|
| 耦合物理机制和深度学习的分布式混合洪水预报模型 |
| Distributed hybrid flood forecasting model coupling physical mechanisms and deep learning |
| 投稿时间：2025-04-15 修订日期：2025-06-29 |
| DOI： |
| 中文关键词: 混合模型 深度学习 可微分建模 洪水预报 |
| 英文关键词: Hybrid model Deep learning Differentiable modeling Flood forecasting |
| 基金项目:国家重点研发计划项目（2024YFC3212300）、国家自然科学基金项目（52479009）、水灾害防御全国重点实验室自主研究项目（5240152I2）、江苏省研究生科研与实践创新计划项目（KYCX25_0936） |
|
| 摘要点击次数: 97 |
| 全文下载次数: 0 |
| 中文摘要: |
| 准确的洪水预报是防洪减灾、水资源优化配置及水环境保护的关键。本文针对洪水预报中过程驱动模型物理机制受限与数据驱动模型可解释性不足的问题，提出了一种耦合物理机制和深度学习的分布式混合洪水预报建模（DHFM）框架。该框架在统一的深度学习平台上实现了物理过程和神经网络的协同编码，并基于反向传播算法实现物理参数与网络权值的同步优化。在此基础上提出了可以与DHFM框架无缝集成的可微分扩散波（DW）和卷积神经网络（CNN）河道汇流方法，同时引入可微分马斯京根（MK）方法作为对比基准。选取湖南省洣水流域为研究区域，系统评估了DHFM框架中三种可微分河道汇流方法在有资料和无资料场景下的性能及可解释性。结果表明，DHFM框架在所有场景下均展现出良好的日流量与洪水模拟能力。在模型其他结构及输入条件保持一致的情况下，可微分CNN河道汇流方法的性能略优于可微分DW河道汇流方法，且二者均优于可微分MK河道汇流方法。DHFM框架内嵌入的神经网络能够捕捉河道静态属性与物理参数的复杂映射关系，其中可微分CNN河道汇流方法还具备基于河道静态属性自适应学习单位线的潜力。文中DHFM框架不仅可以有效提高洪水模拟精度，还为深入理解洪水模拟的物理机制提供了新思路。 |
| 英文摘要: |
| Accurate flood forecasting is crucial for flood prevention and mitigation, optimal allocation of water resources, and protection of the water environment. Addressing the limitations of process-driven models in terms of physical mechanisms and the lack of interpretability in data-driven models for flood forecasting, this study proposes a Distributed Hybrid Flood Modeling (DHFM) framework that deeply integrates physical mechanisms with deep learning. By integrating physical constraints with neural networks on a unified deep learning platform, the framework enables collaborative encoding of hydrological processes and synchronous optimization of physical parameters and network weights via backpropagation. Building on this foundation, the DHFM framework seamlessly incorporates differentiable Diffusion Wave (DW) and Convolutional Neural Network (CNN)-based river routing methods, while integrating the differentiable Muskingum (MK) method as a comparative baseline. Taking the Mishui River Basin in Hunan Province as a typical study area, we systematically evaluate the performance and interpretability of the three differentiable river routing methods within the DHFM framework under both gauged and ungauged scenarios. The results indicate that the DHFM framework demonstrates robust capabilities in simulating daily flows and floods across all scenarios. With other model structures and input conditions held constant, the differentiable CNN river routing method slightly outperforms the differentiable DW method, and both outperform the differentiable MK method. Furthermore, the neural networks embedded within the DHFM framework can capture complex mapping relationships between static river attributes and physical parameters, with the differentiable CNN river routing method also showing potential for adaptively learning unit hydrographs based on static river attributes. Beyond enhancing simulation accuracy, the DHFM framework offers a novel paradigm for bridging physical mechanisms and data-driven modeling in flood research. |
|
查看/发表评论 下载PDF阅读器 |
| 关闭 |