A Novel mRMR-RFE-RF Method for Enhancing Medium- and Long-Term Hydrological Forecasting: A Case Study of the Danjiangkou Basin

Abstract

In machine learning (ML)-based hydrological forecasting, particularly in medium- and long-term prediction, judicious predictor selection is paramount, as it ultimately determines the forecast accuracy. This study pioneered an advanced predictor-screening method that synergizes the mutual information (MI) and random forest (RF) technologies through minimum-redundancy-maximum-relevance-recursive feature elimination-random forest (mRMR-RFE-RF) method, blending both filtering and wrapping techniques. This method was rigorously tested through a detailed case study in the Danjiangkou basin, where a comprehensive analysis of 1560 meteorological factors was conducted. Employing three sophisticated ML algorithms—RF, eXtreme Gradient Boosting (XGB), and Light Gradient Boosting (LGB)—we developed precipitation forecasting models. Furthermore, we performed an in-depth rationality analysis of high-frequency predictors. The findings from our study show that this novel hybrid screening strategy markedly outperformed conventional singular predictor-screening methods in enhancing the accuracy of precipitation forecasting when integrated into these forecasting models. Moreover, it assured the validity of the high-frequency forecast factors employed. Therefore, this innovative method not only elevates the accuracy of medium- and long-term precipitation forecasting but also contributes a novel perspective to the methodology of predictor selection in hydrological forecasting models.