Abstract: Runoff forecast provides a basis for the rational utilization and distribution of river basin water resources. This paper presents multiple linear regression, principal component regression, BP neural network model and a new model which combining the principal component analysis with the BP neural network. And those methods are used to predict the monthly runoff of the Hutubihe River in 2009-2011 collected at the Shimen Hydrological Station of Xinjiang. The prediction accuracy of each model compared by correlation coefficient, determination coefficient and root mean square error. The results show that: (1) Intelligent algorithm such as neural network has the ability of optimization in high speed, which gets better result on short time scales of monthly runoff forecast; (2) Conventional algorithm such as principal component regression can fully reflect the stability of the annual runoff in a given area, the simulation accuracy of total annual runoff is relatively higher than other methods; (3) The method combining the principal component analysis with the BP neural network which can improve the convergence speed of neural networks, while reduces the impact of local extremum, is better than simple BP neural network, and it is suitable for the monthly runoff forecast for the Hutubi River.

Key words: principal component regression, principal component analysis(PCA), BP neural network model, runoff forecast