基于深度学习和水动力模型的洪水演进快速模拟方法  

Rapid simulation of flood routing using deep learning and hydrodynamic model

作　　者：廖耀星 高玮志 张轩 赖成光[1,2] 王兆礼 LIAO Yaoxing;GAO Weizhi;ZHANG Xuan;LAI Chengguang;WANG Zhaoli(School of Civil Engineering&Transportation,South China University of Technology,Guangzhou 510641;Guangdong Artificial Intelligence and Digital Economy Laboratory(Guangzhou),Guangzhou 510330;The National Key Laboratory of Water Disaster Prevention,Nanjing Hydraulic Research Institute,Nanjing 210029)

机构地区：[1]华南理工大学土木与交通学院,广州510641 [2]人工智能与数字经济广东省实验室(广州),广州510330 [3]南京水利科学研究院水文水资源与水利工程科学国家重点实验室,南京210029

出　　处：《中国防汛抗旱》2024年第2期16-22,共7页China Flood & Drought Management

基　　金：国家自然科学基金项目(52379010);广东省基础与应用基础研究基金项目(2023B1515020087、2022A1515010019)。

摘　　要：洪涝灾害的快速模拟及预警预报是洪涝防灾减灾的重点,但目前基于物理机制的城市洪涝模型的模拟时效仍然过低。通过结合水动力模型模拟生成的洪水淹没数据和深度学习技术,构建基于卷积神经网络(CNN)的深度学习洪水模拟模型,对城市洪水淹没演进情况进行快速模拟。结果表明,所构建的CNN模型能较好地模拟洪水淹没的演进情况,预测峰值水深误差在8%以内,对淹没范围模拟效果良好。CNN模型具有极高的洪水淹没模拟效率,在保持和水动力模型近似精度的同时,计算效率可提升约400倍。研究结果可为流域与城市洪水淹没快速模拟、洪涝灾害预警预报和数字孪生流域建设等提供技术支撑。The rapid simulation and early warning of flood disasters are crucial for flood prevention and mitigation.However,the current simulation efficiency of urban flood models based on physical mechanisms remains low.In this study,a deep learning model based on convolutional neural network(CNN)is constructed by combining flood inundation data generated by hydrodynamic model and deep learning techniques to rapidly simulate urban flood routing.The results show that the developed CNN model can effectively simulate the flood inundation,with a peak water depth prediction error within 8%,and a good performance in simulating the inundation extent.The CNN model demonstrates a significantly higher efficiency in flood inundation simulation,achieving approximately 400 times faster computation while maintaining comparable accuracy to hydrodynamic models.This study can provide valuable insights for rapid simulation of urban flood inundation,early warning and forecasting of flood disasters,and the development of digital twin basins.

关 键 词：洪水淹没 水动力模型 深度学习 快速模拟 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TV122[自动化与计算机技术—控制科学与工程]