[1]司徒祖祥,冯婉恩,钟琪升,等.基于时空特征融合的城市洪涝混合深度学习预测[J].中国给水排水,2024,40(17):131-136.
　SITUZu-xiang,FENGWan-en,ZHONGQi-sheng,et al.Urban Flood Prediction Using Hybrid Deep Learning Model Based on Spatio-temporal Feature Fusion[J].China Water & Wastewater,2024,40(17):131-136.
点击复制

基于时空特征融合的城市洪涝混合深度学习预测

中国给水排水[ISSN:1000-4062/CN:12-1073/TU] 卷: 第40卷 期数: 2024年第17期 页码: 131-136 栏目: 出版日期: 2024-09-01
Title:
Urban Flood Prediction Using Hybrid Deep Learning Model Based on Spatio-temporal Feature Fusion
作者:
司徒祖祥， 冯婉恩， 钟琪升， 廖晓婷， 饶开昕， 张佳博， 周倩倩
（广东工业大学 土木与交通工程学院，广东 广州 510006）
Author(s):
SITU　Zu-xiang， FENG　Wan-en， ZHONG　Qi-sheng， LIAO　Xiao-ting， RAO　Kai-xin， ZHANG　Jia-bo， ZHOU　Qian-qian
（School of Civil and Transportation Engineering， Guangdong University of Technology， Guangzhou 510006， China）
关键词:
城市洪涝预测;  时空特征融合;  深度学习;  混合模型
Keywords:
urban flood prediction;  spatio-temporal feature fusion;  deep learning;  hybrid model
摘要:
传统物理模型在二维地表淹没计算中虽然实现了高精度的求解，但其构建过程复杂、校准难度高，且计算效率低，在实时洪涝预报方面存在局限性。深度学习作为人工智能领域的重要分支，具有强大的数据处理和学习能力，可为洪涝预测提供变革性、创新性的技术手段。为此，提出了一种基于时空特征融合技术的城市洪涝混合深度学习预测模型，充分结合卷积神经网络和循环神经网络对空间和时间数据的学习优势，建立了SegNet-GRU混合模型。该模型能够准确预测研究区域在不同降雨情景下的极值水深和地表积水演变过程，实现了良好的预测精度（平均绝对误差、均方根误差、纳什效率系数和克林-古普塔效率系数分别为0.008 5、0.030 6、0.962 7、0.694 9）和处理速度（较一、二维模型预测速率提升近160倍）。
Abstract:
The traditional physical models, while capable of accurately solving two-dimensional surface inundation calculations, are hindered by a complex construction process, high calibration difficulty, and low calculation efficiency. These limitations impede real-time flood prediction. As a crucial subset of artificial intelligence, deep learning has robust data processing and learning capabilities, offering a transformative and innovative technical approach for flood prediction. Consequently, a hybrid deep learning model for urban flood prediction was proposed based on spatio-temporal feature fusion technology. The SegNet-GRU hybrid model was developed by fully integrating the advantages of convolutional neural network and recurrent neural network in learning spatial and temporal data. The model demonstrated high accuracy in predicting the evolution process of the maximum water depth and surface water ponding under varying rainfall scenarios within the study area, with good prediction accuracy (mean absolute error, root-mean-square error, Nash-Sutcliffe efficiency coefficient, and Klin-Gupta efficiency coefficient at 0.008 5, 0.030 6, 0.962 7 and 0.694 9 respectively) as well as significantly improved processing speed (nearly 160 times faster than one- and two-dimensional models).
更新日期/Last Update: 2024-09-01