By establishing flood forecasting models based on the standard BP algorithm, the Levenberg-Marquart algorithm and the genetic algorithm respectively, the paper studies the floods of the Zhouhe River in Dazhou city of Sichuan province, and analyzes and compares the outcomes of these three models. It shows that the model based on the genetic algorithm has the fastest convergence rate and highest accuracy in flood forecasting and can effectively prevent the network from getting into local minimum point.