Improving the rapid EOP products
Introduction to the methods
The following diagram shows the overall scheme of the framework
The inputs to all the algorithms are rapid IERS data, EAM observations and 14-day forecasts, tides, and multivariate ENSO index.
Evaluation is done against IERS 14 C04 series. The forecasting horizon for all the algorithms is 63 days.
Separate models are trained for polar motion and dUT1. Baseline is the rapid IERS EOP itself.
In full ensemble approach, a weighted average of all the types of ResLearner is computed. Weights are based on the performance on either IERS 20 C04, or IERS 14 C04.
The EAM data used are the component-wise summation of AAM, HAM, OAM, and SLAM (i.e., EAM = AAM + HAM + OAM + SLAM).
Acknowledgements
The data sources used in the current framework are taken from different datacenters. These are as follows.
IERS EOP rapid and final 14 C04 series can be accessed here.
IERS EOP 20 C04 series can be accessed here.
JPL EOP series can be accessed here.
ETH EAM 14-day forecasts can be accessed here.
GFZ EAM observations can be accessed here.
NOAA multivariate ENSO index be accessed here.
Methods description
The specifications of the methods that are currently operational are mentioned in the following table.| Method | Target EOP in training | Algorithm |
|---|---|---|
| PhycoRNNreslearner_gps_rapid-63d-iers-eth | IERS 14 C04 | ResLearner physically-constrained coupled oscillatory recurrent neural networks |
| PhycoRNNreslearner_gps_rapid-63d-iers20-eth | IERS 20 C04 | ResLearner physically-constrained coupled oscillatory recurrent neural networks |
| PhycoRNNreslearner_gps_rapid-63d-jpl-eth | JPL final | ResLearner physically-constrained coupled oscillatory recurrent neural networks |
| reslearner_gps_rapid-63d-iers-eth | IERS 14 C04 | ResLearner self-calibration |
| reslearner_gps_rapid-63d-iers20-eth | IERS 20 C04 | ResLearner self-calibration |
| reslearner_gps_rapid-63d-jpl-eth | JPL final | ResLearner self-calibration |
| reslearner_full_ensemble_iers | IERS 14 C04 | ResLearner full ensemble |
| reslearner_full_ensemble_iers20 | IERS 20 C04 | ResLearner full ensemble |
The methods are all developed by the Space Geodesy group at ETH Zurich and Earth System Modelling GFZ. For instance, ResLearner PhycoRNN is based on the concept of physically-constrained neural networks. Figure below shows how the method can be trained based on final EOP data to predict well.
Products
The predictions are avaliable here
Quality Control
The daily evaluation report is avaliable at EOP prediction accuracy