High-precision polar motion prediction is crucial for applications such as real-time satellite orbit determination and deep space probes navigation. We combine the singular spectrum analysis(SSA) and weighted least squares(WLS) models, which takes into account the temporal variability of periodic components of polar motion. Combining with the auto-regressive(AR) model, we prove the effectiveness of SSA+WLS+AR model in improving the accuracy of polar motion prediction. We analyze the impact of detrending, the length of basic sequence, and the weighting method on the final prediction accuracy of the SSA+WLS+AR model. The results show that in the X direction, the model combination of "no detrending+50-year basic sequence+inverse weighting" has the highest prediction accuracy, while in the Y direction, the model combination of "detrending +50-year basic sequence+inverse weighting" has the highest prediction accuracy. When the prediction span is from 150 to 360 days, the prediction accuracy of SSA+WLS+AR model is better than that of bulletin A provided by international Earth rotation service(IERS), with maximum improvements of 33% and 26% in the X and Y directions, respectively. The prediction results show that the first-day prediction accuracy of polar motion is significantly better using the C04 20 sequence compared to the C04 14 sequence.