Due to the lag in the calculation of polar motion, it fails to meet the real-time requirements in space engineering fields such as deep space exploration and satellite orbit determination. Therefore, it is of great significance to obtain predicted values of polar motion. This study proposes a hybrid prediction model that integrates the Least Squares (LS) method and a two-stage attention mechanism, aiming to achieve a 30 days short term prediction of polar motion. First, the LS model is used to perform trend fitting and extrapolation on the polar motion observation sequence. Then, a neural network model based on the two-stage attention mechanism is constructed to train and predict the residual part. Finally, the final predicted value of polar motion is obtained by adding the LS extrapolated value and the predicted residual value. Experimental results indicate that, when compared with the prediction effects of the least squares and autoregressive methods, the accuracy of the proposed hybrid method is significantly improved in the short-term prediction of the polar motion in the Y-direction. Meanwhile, the experiment reveals that increasing the amount of basic data can further enhance the prediction performance of this hybrid model, verifying that an increase in data volume can improve the short-term accuracy of the hybrid model. The hybrid prediction method of LS and the two-stage attention mechanism proposed in this study provides new technical ideas and method references for improving the prediction accuracy of Earth rotation parameters.