Medium- and Long-Term Hydrological Forecasting Using Mutual Information and Random Forest Predictors Selector

The screening of predictors is a necessary step to implement a medium- and long-term hydrological forecasting model based on machine learning, which determines the information input by the forecast model and directly affects the forecasting accuracy. In this paper, a new method of predictors screening based on mutual information and random forest is proposed (mRMR-RFE-RF), which combines the filtering method and the wrapping method. A case study is carried out in the Danjiangkou basin. After selection of 1560 meteorological factors, precipitation forecast models based on three machine learning methods (RF, XGB and LGB) are constructed. The comparison of the precipitation forecast accuracy with that of the single predictors screening method shows that the predictors screened by the proposed new hybrid method have higher precipitation forecast accuracy and better effect after being input into the three machine learning precipitation forecast models. The proposed method effectively improves the accuracy of medium and long-term precipitation forecasting, and provides a new idea for the predictors selection method for medium and long-term hydrological forecasting.