Scholarlink

<p>Accurate monthly streamflow prediction is critical for effective flood mitigation and water resource management. This study presents a novel approach that incorporates penalty terms over statistical features of input data into the loss functions of two models, LSTM-P and ANN-P, aiming to improve the predictive accuracy of monthly streamflow models during testing periods. Four specific penalty types were proposed: minimum boundary, maximum boundary, mean interval, and standard deviation interval penalties. Using historical monthly streamflow data from a hydrological station in China, the study analyzes to determine the optimal weights for each penalty and tests combinations to assess their collective impact on model performance. Comparative analysis under different penalty conditions reveals that incorporating statistical feature-based penalties during training improves predictive accuracy and enhances consistency in performance between training and testing periods—an outcome rarely achieved in previous approaches.</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

Enhancing Monthly Streamflow Prediction with LSTM-P and ANN-P Models using Statistical Feature-Based Penalty Factors

  • 期刊论文

  • 发表日期：

    2025年4月30日

查看全文
  • Zifan Xu,
  • Hao Zheng,
  • Hong Zhang,
  • Xuguang Wang,
  • Xinzhe Xu,
  • Peng Liu,
  • Suzhen Feng,
  • Jinwen Wang

摘要

Accurate monthly streamflow prediction is critical for effective flood mitigation and water resource management. This study presents a novel approach that incorporates penalty terms over statistical features of input data into the loss functions of two models, LSTM-P and ANN-P, aiming to improve the predictive accuracy of monthly streamflow models during testing periods. Four specific penalty types were proposed: minimum boundary, maximum boundary, mean interval, and standard deviation interval penalties. Using historical monthly streamflow data from a hydrological station in China, the study analyzes to determine the optimal weights for each penalty and tests combinations to assess their collective impact on model performance. Comparative analysis under different penalty conditions reveals that incorporating statistical feature-based penalties during training improves predictive accuracy and enhances consistency in performance between training and testing periods—an outcome rarely achieved in previous approaches.