基于长短时记忆网络的山区中小流域降雨径流模拟  

Rainfall-runoff Simulation in the Small and Medium-sized Catchments Using the Long Short-Term Memory Network

作　　者：张锦堂[1] 任明磊[2,3] 李京兵 唐榕[2,3] 钟小燕 王刚 王玉丽[1] ZHANG Jin-tang;REN Ming-lei;LI Jing-bing;TANG Rong;ZHONG Xiao-yan;WANG Gang;WANG Yu-li(Anhui Provincial Bureau of Hydrology,Hefei 230022,China;China Institute of Water Resources and Hydropower Research,Beijing 100038,China;Engineering Technology Research Center for Flood Control,Drought Relief and Disaster Reduction,Ministry of Water Resources(Flood and Drought Disaster Prevention Center),Beijing 100038,China)

机构地区：[1]安徽省水文局,安徽合肥230022 [2]中国水利水电科学研究院,北京100038 [3]水利部防洪抗旱减灾工程技术研究中心(水旱灾害防御中心),北京100038

出　　处：《水电能源科学》2024年第8期33-37,共5页Water Resources and Power

基　　金：国家重点研发计划(2022YFC3006400,2022YFC3006403)。

摘　　要：洪水预报是流域防洪减灾的重要非工程措施之一。目前我国中小河流暴雨洪水灾害频发,但应对短历时强降雨的洪水预报能力仍不强。以安徽省东部山区中小流域为研究对象,引入长短时记忆网络建立流域降雨径流模型,探讨其在山区中小流域的洪水模拟效果。结果表明,考虑降雨输入的空间差异可提升深度学习模型降雨径流模拟预测性能,且长短时记忆网络能够取得优于传统人工神经网络的精度;长短时记忆网络模型有效建立了流域降雨与径流间的复杂非线性关系,模型在所选流域内场次洪水的峰值模拟效果较好,训练、测试集场次洪水峰值合格率均在90%以上;长短时记忆网络内部结构特征与流域水文过程具有较好的相似性,对山区中小流域暴雨洪水非线性关系拟合效果突出。Flood forecasting is one of the important non-engineering measures for flood prevention and mitigation in river basins.Presently,heavy rainfall and flood disasters occur frequently in the small and medium-sized catchments,and the capability of flood forecasting remains limited when facing short-term rainstorm.Taking a small and medium-sized catchment in the eastern mountainous area of Anhui Province as the study site,this paper built a rainfall-runoff model using the Long Short-Term Memory(LSTM)network,and the modelling performance of the LSTM model in the selected study site was explored.The results show that the selecting the stational measured rainfall as model inputs can improve its performance on rainfall-runoff modelling and prediction,and the LSTM model achieves higher precision than traditional artificial neural networks.Besides,the LSTM model effectively captures the complicated rainfall-runoff relationship in the selected river basin,and can achieve satisfactory performance on the flood peak simulation.For the training and testing flood events,the calculated qualified rates of flood peak simulation are both over 90%.Meanwhile,the internal structural characteristics of the LSTM model have strong similarity with the hydrological process over the river basin,which contributes to a good fitting performance of the nonlinear rainfall-runoff relationship in the selected catchment.

关 键 词：山丘区 长短时记忆网络 中小河流 降雨径流模拟 

分 类 号：TV121.1[水利工程—水文学及水资源]