Extending the Predictive Horizon of Earth’s Polar Motion Using a Hybrid Deep‑Learning Framework

Abstract

Polar motion (PM) describes the motion of the Earth’s rotation axis as it wanders across the Earth’s crust and is essential for space geodesy, navigation, and precise geophysical measurements. Real‑time PM data are not directly observable, and existing prediction models, such as those in IERS Bulletin A, typically provide forecasts limited to one year. This study proposes a deep learning framework based on a Long Short‑Term Memory (LSTM) network augmented with a multi‑head attention mechanism for long‑term PM prediction. The model jointly predicts PMX and PMY components using differenced input sequences to capture nonlinear temporal dependencies and multi‑frequency periodic behaviors. Experiments on the IERS C04 dataset demonstrate that the proposed model substantially outperforms classical linear predictors. Specifically, the proposed model reduces the total prediction error by approximately 61% at a 600‑day horizon and 39% at 1100 days compared with a conventional baseline, achieving a mean absolute error of 16 mas at 600 days and 32 mas at 1100 days. This demonstrates the effectiveness of hybrid LSTM‑attention architectures in capturing long‑range temporal dynamics in geophysical time series and their potential for extended Earth rotation forecasting.