Scholarlink

<p>The complex processes of runoff pose significant challenges for accurate prediction. Traditional static input models, which rely on a single input factor scheme, cannot effectively capture the dynamic characteristics of streamflow changes, leading to reduced predictive accuracy and interpretability. This study conducted a two-stage clustered dynamic input framework, providing a more accurate and reliable approach to predict complex streamflow series. In the first clustering stage, the streamflow series were divided into independent time series based on the non-flood season, flood season, and different months. Historical streamflow factors with different lags were analyzed in each independent series using a factor importance ranking method. Subsequently, based on the factor importance ranking, each factor was gradually input into a kernel-based machine learning model to optimize the input scheme for each independent series, thereby obtaining initial prediction results. The higher the prediction accuracy, the more precisely it capture the characteristics of the original streamflow. Therefore, in the second clustering stage, K-means was used to cluster the optimal prediction from the first stage, capturing the magnitude characteristics of the original streamflow series. The optimal input factor scheme for each cluster was then selected to correct the prediction results. Monthly streamflow series from hydrological stations in the upper and lower reaches of the Fenhe River basin were used as a case study. The prediction results indicated that dynamically adjusting factors in the monthly clustering scenarios during the first stage improved prediction performance compared to using static input models. The Kling-Gupta efficiency (KGE) value exceeded 0.717. After a second clustering of the streamflow prediction results and adjustment of input factors, the KGE increased by 4.284% and 6.416% compared to the uncorrected predictions. These results demonstrate the effectiveness of two-stage clustered dynamic input framework in enhancing predictive capacity for complex streamflow series.</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

Advancing complex streamflow prediction through a two-stage clustering framework and dynamic input integration

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Qiucen Guo,
  • Xuehua Zhao,
  • Bowen Zhu,
  • Xueping Zhu

摘要

The complex processes of runoff pose significant challenges for accurate prediction. Traditional static input models, which rely on a single input factor scheme, cannot effectively capture the dynamic characteristics of streamflow changes, leading to reduced predictive accuracy and interpretability. This study conducted a two-stage clustered dynamic input framework, providing a more accurate and reliable approach to predict complex streamflow series. In the first clustering stage, the streamflow series were divided into independent time series based on the non-flood season, flood season, and different months. Historical streamflow factors with different lags were analyzed in each independent series using a factor importance ranking method. Subsequently, based on the factor importance ranking, each factor was gradually input into a kernel-based machine learning model to optimize the input scheme for each independent series, thereby obtaining initial prediction results. The higher the prediction accuracy, the more precisely it capture the characteristics of the original streamflow. Therefore, in the second clustering stage, K-means was used to cluster the optimal prediction from the first stage, capturing the magnitude characteristics of the original streamflow series. The optimal input factor scheme for each cluster was then selected to correct the prediction results. Monthly streamflow series from hydrological stations in the upper and lower reaches of the Fenhe River basin were used as a case study. The prediction results indicated that dynamically adjusting factors in the monthly clustering scenarios during the first stage improved prediction performance compared to using static input models. The Kling-Gupta efficiency (KGE) value exceeded 0.717. After a second clustering of the streamflow prediction results and adjustment of input factors, the KGE increased by 4.284% and 6.416% compared to the uncorrected predictions. These results demonstrate the effectiveness of two-stage clustered dynamic input framework in enhancing predictive capacity for complex streamflow series.