The accurate analysis and reasonable prediction of rule of runoff variation have a great significance for reasonable development of water resources, construction of water conservancy project and development of social economy. Using Mann-Kendall method and wavelet analysis method, the paper analyzed and researched the variety regulation of runoff in Tuwei river basin,and established the neural network model to forecast the runoff variation. The results show that the runoff in Tuwei river basin has a obvious decreasing trend; from the wavelet coefficient chart we can know that the annual runoff process primarily have the period of 2 years, 8 years and 19 years, of which about time scale of 19 years as the first cycle, also found that the current annual runoff is in the late dry season, the amount of water has trend of rising; by using the neural network model to forecast the annual runoff at Gaojiachuan station in Tuwei river basin,the relative error of prediction is only 5.92%,which shows that the forecast precision of annual runoff in the basin by neural network model is higher and a very effective method of annual runoff forecast.