Abstract: Accurate ultra-short-term prediction of UT1-UTC is of critical importance for geodesy, satellite navigation, and space geophysics. This study presents a novel forecasting approach that integrates Effective Angular Momentum functions (EAM) with the Extreme Gradient Boosting (XGBoost) model. The method first preprocesses the UT1-UTC series by removing skip seconds, correcting tidal effects, and applying second-order differencing. Trend and periodic components are then separated using a Least Squares (LS) model to obtain stationary residuals. Historical and predicted EAM data are incorporated as external excitation factors to capture the influence of non-tidal geophysical processes. The residual sequence is subsequently modeled using XGBoost to capture nonlinear dependencies, enabling multi-step rolling forecasts. Experiments conducted on the EOP 14 C04 dataset from the International Earth Rotation and Reference Systems Service (IERS) show that, within a 10-day prediction horizon, the EAMz-assisted LS+XGBoost method reduces the mean absolute error by approximately 29.6% compared with models relying solely on the UT1-UTC historical series. Compared with methods from the Second EOP Prediction Comparison Campaign, the proposed approach achieves the highest accuracy on Days 1, 3, and 4. These results demonstrate that integrating physical excitation information with machine learning substantially enhances the ultra-short-term prediction capability of UT1-UTC.