Objectives The difference between universal time and coordinated universal time (UT1-UTC) is a critical component of the Earth orientation parameters (EOP). Accurate and rapid prediction of UT1-UTC is essential for real-time applications such as GNSS meteorology and precise orbit determination of artificial satellites. Traditional methods for predicting UT1-UTC often experience significant accuracy degradation in medium- and long-term forecasts, making them inadequate for high-precision requirements in applications like the BeiDou navigation satellite system and precision guidance in military operations.

Methods We propose a novel method for predicting UT1-UTC by integrating the axial component data ${\chi }_{\mathrm{3}}$ of effective angular momentum (EAM) with the EOP14 C04 series using a convolutional long short-term memory (ConvLSTM) model. Initially, leap seconds and solid Earth tide terms are removed from the original UT1-UTC series to obtain the difference between the tidally reduced universal time and international atomic time (UT1R-TAI) data. Spectral analysis is then performed on the axial component data ${\chi }_{\mathrm{3}}$ of EAM and UT1R-TAI data using the fast Fourier transform to investigate whether the axial component data ${\chi }_{\mathrm{3}}$ of EAM can comprehensively describe the excitation of UT1-UTC. Subsequently, a ConvLSTM model incorporating the axial component data ${\chi }_{\mathrm{3}}$ of EAM is constructed to predict the UT1-UTC time series. After prediction, the leap seconds and solid Earth tide terms are reintroduced, while the accuracy of predictions is evaluated.

Results Analysis of observations reveals a strong correlation between the axial component ${\chi }_{\mathrm{3}}$ of EAM and UT1-UTC data after applying leap-second and tidal corrections, with consistent amplitude and phase characteristics in their frequency spectra. This indicates that the axial component ${\chi }_{\mathrm{3}}$ of EAM serves as a primary excitation source for UT1-UTC. Compared to the prediction accuracy of participants in the second Earth orientation parameters prediction comparison campaign, the ConvLSTM model demonstrates superior performance in medium- to long-term predictions spanning 90 d to 360 d, with accuracy improvements ranging from 30.27% to 92.44%. Additionally, compared to Bulletin A, the ConvLSTM model achieves accuracy enhancements of 41.46%, 70.07%, and 59.43% for prediction spans of 60 d, 180 d, and 360 d, respectively.

Conclusions The results confirm that the ConvLSTM model significantly improves the medium- to long-term prediction accuracy of UT1-UTC. These findings are crucial for autonomous determination of EOP and real-time applications, as well as for precise satellite orbit determination and other related fields.