Optimized Ensemble Method of SVMD-SVR+BiLSTM for Short-Term Forecasting of Earth Rotation Parameters

Abstract

Earth rotation parameters (ERP), including the polar motion components in the X and Y directions (PMX and PMY) and the difference between universal time 1 and coordinated universal time (UT1-UTC), contain important geodynamic information, and their accurate short-term prediction is of great significance. In this study, an optimized hybrid forecasting method integrating successive variational mode decomposition (SVMD), support vector regression (SVR), and bidirectional long short-term memory (BiLSTM) is proposed for short-term ERP prediction. First, the original ERP series are decomposed by SVMD into several intrinsic mode functions and a residual term. Then, the augmented Dickey-Fuller test and multiscale fuzzy entropy are used to evaluate the stationarity and complexity of the decomposed components, respectively. Based on these results, non-stationary or low-complexity components are assigned to the SVR branch, whereas stationary components with higher complexity are assigned to the BiLSTM branch. The proposed method was trained and evaluated using IERS 20 C04 ERP data from 2016 to 2026. Compared with the best reported results of the 2nd Earth Orientation Parameters Prediction Comparison Campaign, the 30-day prediction results showed the mean absolute error (MAE) improvements reach 25% for PMX, 59% for PMY, and 40% for UT1-UTC, respectively. Improvements are also observed for PMY at the 1-day and 10-day horizons and for PMX at the 10-day horizon. These results indicate that the proposed method can capture the multi-scale characteristics of ERP series and be effective for the short-term ERP prediction.