A predictive model was set up to improve the prediction precision of polar motion of earth orientation parameters. The periodic characteristics of interpolated basic series was studied by Fourier analysis, the feasibility of basic series resampling was verified, then the trend terms were derived from the interpolation basic series, and the multiple input-single output BP(Back Propagation)neural network model was used to predict the residual series for different time spans. Finally the predicted polar motion was achieved by combining the trend terms with residual series. Prediction results indicate that the appropriate selection of interpolation basic series can realize high precision prediction of polar motion. Moreover, the BP neural network can be applied to the prediction of polar motion of earth orientation parameters effectively.