Scholarlink

<p>Accurate streamflow prediction is essential for flood mitigation and water-resources planning but remains challenging due to nonlinear hydrological processes and limited data availability across large and complex basins. Traditional physics-based models and data-driven models often struggle to capture spatial variability and extreme peak flows. This study proposes a unified Enhanced LSTM (En-LSTM) framework that augments a conventional LSTM with Temporal Convolutional Networks (TCN), an attention mechanism, a peak-aware hybrid loss function, and multi-scale lag and rolling-window features, while explicitly incorporating hydrologically meaningful predictors for the Mahanadi River Basin in India. Model shows substantial improvements over a conventional LSTM(i), achieving KGE values of 0.84 in training and 0.71 in testing, and demonstrates strong transferability at completely unseen stations, where En-LSTM(i) attains NSE ≈ 0.79 and KGE ≈ 0.76, indicating robust performance in data-scarce locations. Furthermore, En-LSTM(i) exhibits marked improvements in peak-flow simulation relative to both the SWAT + process-based model and the conventional LSTM, thereby addressing a well-known weakness of many existing hydrological modeling approaches. These results position En-LSTM(i) as a robust and transferable framework for operational hydrological forecasting and basin-scale streamflow assessment, with strength in capturing extreme events and spatially variable responses.</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

Improving Daily Streamflow Predictions over Large Watersheds: Introducing a Novel Enhanced Long Short-Term Memory (En-LSTM) Model

  • 期刊论文

  • 发表日期：

    2026年1月24日

查看全文
  • Purshottam Sharma,
  • Vaibhav Tripathi,
  • Mohit Prakash Mohanty

摘要

Accurate streamflow prediction is essential for flood mitigation and water-resources planning but remains challenging due to nonlinear hydrological processes and limited data availability across large and complex basins. Traditional physics-based models and data-driven models often struggle to capture spatial variability and extreme peak flows. This study proposes a unified Enhanced LSTM (En-LSTM) framework that augments a conventional LSTM with Temporal Convolutional Networks (TCN), an attention mechanism, a peak-aware hybrid loss function, and multi-scale lag and rolling-window features, while explicitly incorporating hydrologically meaningful predictors for the Mahanadi River Basin in India. Model shows substantial improvements over a conventional LSTM(i), achieving KGE values of 0.84 in training and 0.71 in testing, and demonstrates strong transferability at completely unseen stations, where En-LSTM(i) attains NSE ≈ 0.79 and KGE ≈ 0.76, indicating robust performance in data-scarce locations. Furthermore, En-LSTM(i) exhibits marked improvements in peak-flow simulation relative to both the SWAT + process-based model and the conventional LSTM, thereby addressing a well-known weakness of many existing hydrological modeling approaches. These results position En-LSTM(i) as a robust and transferable framework for operational hydrological forecasting and basin-scale streamflow assessment, with strength in capturing extreme events and spatially variable responses.