A Deep Learning Approach for Improving Near Real-Time LOD Prediction Accuracy by Integrating IGS Rapid Products and EAM Datasets

Length of day (LOD), a fundamental component of Earth orientation parameters (EOP), reflects variations in Earth's rotation rate. Due to the influenced of atmospheric circulation, ocean currents, hydrological processes, and internal Earth dynamics, LOD exhibits complex nonlinear characteristics, making accurate prediction challenging Current LOD prediction models primarily depend on the high-precision, smoothed EOP C04 series provided by the IERS. However, this series has an inherent delay of approximately 30 days, causing it unsuitable for real-time applications such as interplanetary spacecraft tracking and navigation, GNSS meteorology, and real-time satellite orbit determination. To address these challenges, a novel approach based on a convolutional long short-term memory (ConvLSTM) method was proposed, which captures the time-varying characteristics of LOD by integrating IGS rapid products and effective angular momentum (EAM) datasets to improve the accuracy of near real-time LOD predictions. Results indicate that incorporating GNSS near real-time (NRT) data improves short-term (10–30 days) LOD prediction accuracy by 55.07%. By incorporating GNSS NRT data and EAM datasets, the ConvLSTM model significantly improves LOD prediction accuracy across various time scales. This enhancement not only strengthens Earth's rotation prediction models but also facilitates critical applications in real-time satellite orbit determination, extreme weather forecasting, and so on.