The objective of this paper is to test the application effect of Bayesian discriminant analysis theory in the annual runoff of Biliu reservoir forecasting. The proposed models are called Naive Bayesian classifier (N-Bayes) and Bayesian Discriminance criterion (D-Bayes), The structure of the paper is as follows: 1) according to the result of clustering analysis and actual conditions, the range of the runoff is divided to five levels that are deficient water, Partial deficient water, medium water, Partial abundant water and abundant water respectively; 2) selecting ten classification predictors as the correlation coefficients method and establishing runoff forecasting models of Biliu reservoir by considering the characteristic and physical genesis of runoff; 3) the forecasting models were experimented and tested by use of the data during 2001 to 2010. The result is compared with the practical attribute category shows that the accuracy of N-Bayes and D-Bayes are more than 0.7, the D-Bayes is superior to the N-Bayes in forecasting efficiency, both the two models can be a reference for reservoir operation management.