Abstract:

Earth polar motion (PM), a pivotal geodynamic parameter governing deep space exploration and satellite precise orbit determination, necessitates high-precision prediction models that persist as a research focus in space geodesy. To address the issues of accumulated prediction errors caused by inconsistencies between training and application scenarios, as well as the effect of signal noise in long short-term memory (LSTM) neural networks, we propose a short-term PM prediction method with a cascaded LSTM architecture based on singular spectrum analysis (SSA) denoising. The proposed method first employs SSA to eliminate high-frequency noise components from polar motion time series signals. Subsequently, it fully considers the evolving scenario characteristics across different future prediction horizons, and constructs an interconnected cascaded LSTM framework where multiple sub-models are sequentially connected for progressive information transfer. The experimental results based on the EOP 20 C04 dataset spanning 1984 to 2024 demonstrate significant improvements: For 1~10 days short-term predictions, the proposed method achieves mean absolute errors (MAE) of 1.70 mas and 0.93 mas in the X and Y polar motion components, respectively. Compared to recursive LSTM baselines, the proposed model achieves 42.8% and 48.0% improvements, respectively. Furthermore, it outperforms existing SSA-recursive LSTM hybrid benchmarks by 11.0%and 28.5%in MAE reductions. Significantly, the cascaded architecture demonstrates superior predictive capability in 6~10 days forecasts, validating its effectiveness in mitigating error propagation while enhancing mid-to-long-term forecast stability. The prediction results are applied to the transformation between celestial and Earth coordinate systems for satellite orbits, significantly improving the accuracy of coordinate conversion.

Key words: Earth's polar motion, short-term prediction, singular spectrum analysis, cascaded LSTM, denoising optimization