Regarding the flood prevention, the concept of ensemble prediction was integrated into artificial neural networks by considering the data quality and initial weights to reduce the uncertainties of single neural network. The ensemble neural networks were built for 1~3 h ahead flood forecasting in Longquan River Basin, by generating the ensemble members through resampling and initialization methods and combining the model outputs through the arithmetic average and Bayesian model average (BMA). Results indicate that the ensemble neural networks significantly improve the flood forecasting accuracy as compared with the single neural network. The accuracy of ensemble neural network is about 15% to 35% higher than that of the single neural network in terms of root mean square error. Among various ensemble strategies, the combination of initialization and arithmetic average has simpler structure and less computational time; the improvement percentage rangs from 16% to 32%. The forecasts accuracy obtained from the combination of boosting and BMA are improved 22% to 35%.