基于误差修正的月径流集成预测——以珠江源头南盘江流域为例OA

The mechanisms governing medium-and long-term runoff are inherently complex,rendering errors in deterministic prediction models inevitable.By extracting and utilizing the runoff-relevant information contained within the error residuals,it is possible to further improve the prediction accuracy of the models.This study pro-poses a monthly runoff ensemble prediction model based on long short-term memory(LSTM)neural network with er-ror correction using extreme gradient boosting(XGBoost).Information flow analyses were constructed in three direc-tions:upstream-midstream,midstream-downstream,and upstream-downstream.Meteorological factors highly corre-lated with monthly runoff were selected as model input variables using the Pearson correlation coefficient.Runoff predictions were generated using the LSTM model,and the resulting prediction errors were subsequently predicted using XGBoost.The SHAP method was employed to analyze the error series at three representative stations in the study area—the Nanpan River Basin—namely,the Xiqiao,Xiaolongtan,and Tianshengqiao stations,located in the upstream,midstream,and downstream sections,respectively.The results indicated that meteorological factors such as temperature,ground temperature,and atmospheric pressure—which were not utilized in the primary prediction—had their inherent characteristics effectively exploited during the secondary error prediction stage.Finally,the XGBoost-predicted errors were added to the LSTM prediction results to generate the final prediction from the LSTM-XGBoost ensemble model.In the runoff predictions for the upstream,midstream,and downstream sections of the Nanpan River Basin,the Nash-Sutcliffe efficiency coefficients of the LSTM-XGBoost ensemble model reached 0.82,0.81,and 0.82,respectively.Compared with the random forest model,BP neural network model,standalone LSTM model,and standalone XGBoost model,the proposed LSTM-XGBoost ensemble model exhibited superior pre-diction accuracy and reliability.