Abstract: Objectives: There is a time lag in the calculation of polar motion data, which is difficult to meet the needs of space engineering such as satellite orbit determination and deep space exploration. Methods: A short-term forecasting method combining Effective Angular Momentum (EAM) and polar motion timing data is proposed. Firstly, the EAM data is preprocessed by weight processing, dimension normalization, etc., and the standard sequence consistent with the spatiotemporal resolution of the polar motion data is obtained. Secondly, grey correlation analysis is used to reveal that the EAM-z component has significant coupling effect on the X and Y directions of the polar motion. Finally, a multi-variable input-univariate output prediction model based on multi-head attention mechanism was constructed, which was characterized by x, y, z component data of EAM and historical data of polar motion to achieve 30-day span of polar motion prediction. The sliding window strategy was used to update the prediction data every 7 days, and a total of 58 prediction experiments were conducted. Results: The results show that compared with the classical LS+AR model, the average absolute error of this model is relatively high in the ultra-short term, but the prediction interval shows a significant advantage and the prediction error decreases after about the 10th day. Conclusions: The multi-head attention mechanism model proposed in this paper can effectively extract the features of EAM and polar motion historical data for short-term prediction, providing a practical reference for subsequent polar motion prediction research.