In order to improve accuracy of rivers runoff forecast, a new method combining the rescaled range(R/S)analysis with the grey theory(GM)was proposed, based on that the R/S analysis provides scientific and effective non-linear prediction and rivers runoff has a grey character. This new method using the R/S-GM(1, 1)model can overcome the defects of low prediction accuracy when there is a larger fluctuation in data. Then it was applied in the runoff forecast of the Yingluoxia and the Zhengyixia stations in the Heihe River, with the use of six series of the annual runoff, flood season and non-flood season runoff. It is validated that the prediction accuracy of annual runoff and flood season runoff series are about 90%, and those of non-flood season runoff series are more than 80%, and the forecast results of the runoff series are consistent with the Mann-Kendall trend test results. Hence, the predication result is reliable, providing a new approach for accurate runoff forecast.