Shandong University researchers build model improving Earth rotation forecasts by up to 40%

Researchers from the College of Geodesy and Geomatics at Shandong University of Science and Technology have developed a hybrid mathematical framework that substantially increases the accuracy of Earth orientation predictions, EurekAlert reported. The study, published in the journal Geodesy and Geodynamics, focuses on calculating variations in Universal Time relative to Coordinated Universal Time (UT1-UTC) and changes in the Length of Day (LOD) eurekalert.orgEurekAlerteurekalert.orgOpen source ↗.

Earth Orientation Parameters are necessary to link Earth-fixed coordinate systems with celestial reference frames, which directly affects deep-space tracking, satellite navigation, and astrometry. The rotational velocity of the planet deviates from atomic time due to continuous angular-momentum exchanges involving atmospheric circulation, oceanic currents, terrestrial water storage variations, and sea-level shifts, as well as tidal and deep internal geophysical dynamics eurekalert.orgEurekAlerteurekalert.orgOpen source ↗. Because definitive geodetic observations involve latency, operational positioning systems rely on predictive modeling.

Algorithmic architecture and physical data

To overcome the drift inherent in traditional linear extrapolation and the limitations of purely statistical time-series forecasting, the team designed a two-stage processing pipeline. The framework pairs segmented least squares (SLS) with Autoformer, a deep architecture specialized for long-sequence time-series analysis eurekalert.orgEurekAlerteurekalert.orgOpen source ↗. In addition to historical geodetic observations, the model explicitly incorporates angular-momentum datasets derived from the atmosphere, oceans, terrestrial water, and sea level eurekalert.orgEurekAlerteurekalert.orgOpen source ↗.

Within this workflow, the segmented least squares module isolates systematic trends and established periodic signals. The residual non-linear dynamics and long-range temporal dependencies are subsequently processed by Autoformer, preventing the rapid accumulation of prediction error over extended horizons eurekalert.orgEurekAlerteurekalert.orgOpen source ↗.

Quantitative prediction performance

Benchmarking demonstrated significant gains across operational forecast windows. For UT1-UTC, the model improved prediction accuracy over 30 to 90 days by approximately 32% to 40% compared with the International Earth Rotation and Reference Systems Service (IERS) Bulletin A product eurekalert.orgEurekAlerteurekalert.orgOpen source ↗.

Similarly, LOD forecasts improved by about 34% to 38% relative to a leading traditional benchmark that utilizes angular-momentum data eurekalert.orgEurekAlerteurekalert.orgOpen source ↗. The authors emphasized that future research will evaluate alternative architectures against Autoformer to systematically map the performance envelope of hybrid geophysical forecasting tools eurekalert.orgEurekAlerteurekalert.orgOpen source ↗.