Machine Learning System Service for Earth Orientation Parameters Forecast

Abstract

Earth Orientation Parameters (EOP) polar motion (xp, yp), the difference between astronomically observed and atomic time (UT1−UTC), length-of-day excess (LOD), and celestial pole offsets (dX, dY ) describe irregularities in the Earth’s rotation that are essential inputs to satellite navigation, precise geodesy, and reference-frame transformations, yet are difficult to predict because they arise from the coupled, only partially observable exchange of angular momentum between the solid Earth, the atmosphere, and the oceans. This thesis presents the design, implementation, and evaluation of an end-to-end software platform that forecasts EOP using machine learning, framed primarily as a contribution to Computer Systems and Networks: the scientific forecasting problem motivates the system, but the central contribution is the architecture and engineering of a reproducible, modular, testable platform that carries a forecast from raw data acquisition through to a served, visualized prediction. The platform ingests real International Earth Rotation and Reference Systems Service (IERS) EOP bulletins and real atmospheric/oceanic angular momentum data (AAM/OAM) from the GFZ Potsdam Earth System Modelling group, alongside a reproducible synthetic data generator used for offline engineering validation. Data are cleaned via robust, Median-Absolute- Deviation-based outlier detection and time-aware interpolation onto a strict daily calendar, merged with AAM drivers, and transformed into a supervised-learning representation using lag features, rolling statistics, and calendar encodings, with an explicit shift(1) guard against target leakage. Five model families – Ridge regression, Random Forest, XGBoost, an LSTM network with an additive attention mechanism, and a Split Conformal Prediction wrapper for uncertainty quantification – are trained independently for each of six EOP targets at four forecast horizons (7, 14, 30, and 365 days) using a strict chronological train/calibration/test split, and served through a FastAPI inference layer and a browser-based visualization interface. The complete Targets × Horizons × Models experimental grid (120 trained artifacts) v was evaluated on the reproducible synthetic dataset, and is reported and analyzed as such; a smaller, targeted subset was additionally trained and evaluated directly on real IERS/AAM data to characterize the gap between synthetic-data and real-data performance. Results indicate that model accuracy degrades with forecast horizon for most targets, that no single model family is uniformly best across all targets, and that the implemented conformal prediction intervals achieve empirical coverage close to their 90% nominal target for most target/ horizon combinations but substantially undershoot it for others – evidence directly attributable to the non-stationarity of EOP series violating conformal prediction’s exchangeability assumption. The realdata subset shows that this gap is target-dependent: real-data forecast error for length-of-day was several times higher than on synthetic data, while polar-motion forecasts were comparable or better on real data. The thesis concludes that the platform successfully demonstrates an operational, reproducible, and honestly evaluated machine-learning forecasting service for EOP, while identifying concrete limitations – most importantly the synthetic-data scope of the full experimental grid – that bound the strength of its scientific claims.

Note on evidentiary status. All quantitative results reported in this thesis are drawn directly from experiment artifacts (generated figures, leaderboard tables, and training logs) produced by the implemented system; no numerical result is invented. Where a specific value could not be recovered from the available project records, it is marked [TO BE VERIFIED FROM EXPERIMENTAL RECORDS] rather than estimated.