Scholarlink

<p>This paper presents a hybrid prediction approach based on Dynamic Mode Decomposition principle and designed to improve 30-day forecasts of the celestial pole offset (CPO), expressed through its components dX and dY. We assess the method in an experimental framework referenced to the IERS EOP 14 C04 and IERS EOP 20 C04 series. The procedure also incorporates the JPL dX, dY time series as an auxiliary input, used as a supplementary source of information during prediction. To ensure consistency with the IERS CPO values, predictions generated from the JPL series are systematically transformed using total least squares (orthogonal) regression, explicitly accounting for measurement uncertainties in both datasets. In parallel, the IERS CPO series is preprocessed with a penalized least-squares smoother (Whittaker smoother); the resulting smoothed series is then treated as an additional information source within the prediction routine. A key feature of this workflow is that it avoids relying on external data to bridge latency gaps in the final IERS product. Instead, gap handling is performed internally as part of the prediction process, making the method well suited to real-time applications. The experimental setup therefore mirrors operational conditions by combining prediction with simultaneous gap filling. We benchmark the approach against results from the 2nd Earth Orientation Parameters Prediction Comparison Campaign (EOPPCC) and also evaluate its long-term performance in a 10-year experiment using two IERS reference series (EOP 14 C04 and EOP 20 C04). In this long-term test, the best-performing variants for dX (EOP 14 C04) achieve an average mean absolute error (MAE) of 61.5 µas at day 10 and 64.2 µas at day 30; for EOP 20 C04, the corresponding values are 68.9 µas and 72.1 µas. For dY, the averaged MAE with EOP 14 C04 is 66.1 µas at day 10 and 71.8 µas at day 30, while for EOP 20 C04 it is 74.4 µas and 78.0 µas, respectively. In the 2nd EOPPCC comparison, the top variants yield MAE ranges of 40–55 µas for dX and 50–80 µas for dY across the full 30-day forecast horizon.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Celestial pole offsets prediction with multiple data sources: a data-driven approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Marcin Ligas,
  • Maciej Michalczak,
  • Santiago Belda,
  • Jose M. Ferrándiz,
  • Maria Karbon,
  • Sadegh Modiri

摘要

This paper presents a hybrid prediction approach based on Dynamic Mode Decomposition principle and designed to improve 30-day forecasts of the celestial pole offset (CPO), expressed through its components dX and dY. We assess the method in an experimental framework referenced to the IERS EOP 14 C04 and IERS EOP 20 C04 series. The procedure also incorporates the JPL dX, dY time series as an auxiliary input, used as a supplementary source of information during prediction. To ensure consistency with the IERS CPO values, predictions generated from the JPL series are systematically transformed using total least squares (orthogonal) regression, explicitly accounting for measurement uncertainties in both datasets. In parallel, the IERS CPO series is preprocessed with a penalized least-squares smoother (Whittaker smoother); the resulting smoothed series is then treated as an additional information source within the prediction routine. A key feature of this workflow is that it avoids relying on external data to bridge latency gaps in the final IERS product. Instead, gap handling is performed internally as part of the prediction process, making the method well suited to real-time applications. The experimental setup therefore mirrors operational conditions by combining prediction with simultaneous gap filling. We benchmark the approach against results from the 2nd Earth Orientation Parameters Prediction Comparison Campaign (EOPPCC) and also evaluate its long-term performance in a 10-year experiment using two IERS reference series (EOP 14 C04 and EOP 20 C04). In this long-term test, the best-performing variants for dX (EOP 14 C04) achieve an average mean absolute error (MAE) of 61.5 µas at day 10 and 64.2 µas at day 30; for EOP 20 C04, the corresponding values are 68.9 µas and 72.1 µas. For dY, the averaged MAE with EOP 14 C04 is 66.1 µas at day 10 and 71.8 µas at day 30, while for EOP 20 C04 it is 74.4 µas and 78.0 µas, respectively. In the 2nd EOPPCC comparison, the top variants yield MAE ranges of 40–55 µas for dX and 50–80 µas for dY across the full 30-day forecast horizon.