Abstract: Methods of PCA (principal component analysis) and PLS (partial least squares) based on RBF (radial basis function)neural network are proposed for the reason that the generalization ability of common neural networks debases when the input data is high-dimension or correlations exist. These two methods can reduce the dimension and extract the correlations of the input data. They are used in the prediction of polypropylene melt index, and the simulation results show that the statistical methods improve the predictive precision successfully.