Short-term prediction of polar motion using an LS-BiLSTM-Attention Model integrated with angular momentum and Kriging interpolation

With the rapid development of space geodesy technology and the aerospace industry, cutting-edge fields such as satellite precise orbit determination, deep space exploration, and autonomous navigation have put forward more stringent requirements for high-precision and high time-resolution polar motion prediction data. Existing ultra-rapid polar motion solutions have relatively low accuracy, while final solutions suffer from low time resolution, making it difficult to meet the growing demand for high-frequency dynamic observations. Moreover, due to the mismatch in time resolution between polar motion observation data and external physical excitation sources, existing studies often suffer from the loss of high-frequency features during the data fusion process. To overcome the loss of high-frequency physical excitation information caused by the downsampling process, this paper employs the Kriging interpolation method to accurately estimate and generate a high-frequency polar motion sequence with a 3-hour resolution. Furthermore, we attempt to directly perform multi-source feature fusion prediction by combining it with the unweighted original 3-hour AOM sequence. Addressing the long-term dependency features of high time-resolution time series data, this paper further constructs an LS-BiLSTM-Attention combined network for refined high-frequency prediction. Comparative experiments show that the prediction results based on the 3-hour high-frequency AOM data are significantly superior in accuracy to the traditional scheme of integrating "AOM data weighted from a 3-hour resolution to a 1-day resolution." For the 30th-day prediction, compared with the LS+AR baseline model, the accuracy of this multi-source high-frequency fusion model is improved by 12.8% and 19.4% in the PMX and PMY directions, respectively. The results fully demonstrate that preserving and deeply mining high time-resolution AOM physical excitation information plays an important role in suppressing error divergence and improving the comprehensive performance of polar motion prediction.