one initial epoch before the forecast starts. For HAM and SLAM with only 24-hourly sampling, the input and output neurons are reduced accordingly.
Neural network to predict level of uncertainty of EAM forecast. Training and input data are the absolute values of the differences between EAM forecast and EAM analysis. The target and output is the level of uncertainty of the EAM forecast. AAM and OAM forecasts are three-hourly sampled with 48 epochs for days. The input series also include one initial epoch before the forecast starts. For HAM and SLAM with only 24-hourly sampling, the input and output neurons are reduced accordingly.

Figure 5.

Neural network to predict the level of uncertainty of EAM forecast. Training and input data is the absolute value of the difference between EAM forecast and EAM analysis. The target and output is the level of uncertainty of the EAM forecast.
Neural network to predict the level of uncertainty of EAM forecast. Training and input data is the absolute value of the difference between EAM forecast and EAM analysis. The target and output is the level of uncertainty of the EAM forecast.

RMS prediction error in polar motion for forecast horizons of 5, 10, 40, and 90 days in x-pole, y-pole coordinate and 2D polar motion vector for different EAM forecast input_ Black line: prediction approach using erroneous EAM forecasts; purple line: using perfect EAM forecast without any deviation from EAM analysis; yellow line: using EAM forecasts with corrected AAM X1 and X2 motion terms_ The percentages give the improvement in RMS compared to the original RMS (black)_

RMS [mas]X poleY poleX+iY pole
2016–20205d10d40d90d5d10d40d90d5d10d40d90d
LS + AR + EAM forecast0.931.928.6515.760.651.305.1410.851.132.3210.0519.13
perfect EAM forecasts0.881.688.5615.800.661.285.1010.741.102.119.9719.10
−5.4%−12.5%−1.0%0.25%1.5%−1.5%−0.8%−1.0%−2.7%−9.1%−0.8%−0.2%
corrected AAM0.891.828.6315.770.651.315.0910.771.122.2510.0319.11
−4.3%−5.2%−0.2%0.1%0.0%0.8%−1.0%−0.7%−0.9%−3.0%−0.2%−0.1%
DOI: https://doi.org/10.2478/arsa-2023-0013 | Journal eISSN: 2083-6104 | Journal ISSN: 1509-3859
Journal RSS Feed
Language: English
Page range: 330 - 340
Submitted on: May 25, 2023
Accepted on: Dec 4, 2023
Published on: Jan 19, 2024
Published by: Polish Academy of Sciences, Space Research Centre
In partnership with: Paradigm Publishing Services
Publication frequency: 4 issues per year
Keywords:
effective angular momentum function,
machine learning,
forecast error,
level of uncertainty
Related subjects:
Geosciences,
Geosciences, other

© 2024 Robert Dill, Henryk Dobslaw, Maik Thomas, published by Polish Academy of Sciences, Space Research Centre
This work is licensed under the Creative Commons Attribution 4.0 License.

Previous articleVolume 58 (2023): Issue 4 (December 2023)Next article