Chinese scientists developed a new model that substantially improves Earth rotation forecast accuracy

Earth spins, but not with the regularity of an atomic clock. Its rotation axis shifts, its orientation in space changes and the duration of a day varies. Scientists describe these movements using Earth Orientation Parameters, or EOP.

EOP include polar motion, which tracks the rotation axis relative to Earth’s surface; precession and nutation, which describe changes in the axis direction in space; and Universal Time and Coordinated Universal Time (UT1-UTC) and Length of Day (LOD), which track rotational timing. Together, they connect Earth-fixed coordinates with space-fixed coordinates, supporting satellite navigation, astrometry and spacecraft tracking.

The study focuses on UT1-UTC and LOD. UT1 is time determined from Earth’s actual rotation, while UTC is the time standard maintained using atomic clocks. Their difference, UT1-UTC, indicates how rotational time is departing from atomic time. Length of Day, or LOD, measures variations around the nominal 24-hour day.

Accurate forecasts are needed because final EOP observations are not immediately available. Yet predicting them is difficult. Atmospheric winds, ocean currents, terrestrial water storage and sea level exchange angular momentum with the solid Earth, while tides and internal dynamics add further variations.

Previous studies have tested forecasting approaches including least-squares extrapolation, autoregressive models and neural networks. Traditional linear methods can represent trends and cycles but struggle with nonlinear behavior and accumulating errors over longer periods. Machine-learning models can capture more complex patterns, yet approaches based only on past geodetic observations may lack information about the physical forces driving rotation changes.

Researchers from the College of Geodesy and Geomatics at Shandong University of Science and Technology developed a hybrid model that combines both perspectives. Published in Geodesy and Geodynamics, the approach pairs segmented least squares (SLS) with Autoformer, a deep-learning model for long time series, while incorporating angular-momentum information from the atmosphere, oceans, land water and sea level.

SLS extracts changing trends and known periodic signals. Autoformer then analyzes the remaining nonlinear patterns and longer-range dependencies.

For UT1-UTC, prediction accuracy over 30 to 90 days improved by approximately 32% to 40% compared with the International Earth Rotation and Reference Systems Service (IERS) Bulletin A product. LOD forecasts improved by about 34% to 38% over a leading traditional benchmark using angular-momentum data.

The results show the value of combining physical information with deep learning. “Autoformer is only one of many machine-learning models, and we cannot assume that it is the best possible option,” the authors noted. “Future work will compare its predictive performance with that of other machine-learning approaches more comprehensively.”

Contact the author: Qiaoli Kong, kqlabc3334@163.com, College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao, China

The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 300 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).