Combined BiLSTM and ARIMA models in middle- and long-term polar motion prediction

Abstract

As one of the main components of the Earth orientation parameters, short-term prediction of the geodetic polar motion series is crucial in the field of deep-space exploration, high-precision positioning, and timing services, which require high real-time performance. Additionally, its middle- and long-term prediction is equally important in climate forecasting and geodynamics research. In this study, we propose the combined BiLSTM+ARIMA model, which is based on bidirectional long- and short-term memory (BiLSTM) and autoregression integrated moving average (ARIMA). First, ensemble empirical mode decomposition (EEMD) is performed as a filter to decompose the polar motion time series to obtain low- and high-frequency signals. The EOP14 C04 time series provided by International Earth Rotation and Reference Systems Service and decomposed by EEMD includes low-frequency signals like the long-term trend, decadal oscillation, Chandler wobble, and prograde annual wobble, along with shorter-period high-frequency signals. Second, low- and high-frequency signals are predicted using BiLSTM and ARIMA models, respectively. Finally, the low- and high-frequency signal forecast components are reconstructed to obtain geodetic polar motion predictions. In middle- and long-term polar motion prediction, the results show that the proposed model can improve the prediction accuracy by up to 42% and 17%, respectively. This demonstrated that the BiLSTM+ARIMA model can effectively improve the accuracy of polar motion prediction.