Based on the measured data of inflow annual runoff to Yuanyangchi Reservoir from 1959 to 2015, the interannual variation and annual distribution characteristics of runoff were analyzed using the mean decadal runoff and the concentration degree of annual runoff. The basic GM (1, 1), improved GM (1, 1) and R/S analysis combined with the GM (1, 1) grey model were used to predict the annual runoff. The results show that the time series of annual runoff has obvious fractal characteristics, and the calculated H-index is 0.922. The runoff time series has state persistence and long-term memory, which means that the runoff changes in the future are consistent with the past trends. The calculated long-term memory of the runoff series is a 41-year cycle. Based on the R/S analysis, the basic GM(1,1) gray model predicts that the annual runoff are 3.60×108 m3, 2.97×108 m3, and 3.67×108 m3 in 2013, 2014, and 2015, respectively. The relative errors are 20.69%, 8.97%, and 8.10%, respectively compared to the measured values, which shows the accuracy is higher than that of the basic GM(1,1) and improved GM(1,1) gray models. The predicted inflow annual runoff to reservoir in 2020 is 8.99 % higher than that in 2015.