Improving the Accuracy of Polar Motion Prediction by Combining Least Squares and the iTransformer Model

Polar motion prediction is an important means to study the dynamics of Earth's rotation and detect the internal structure of the Earth. This paper proposes an LS + iTransformer method that combines the least squares (LS) method with the iTransformer deep learning model. Using IERS EOP 20 C04 data from 2000 to 2022, Fast Fourier Transform (FFT) to identify the dominant period, and Singular Spectrum Analysis (SSA) to decompose the signal, it is found that the annual cycle contributes more variance than the Chandler wobble, with the Chandler wobble showing a decaying trend during this period. Based on these, a 360‐day continuous forecasting experiment was conducted on four scenarios of LS + AR, LS + iTransformer, iTransformer, and IERS Bulletin A to evaluate the performance of the proposed model. The results showed that the forecast errors exhibit seasonal variations. Compared with IERS Bulletin A, the LS + iTransformer model exhibited the greatest improvement, followed by the iTransformer model, and the LS + AR model showed the least improvement. The optimal LS + iTransformer model improves the prediction accuracies of PMX and PMY by about 33% and 44% on average, with the highest improvement being about 45% and 56%, respectively. It performs well in all forecast horizons and has a particularly significant advantage in middle‐ and long‐term forecasts. Due to the dominant cycle component of PMY (96.1%), the accuracy of PMY is significantly improved compared with PMX. In summary, the LS + iTransformer model effectively captures the evolution law and complex change patterns of polar motion signals, providing a promising method for enhancing polar motion forecasting.