To provide more reliable simulations and forecasts using the Xinanjiang model in the semi-humid watersheds, this study introduced three real-time correction methods into the flood forecasting, respectively, including the K-nearest neighbor algorithm(the KNN method), the traditional error autoregression method(the AR method)and the simulating feedback method. The Chenhe Basin, in Shaanxi Province, was selected as the experimental basin. Considering the relative error of flood peak and the coefficient of Nash-Sutcliffe efficiency as evaluation indicators, this study analyzed the results of three correction methods. The results show that all three kinds of correction methods can improve the coefficient of Nash-Sutcliffe efficiency and the simulating feedback method was optimal on the Nash-Sutcliffe efficiency coefficient, while the AR method and the KNN method were the second best. The simulating feedback method allow a remarkable improvement compared with the KNN method in a short forecast period, and both of them can effectively avoid the defect of the AR method in terms of the error correction of flood peak. The results also indicate the KNN method with historical samples yielded better results than the simulation of feedback method on the error correction of flood peak, which can effectively improve the accuracy of flood forecasting.