High-precision polar motion prediction using EOP_20_C04 and EAM based on CSLS+AR and CSLS+LSTM methods

High-precision prediction of polar motion (PM) plays an important role in fields such as astronomy, geodesy, navigation, and aerospace. Incorporating information on the effective angular momentum of the geophysical fluid is an effective way to improve the precision of PM prediction. Based on the EOP_20_C04 dataset and the Effective Angular Momentum function (EAM), this study applies Complex Segmented Least-squares (CSLS) + the Long Short Term Memory (LSTM) neural network and CSLS + Autoregression (AR) models to predict PM. For the 6-day polar motion prediction, the mean absolute errors (MAEs) achieved by CSLS+AR are 1.03mas and 0.8 mas in the X and Y directions, respectively, corresponding to reductions of 45.80% and 31.97% when compared to predictions reported routinely in Bulletin A of the International Earth Rotation and Reference Systems Service (IERS). For the 365-day polar motion prediction, the MAEs gained by CSLS+LSTM model are 14.58mas and 10.59mas in the X and Y directions respectively, corresponding to reductions of 28.17% and 51.09% compared to predictions of the Bulletin A, and the prediction accuracy attained by CSLS+LSTM is higher than other prediction schemes. The experimental results show that, when considering EAM information, the CSLS+AR model can achieve better prediction accuracy for short-term PM prediction, however, the CSLS+LSTM model is more effective for medium and long-term PM prediction owning to the excellent non-linear fitting capability of the LSTM deep learning algorithm.