Abstract: Abstract: Crop growth simulation models can simulate the processes of crop growth, development, yield formation, and its response to environment, which provides an effective method for crop yield forecast. However, how to select suitable weather data for the forecast is still an open question. In this study, we established a method for maize yield forecast based on maize growth simulation model of CERES-Maize and historical weather data from the year of 1956 to 2015. Two year's experimental data from 3 sites of Yangling (2014 and 2015), Heyang (2009 and 2011) and Changwu (2010 and 2011) in Shaanxi Province were used to test the reliable and accuracy of the method established. The weather data needed for model simulation were divided into 2 different groups including the known weather data and unknown weather data during the whole growth season of spring maize. The known weather data were obtained from local weather stations, while unknown data were supplemented with historical weather data of multiple years in the local experimental sites. Multiple complete climatic data series were then created and used to run the CERES-Maize model to forecast maize yield for a given year. As the advancing of maize growth season, the daily weather data were gradually merged into the observed weather data in a target year. Consequently, the daily maize yield was forecasted from sowing day to harvest. In addition, in order to reduce the times of model runs and reduce the uncertainties in yield forecasts, this study compared the daily meteorological data of historical and target years with normal K nearest neighbor (K-NN) and a modified K-NN algorithm to select several historical analogue years whose weather data were similar to the target year. The results showed that: 1) the model was suitable for the yield simulation since the absolute relative error was smaller than 15%; 2) the data distribution of predicted yields began to converge and the uncertainty decreased rapidly after the tasseling stage. For example, the predicted yield after 30, 60 and 90 days (the tasseling stage) of sowing was 3 531-14 461, 3 413-14 828 and 961-13 210 kg/hm2, respectively. But, the yield was 49 33-10 826, 8 484-10 565 kg/hm2, respectively after 100 and 130 days of sowing. The coefficient of variation had a sudden fall around the tasseling stage; 3) Yield forecast accuracy was generally lower than expectation for the method based on all historical data and climatic analogue years selected with historical data. The model run cost 61 min for a yield prediction during a complete growth stage of spring maize, indicting a necessary change in the prediction method optimization; 4) Among the 3 methods, the modified K-NN method showed a higher prediction accuracy and shorter run time than the other methods. The coefficient of variation was 11.7%-23.8% for the modified K-NN method, 15.1%-29.1% for the historical data, and 14.7%-26.9% for the K-NN method, respectively. To complete the yield prediction of a growth stage of spring maize, the modified K-NN method only took 14 min, which was shorter than the normal K-NN method. Thus, the modified K-NN method in this study had a big potential for the yield prediction by the CERES-Maize model. The study provides an effective method for selecting precipitation factor used for the yield prediction by crop models.