The grain output fluctuation is a result of several factors. And there is a very complex nonlinear relation between these factors. Lacking the ability to reflect the nonlinear regulation, most of traditional prediction method leads to low accuracy of prediction. BP neural network model has good nonlinear approximation capacity and it does well in prediction of Chinese grain output. Principal component analysis can be associated with the fuzzy variable data for dimension reduction. The combination of PCA and BPNN can optimize the network structure and improve the prediction precision. The results show that the accuracy of combined model is improved by 3% and the efficiency of network training performance also has been improved in different degree.