High-precision Earth Orientation Parameter (EOP) forecasting, which encompasses parameters such as Polar Motion (PM), universal time, and length of day, is crucial for facilitating the transformation between terrestrial and celestial reference frames in various applications (e.g., satellite autonomous navigation, deep space exploration, and geodynamic research). However, the mainstream predictive methods (including Least squares, numerical decompositions) have serious limitations, such as tail effect, prior periods and poor estimations of model parameters. To improve the accuracy of EOP forecasting (1~360 days), this paper introduces an integrated model combining Singular Spectrum Analysis (SSA), Prony's method, and Autoregressive (AR) models, demonstrated through the case of PM parameter prediction: Firstly, the SSA is used to separate the principal components (e.g., trend, annual, and Chandler terms) and residual components from the original PM observations. Secondly, combined with the Prony method to model and extrapolate these principal components based on complex exponentials functions; and we combine the widely used AR method to predict their residuals. To verify this combined approach, we conduct multiple prediction experiments based on the IERS EOP 20 C04 data. The experimental results showed that the SSA+Prony+AR model effectively captures the time-varying characteristics and significantly mitigates the tail effects of the PM components. Compared with traditional LS+AR models and IERS Bulletin A forecast products, our proposed model exhibits superior performance in medium to long-term polar motion forecasting, particularly reducing forecast errors by nearly 40% in the X direction. These findings can also provide valuable insights into the forecasting of other EOP parameters.