时空传播约束下洮河流域水沙过程预测研究OA

Traditional single-station modeling paradigms ignore the spatiotemporal propagation characteristics of watershed water and sediment processes,leading to insurmountable physical deficiencies.Taking the Taohe River Basin as a case study,this research develops a deep learning prediction framework that incorporates spatiotemporal propagation constraints for water and sediment processes,achieving a transition from single-point independent prediction to multi-point collaborative prediction.The Dempster-Shafer(D-S)evidence fusion method integrates the results of Pearson correlation coefficient(PCC)analysis,maximal information coefficient(MIC)mutual information,and recursive feature elimination(RFE)to select 6～7 core predictive factors from 10 candidate factors,achieving a dimensionality reduction rate of 30%～40%.A Stacking ensemble model based on multilayer perceptron(MLP),long short-term memory network(LSTM),and Transformer is constructed,with a three-layer physical constraint mechanism designed to include temporal propagation time constraints,spatial continuity constraints,and spatiotemporal water-sediment coupling constraints.A four-stage spatial propagation prediction chain is established connecting Xibagou Station,Minxian Station,Lijiacun Station,and Hongqi Station.The SHapley Additive exPlanations(SHAP)method is employed for multi-dimensional driving mechanism analysis to identify the spatiotemporal evolution patterns of key driving factors.The results demonstrate that:the Nash-Sutcliffe efficiency coefficient(NSE)for runoff prediction consistently exceeds 0.85,while NSE for sediment prediction reaches 0.782～0.815,which are increased by 18.2%and 21.0%respectively compared to traditional methods,achieving excellent prediction performance;the physical constraint mechanisms effectively ensure the rationality of prediction results with constraint violation rates below 2.7%;current-month precipitation,antecedent runoff,and temperature are the most important driving factors,though their importance exhibits significant spatiotemporal differentiation;the driving mechanism has shifted from natural dominance to artificial intervention,with climate factor weights decreasing from 92.4%to 85.8%and artificial factor weights increasing fourfold.This study provides novel theoretical methods for watershed water and sediment process prediction and holds significant scientific value for accurate and reliable long-term runoff and sediment forecasting as well as watershed water resources management.