水资源与水工程学报

文章摘要

张书齐，左其亭，臧超，张乐开，巴音吉.基于CNN-LSTM-Attention模型的沁河流域径流模拟及未来多情景预测Journal of Water Resources and Water Engineering[J].,2024,35(5):73-81

基于CNN-LSTM-Attention模型的沁河流域径流模拟及未来多情景预测

Runoff simulation and future multi-scenario prediction in the Qinhe River Basin based on the CNN-LSTM-Attention Model

DOI：10.11705/j.issn.1672-643X.2024.05.09

中文关键词: 径流模拟及预测深度学习模型 CNN-LSTM-Attention 气候变化沁河流域

英文关键词: runoff simulation and prediction deep learning model CNN-LSTM-Attention climate change the Qinhe River Basin

基金项目:国家重点研发计划项目（2021YFC3200201）；中国工程科技发展战略河南研究院战略咨询研究项目（2024HENYB01）；中国地质调查局项目（DD20220885）

Author Name	Affiliation
ZHANG Shuqi¹, ZUO Qiting^1,2, ZANG Chao^1,2, ZHANG Lekai¹, BA Yinji³	（1.郑州大学水利与交通学院，河南郑州 450001； 2.河南省水循环模拟与水环境保护国际联合实验室，河南郑州 450001； 3.中国地质调查局烟台海岸带地质调查中心, 山东烟台 264000）

Hits: 134

Download times: 48

中文摘要:

为提升深度学习模型对变化环境下流域的径流模拟精度，以沁河流域为例，构建了基于卷积神经网络（CNN）、长短期记忆网络（LSTM）和注意力机制（Attention）的CNN-LSTM-Attention耦合模型，加入多种优化算法，结合第六次国际耦合模式比较计划CMIP6中的BCC-CSM2-MR气候模式并考虑多种情景，应用于流域的径流模拟和预测，同时比较了多种深度学习模型的模拟精度。结果表明：CNN-LSTM-Attention模型在沁河流域表现出了较好的径流模拟效果，模拟精度均优于其他深度学习模型，纳什效率系数（NSE）为0.883，均方根误差（RMSE）为2.317，平均绝对误差（MAE）为1.098；不同气候变化情景下，沁河流域在2025—2050年的年径流量均呈现缓慢衰减趋势且波动程度较大，尤其在SSP1-2.6情景下，径流量衰减和波动程度突出。研究可为深度学习模型在人水关系智能化计算模拟领域的应用提供新思路，并为流域后续的水资源开发利用和管理提供科学参考价值。

英文摘要:

To enhance the accuracy of deep learning models in simulating watershed runoff under changing environmental conditions, a coupled model of convolutional neural network (CNN), long short-term memory (LSTM) and Attention mechanism was constructed for the study of the Qinhe River Basin. Integrated with multiple optimization algorithms and multiple scenarios in BCC-CSM2-MR climate model from the Coupled Model Intercomparison Project Phase 6 (CMIP6), this model was applied to watershed runoff simulation and prediction. Its simulation accuracy was then compared with that of various deep learning models. The results demonstrate that the CNN-LSTM-Attention model exhibits superior performance in simulating runoff in the Qinhe River Basin, with Nash-Sutcliffe efficiency coefficient (NSE) of 0.883, root mean square error (RMSE) of 2.317, and mean absolute error (MAE) of 1.098, outperforming other deep learning models. Notably, the annual runoff of the Qinhe River Basin from 2025 to 2050 shows a slow decreasing trend with significant fluctuations under different climate change scenarios, especially in the SSP1-2.6 scenario. This study provides new insights into the application of deep learning models in intelligent simulation of human-water relationships and offers a referential value for subsequent water resources development and management in the watershed.

View Full Text View/Add Comment Download reader