| 摘要: |
| 为了探究概念性水文模型（GR4J）与长短时记忆模型（LSTM）耦合过程中物理模型参数反馈机制以及考虑土壤含水量作为中间变量对物理引导机器学习（PIML）模型径流模拟的影响，本研究构建了PIML模型并设置考虑参数反馈、考虑中间变量和两者同时考虑的三种方案（依次简称为H1、H2、H3），以安康站为控制站的汉江上游流域开展实例研究。结果表明：（1）三种PIML模型径流模拟效果均优于LSTM模型，其中验证期NSE平均提升10.6%，而PIML-H1与PIML-H3模型径流模拟效果优于GR4J模型，验证期NSE平均提升4.2%，其中PIML-H3模型在三种PIML模型中表现最佳，表明基于LSTM和GR4J模型构建的PIML模型对径流模拟效果有所改善，且同时考虑中间变量和参数反馈因素对径流模拟效果改善最佳；（2）三种PIML模型对低水流量的模拟效果均优于GR4J和LSTM模型，且PIML-H3模型效果最佳，而对于高水流量，三种PIML模型均表现不佳，说明PIML模型往往在模拟低流量事件中更占优势；（3）三种PIML模型的径流模拟效果均表现出较强的季节性变化，PIML-H2和PIML-H3模型相较于PIML-H1模型受到的季节性变化影响更为明显，说明PIML模型模拟径流结果的季节性变化受到中间变量的影响。研究可为汉江上游流域径流模拟预报提供技术支撑。 |
| 关键词: 物理引导机器学习 径流模拟 中间变量 GR4J LSTM |
| DOI： |
| 分类号: |
| 基金项目:国家重点研发计划资助项目、国家自然科学基金项目 |
|
| Runoff simulation of physics-informed machine learning model in the upper Han River Basin |
|
DENG Chao1, SUN Peiyuan1, YIN Xin2, ZOU Jiacheng3, WANG Weiguang1
|
|
1.The National Key Laboratory of Water Disaster Prevention,Hohai University;2.The National Key Laboratory of Water Disaster Prevention,Nanjing Hydraulic Research Institute;3.Hydrology and Water Resources Monitoring Center of Lower Ganjiang River
|
| Abstract: |
| This study investigates the impact of coupling the conceptual hydrological model (GR4J) with the Long Short-Term Memory model (LSTM) in a Physics-Informed Machine Learning (PIML) framework for runoff simulation. Three scenarios, denoted as H1, H2, and H3 respectively, are designed to examine the effects of the physical model parameter feedback mechanism, the consideration of soil moisture as an intermediate variable and the former both on the PIML models, respectively. The case study is conducted in the upper Han River Basin, with the Ankang hydrological station as the control station. The main findings are as follows: (1) All three PIML models demonstrate improved runoff simulation performance compared to the LSTM model, with a 10.6% increase in average Nash-Sutcliffe efficiency (NSE) during the validation period. Additionally, both the PIML-H1 and PIML-H3 models exhibit superior performance to the GR4J model, with a 4.2% increase in average NSE during validation. Notably, the PIML-H3 model outperforms the other PIML models, indicating that coupling GR4J and LSTM models while simultaneously considering intermediate variables and parameter feedback yields the most significant improvement in runoff simulation effectiveness. (2) For low flows, all three PIML models outperform the GR4J and LSTM models, and the PIML-H3 model achieves the best performance. While for high flows, the performance of all three PIML models is subpar. This indicates that PIML models tend to have an advantage in simulating low flows events. (3) The runoff simulations from the three PIML models exhibit significant seasonal variations during both the training and validation periods. The seasonal variations in the PIML-H2 and PIML-H3 models are more pronounced compared to the PIML-H1 model, indicating that the seasonal variations in simulated runoff results of the PIML model are influenced by intermediate variables. This study contributes to a better understanding of the performance differences among various PIML model schemes in runoff simulation, providing technical support for runoff simulation and forecasting in the study area. |
| Key words: physics-informed machine learning runoff simulation intermediate variable GR4J LSTM |
