Scholarlink

<p>Streamflow is crucial for evaluating regional water resources and planning, with accurate predictions essential for assessing climate change impacts. While deep learning models show promise, many lack interpretability and fail to establish causal relationships, reducing confidence in their predictions. To improve prediction accuracy and interpretability, we propose three Transformer-integrated models (TCN-Transformer, KAN-Transformer, xLSTM-Transformer), coupled with the SWAT hydrological model. These models uncover the complex relationships among hydrological attributes, meteorological features, and streamflow. Our results show these models outperform the benchmark SWAT model, with <i>PBIAS</i> values under 10%, <i>NSE</i> between 0.76 and 0.86, and <i>R</i><sup><i>2</i></sup> from 0.81 to 0.90. SHAP analysis identifies precipitation as the key driver of streamflow, with groundwater discharge and surface runoff also playing significant roles. The SWAT-derived hydrological features were most influential in the KAN-Transformer model. This study demonstrates that integrating SWAT with Transformer models enhances model reliability, supporting better water resource planning.</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

Coupling SWAT and Interpretable Deep Learning to Improve Streamflow Simulation in Arid Regions

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Qingyang Zhang,
  • Junhu Wu,
  • Tianle Wang,
  • Chenzhan Tang,
  • Weibin Liu

摘要

Streamflow is crucial for evaluating regional water resources and planning, with accurate predictions essential for assessing climate change impacts. While deep learning models show promise, many lack interpretability and fail to establish causal relationships, reducing confidence in their predictions. To improve prediction accuracy and interpretability, we propose three Transformer-integrated models (TCN-Transformer, KAN-Transformer, xLSTM-Transformer), coupled with the SWAT hydrological model. These models uncover the complex relationships among hydrological attributes, meteorological features, and streamflow. Our results show these models outperform the benchmark SWAT model, with PBIAS values under 10%, NSE between 0.76 and 0.86, and R2 from 0.81 to 0.90. SHAP analysis identifies precipitation as the key driver of streamflow, with groundwater discharge and surface runoff also playing significant roles. The SWAT-derived hydrological features were most influential in the KAN-Transformer model. This study demonstrates that integrating SWAT with Transformer models enhances model reliability, supporting better water resource planning.