Abstract: To improve the accuracy of the medium-term runoff forecasting of the watershed, a multi-model fusion method of medium-term runoff prediction method based on machine learning is proposed, which is applied to the Huanren Reservoir Basin. Firstly, BP neural network model, multiple linear regression model, and support vector machine model are used to forecast the ten-day runoff. The above models are merged based on information entropy, BP neural network, and SVM. Then, a ten-day runoff forecasting model for the spring flooding period that considers the snowmelt effect is constructed.Three error evaluation indicators, MAE, RMSE and QR, are introduced to evaluate the runoff forecasting accuracy of each model in the flood and non-flood periods. The results in Huanren Reservoir show that all models can accurately simulate the runoff change process, but the single forecasting models are poorly fitted in the peak section. The information fusion models based on machine learning algorithms can well combine the advantages of different forecasting models, and outperform each single forecasting model and the fusion model based on information entropy, which can improve the runoff forecast accuracy for 10 periods throughout the year and increase the forecast qualification rate to 100% for 6 periods with a maximum increment of 24%. The ten-day runoff forecasting model considering the effect of snowmelt has a passing rate of over 90% in both March and April, improving the non-flood runoff forecasting capability of the basin.The proposed information fusion prediction method based on machine learning can obtain runoff prediction models with high accuracy and reliability, providing data support and theoretical support for runoff forecasting work and efficient water resources management in the Huanren Reservoir.