Scholarlink

<p>Runoff simulation is crucial for water resource management. However, existing research on integrating process-driven hydrological models with machine learning mainly focuses on enhancing simulation accuracy, while the roles of hydrological models with different levels of physical mechanism within coupling frameworks, as well as the influence of model structure on simulation stability, have not been systematically evaluated. To address this issue, this study takes the Yalong River Basin as the study area to investigate whether hydrological models with stronger physical mechanisms can enhance runoff simulation performance of data-driven models within a unidirectional coupling framework, and further examines the effect of introducing a Stacking structure on simulation robustness. The distributed hydrological models CREST and VIC, as well as the lumped hydrological model Xinanjiang, are employed to generate simulated runoff series, which are used as physically constrained inputs for the LSTM to construct a unidirectional coupling framework. In addition, the Stacking structure is introduced by combining simulated runoff from two hydrological models as inputs to the LSTM to examine the robustness of runoff simulation. Results show that models with stronger physical mechanism perform better in the coupling framework, with their average NSE, R², and KGE higher by 0.03, 0.03, and 0.01, respectively. Under the Stacking framework, the mean RMSE and Pbias of the Annual Maximum Daily Runoff are 1437.74&#xa0;m³/s and − 16.41%, showing no clear improvement over the Original models. However, Stacking reduces uncertainty while maintaining accuracy and improves consistency and stability in runoff distribution and temporal variation.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exploration on Coupling Machine Learning with Hydrological Model to Enhance Runoff Simulation

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Yinmao Zhao,
  • Ningpeng Dong,
  • Chao Ma,
  • Hao Wang

摘要

Runoff simulation is crucial for water resource management. However, existing research on integrating process-driven hydrological models with machine learning mainly focuses on enhancing simulation accuracy, while the roles of hydrological models with different levels of physical mechanism within coupling frameworks, as well as the influence of model structure on simulation stability, have not been systematically evaluated. To address this issue, this study takes the Yalong River Basin as the study area to investigate whether hydrological models with stronger physical mechanisms can enhance runoff simulation performance of data-driven models within a unidirectional coupling framework, and further examines the effect of introducing a Stacking structure on simulation robustness. The distributed hydrological models CREST and VIC, as well as the lumped hydrological model Xinanjiang, are employed to generate simulated runoff series, which are used as physically constrained inputs for the LSTM to construct a unidirectional coupling framework. In addition, the Stacking structure is introduced by combining simulated runoff from two hydrological models as inputs to the LSTM to examine the robustness of runoff simulation. Results show that models with stronger physical mechanism perform better in the coupling framework, with their average NSE, R², and KGE higher by 0.03, 0.03, and 0.01, respectively. Under the Stacking framework, the mean RMSE and Pbias of the Annual Maximum Daily Runoff are 1437.74 m³/s and − 16.41%, showing no clear improvement over the Original models. However, Stacking reduces uncertainty while maintaining accuracy and improves consistency and stability in runoff distribution and temporal variation.

Graphical Abstract