In this paper, the characteristics of flood forecasting by Xin'anjiang model were studied in Chenhe Basin and Daheba Basin, in order to solve the problem of rainfall uncertainty in real-time flood forecasting. Four methods, namely considering forecasting precipitation, discarding forecasting precipitation, adding rainfall forecasting error and adding rainfall time error, are applied to analyze the influences of prediction accuracy resulting from different forecasting periods and uncertainty of rainfall input. The results show that:(1) compared with discarding forecasting precipitation in the future, the accuracy of flood prediction has a significant enhancement by considering the forecasting precipitation in a long forecast period while a slight improvements in a short forecast period; (2) the central location of rainfall has a great influence on prediction accuracy, and the allowable error of rainfall center in the upstream is greater than that in the downstream; (3) the greater the forecast rainfall forecast error, the lower the forecast accuracy; (4) the rainfall error has liner influences on the flood volume error and flood peak error, but nonlinear influences on the Nash-Sutcliffe efficiency. In addition, the flood peak time error is mainly affected by the rainfall time error.