The short-term prediction of Polar Motion and LOD using the Multivariate Multistep 1D Convolutional Neural Networks with Multioutput strategy.
Guessoum Sonia  1, 2, 3, 4, 5, 6, 7@  
1 : Sonia Guessoum
2 : Santiago Belda
3 : Jose M. Ferrandiz
4 : Sadegh Modiri2
5 : Sujata Dhar
6 : Robert Heinkelmann
7 : Harald Schuh

Accurate prediction of Earth orientation parameters (EOPs) is critical for astro-
geodynamics, high-precision space navigation, and positioning, onboard au-
tonomous orbit determination, and deep space exploration. However, the cur-
rent models' prediction accuracy for EOPs is significantly lower than that of

geodetic technical solutions, which can adversely affect certain high-precision

real-time users. To address this challenge, deep learning neural networks, specif-
ically the 1DCNN and LSTM models, have emerged as powerful tools. These

models possess the capability to automatically learn complex mappings from
inputs to outputs, This feature holds great promise for time series forecasting,

especially for problems with intricate nonlinear dependencies, multivariate in-
puts, multi-output, and multi-step forecasting. Consequently, these models are

well-suited for our study, where our objective is to predict three parameters.

simultaneously and enhance the short-term prediction accuracy of PM (polar
motion) and LODR (length of day rate).

The computational strategy follows multiple steps, first, using the SSA the de-
terministic time-varying signal of the EOP time series can be more precisely and

reasonably detected and modeled. Then the reconstructed series and its corre-
sponding residuals are used for 1DCNN training and prediction. However, we

develop a Multivariate Multi step 1DCNN model with a multi-output strategy
using three different scenarios including the Ocean Angular Momentum (OAM),

Atmospheric Angular Momentum (AAM), and Hydrological Angular Momen-
tum (HAM), to predict both the deterministic and the stochastic part for (Xp,

Yp). Then the best case with fewer errors is chosen to predict the Polar motion
and length of day (LOD) at the same time in short term.

The results of two years of prediction experiments based on the EOP 14 C04 se-
ries using 1DCNN are compared with LSTM and show that the proposed model

can predict both the deterministic and the stochastic parts for the three param-
eters at the same time with significant improvements in polar motion (PM) and

length of day (LOD) for short-term prediction.

Online user: 1 Privacy