耦合知识嵌入与数据驱动的黄河径流量预测OA

[Objective]This study constructs a data-coupled runoff prediction model that integrates hydrological physical principles with deep learning,improves the predictive accuracy and physical credibility of data-driven methods under extreme hydrological events,and provides practical guidance for refined runoff prediction and intelligent management of water re-sources in the Yellow River Basin.[Methods]Taking the Tangnaihai hydrological station in the upper reaches of the Yellow River as the study area,a technical framework of "hydrological prior extraction—knowledge embedding—deep learning predic-tion—model validation" was established with runoff prediction as the core task.The joint probability density distribution of hydrometeorological factors,including precipitation,relative humidity,and sunshine duration,was used to extract prior knowledge characterizing the correlations in hydrological processes.This prior knowledge was then embedded into a temporal convolutional network(TCN)to construct a knowledge-guided runoff prediction model.Comparative experiments with a base-line TCN model were conducted to systematically evaluate the impact of knowledge embedding on runoff prediction perform-ance.[Results](1)Knowledge embedding significantly improved the model's ability to fit peak flows,effectively alleviating the systematic bias of purely data-driven models under extreme hydrological events.(2)The coupled model achieved increa-ses of 1.71%and 2.75%in the Nash-Sutcliffe efficiency(NSE)on the training and testing sets,respectively,indicating that the generalization capability of the model was significantly enhanced.(3)The joint probability density constraints enhanced the physical consistency of the prediction results while preserving the nonlinear fitting advantage of deep learning.(4)Com-pared with the baseline model,the coupled model showed better performance in overall stability and responsiveness under ex-treme scenarios.[Conclusions]Knowledge embedding effectively compensates for the lack of physical constraints in data-driven runoff prediction models,significantly improving both prediction accuracy and physical credibility.In the future,by incoporating multi-source data and optimized model structure,the proposed approach can further expand its application poten-tial in different river basins and real-time prediction scenarios.