Improved polar motion prediction with denoised geophysical fluid excitation information and long short term memory neural network

Abstract Polar Motion (PM) is a crucial parameter linking the celestial and terrestrial reference frames. Integrating Earth's fluid effective angular momentum (EAM) and neural network is an effective way to improve PM prediction accuracy. However, EAM data is polluted by high-frequency noise and the neural network models are highly sensitive to the input EAM dataset. To improve PM prediction accuracy, this study proposes a method combining Complex Segmented Least Squares (CSLS) with Long Short-Term Memory (LSTM) using the noise-reduced EAM data and the filtered 6-day forecast EAM data. Geodetic angular momentum function is employed to reduce the high frequency noise signal of EAM, Kalman Filter is adapted to denoise the 6-day forecast product, and CSLS + Autoregression (AR) and CSLS + LSTM are used for fitting and prediction. The 6-day predicted outcomes with denoised EAM achieved the reductions of the Mean Absolute Error (MAE) in x direction by 48.68% and 53.47% compared to those of Bulletin A and those gained with the original EAM, and in y direction, the reductions reached 37.16% and 41.24%, respectively. The 365-day predicted results with CSLS + LSTM achieved MAE reduction of 43.02% and 32.65% in x and y directions, respectively, compared to Bulletin A. The results confirm that, for short-term PM prediction, CSLS + AR with the denoised EAM data outperforms other methods, while for long-term, CSLS + LSTM with the original EAM data performs better. It is also found that the denoised-EAM can improve PM prediction accuracy within future 1–30 days, while for longer time, there is no significant improvement. Graphical Abstract