Based on Matlab, BP neural network model for efficiency and head of centrifugal pumps with compound impeller predicting was established. Seventy-three groups of experimental data were selected as samples for BP neural network training with Levenberg-Marquardt law. Then twelve experimental data extra was random selecting to test the trained BP neural network. The main parameters for experimentation are flow rate Q, the number of blade z, outlet angle of blade β2, inlet diameter of splitter blade Di, outlet width of impeller b2, efficiency η and head H. Select Q, z, β2, Di , b2 as input layer, η and H as output layer. The results show the predicted value favourably accorded with experiment. So it is possible to use BP neural network for predicting performance of centrifugal pumps with compound impeller. BP neural network can be applied to compound impeller designing, which can shorten experimental time and reduce cost.