Scholarlink

<p>Differentiable modeling techniques have enabled end-to-end calibration of hydrological models and facilitated their integration with data-driven approaches. However, the potential of differentiable parameter learning across diverse climatic conditions remains underexplored, and a systematic evaluation of differentiable module replacement strategies remain limited. To address this, we develop a differentiable parameter learning model (D-HBV) based on the HBV-96 framework, in which Temporal Convolutional Network (TCN) and Convolutional Long Short-Term Memory (ConvLSTM) modules are incorporated to replace the runoff generation and routing components, respectively. Three hybrid models (D-T-M, D-H-C, and D-T-C) are constructed and applied to river basins spanning ten global climate types. Compared with the traditional HBV model, the D-HBV model improves <i>NSE</i> and <i>KGE</i> by 37.3% and 46.9%, respectively. The hybrid models achieve even greater gains, with D-T-C performing best, increasing <i>NSE</i> and <i>KGE</i> by 45.1% and 65.3%, and reducing <i>RMSE</i> by 41.4%, leading to more reliable reproduction of flood peaks, recession dynamics, and low-flow conditions. Interpretability analyses reveal that the TCN and ConvLSTM modules effectively capture nonlinear runoff responses and routing delays with strong physical consistency, allowing the model to reflect basin-specific runoff formation and flow concentration behaviors. The performance improvements of the hybrid models stem from their structural adaptability to basin-scale hydrological processes, with notable variation across climate zones. These findings underscore the importance of climate-aware neural integration design and demonstrate a promising pathway toward robust, transferable, and physically consistent streamflow simulation in a changing environment.</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

Streamflow Simulation and Interpretability Analysis in Multi-Climatic Basins Using Physics-Based and Data-Driven Hybrid Models

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Jijun Chen,
  • Jiali Guo,
  • Jiahui Gu,
  • Kunteng Zhang

摘要

Differentiable modeling techniques have enabled end-to-end calibration of hydrological models and facilitated their integration with data-driven approaches. However, the potential of differentiable parameter learning across diverse climatic conditions remains underexplored, and a systematic evaluation of differentiable module replacement strategies remain limited. To address this, we develop a differentiable parameter learning model (D-HBV) based on the HBV-96 framework, in which Temporal Convolutional Network (TCN) and Convolutional Long Short-Term Memory (ConvLSTM) modules are incorporated to replace the runoff generation and routing components, respectively. Three hybrid models (D-T-M, D-H-C, and D-T-C) are constructed and applied to river basins spanning ten global climate types. Compared with the traditional HBV model, the D-HBV model improves NSE and KGE by 37.3% and 46.9%, respectively. The hybrid models achieve even greater gains, with D-T-C performing best, increasing NSE and KGE by 45.1% and 65.3%, and reducing RMSE by 41.4%, leading to more reliable reproduction of flood peaks, recession dynamics, and low-flow conditions. Interpretability analyses reveal that the TCN and ConvLSTM modules effectively capture nonlinear runoff responses and routing delays with strong physical consistency, allowing the model to reflect basin-specific runoff formation and flow concentration behaviors. The performance improvements of the hybrid models stem from their structural adaptability to basin-scale hydrological processes, with notable variation across climate zones. These findings underscore the importance of climate-aware neural integration design and demonstrate a promising pathway toward robust, transferable, and physically consistent streamflow simulation in a changing environment.