水电能源科学

利用数值天气预报(NWP)产品作为模型输入进行洪水预报，可以显著延长洪水预报的预见期。若直接将未经校正的预报降水数据用于水文模型，会引入额外误差，进而影响洪水预报精度及预警效果。为校正NWP数据的偏差从而提升洪水预报的准确性，提出一种Stacking集成方法对预报降水进行统计后处理，该方法分为2层，第1层采用RF、KNN、XGB、LGB、Catboost、GBM共6种机器学习模型对欧洲中期天气预报中心(ECMWF)提供的预报降水产品分别进行校正，然后在第2层构建RF模型将6种机器学习校正结果进行集成，得到最终的校正预报降水数据。在山东省大汶河流域的应用结果表明，该方法能够有效融合多个机器学习模型的优势，优于单一机器学习模型，校正后降水的均方根误差从13.75 mm降至9.33 mm,相关系数由0.41提高至0.78,相对偏差由-10.95%减小至-1.40%,同时更准确地反映了降水强度的梯度特征。

Using Numerical Weather Prediction(NWP) products as model inputs for flood forecasting can significantly extend the lead time of flood forecasts. However, directly using uncorrected forecast precipitation data in hydrological models can cause additional errors, thereby affecting the accuracy of flood forecasts and the effectiveness of early warnings. Therefore, correcting the bias of NWP data is crucial for improving the accuracy of flood forecasting. This study proposes a Stacking ensemble method for statistical post-processing of forecast precipitation. The method consists of two layers. In the first layer, six machine learning(ML) models, namely RF, KNN, XGB, LGB, Catboost, and GBM, are used to correct the forecast precipitation products provided by the European Centre for Medium-Range Weather Forecasts(ECMWF) respectively. Then, in the second layer, a random forest model is constructed to integrate the correction results of the six ML models, obtaining the final corrected forecast precipitation data. The application results in the Dawenhe Basin of Shandong Province show that this method can effectively integrate the advantages of multiple ML models, outperforming a single ML model. The root mean square error of the corrected precipitation decreased from 13.75 mm to 9.33 mm, the correlation coefficient increased from 0.41 to 0.78, and the relative bias decreased from-10.95% to-1.40%. At the same time, it more accurately reflects the gradient characteristics of precipitation intensity.