High-precision polar motion forecasting using a hybrid linear–skipGRU model

Accurate and rapid prediction of Earth orientation parameters (EOP), particularly polar motion (PM), is essential for satellite navigation and positioning, orbit determination, deep space exploration, and geodynamics research. However, traditional linear forecasting methods are limited in their ability to capture complex nonlinear characteristics. Conversely, while pure data-driven models possess powerful pattern-matching capabilities, they often lack robustness unless deterministic components are explicitly decoupled from stochastic variations. To reconcile this, this study proposes a hybrid neural network model that integrates a parallel linear module and a skip-connected gated recurrent unit (skipGRU) module. First, the deterministic components of the PM series are modeled using an extended least squares (LS) method with an optimized 20-year baseline and an auxiliary 450-day periodic term to mitigate the significant spectral residuals observed near the Chandler frequency band. The LS-derived residuals are subsequently used to train the linear–skipGRU model, where the linear branch captures autoregressive regularity while the skipGRU focuses on nonlinear dynamics. Experimental results demonstrate that the proposed model achieves comparable short-term accuracy to the International Earth Rotation and Reference Systems Service (IERS) Bulletin A, while significantly outperforming it in medium- and long-term forecasts. Specifically, for X coordination of PM prediction, the hybrid model improves accuracy by approximately 12, 39, and 40% at forecast horizons of 90, 180, and 360 days, respectively, compared with the IERS Bulletin A. Moreover, the modular architecture of the proposed framework provides considerable potential for integrating more advanced sub-models and offers a new perspective for improving existing EOP prediction approaches.