摘要
传统内涝数值模拟方法主要基于流体力学和水文学原理实现内涝模拟。这种方法精度高但计算耗时,难以满足城市内涝预报的时效性要求。以苍南县灵溪镇为例,该文在仿真模型产生8000条降雨时序对应的城市内涝淹没数据集的基础上,通过耦合长短时记忆网络(long short-term memory,LSTM)与卷积神经网络(convolutional neural network,CNN),构建基于数据驱动的城市暴雨内涝多步提前预测代理模型。代理模型通过输入过去6 h实测降雨和未来6 h预报降雨时序,实现对未来6 h的城市内涝淹没时空预测。结果表明:代理模型在测试集中预测值与标签值的回归线可决系数(R2)达到0.9574;在台风“菲特”(2013年第23号强台风)案例中,代理模型仅耗时10 s完成24 h的内涝精准预测。该模型实现了对暴雨-内涝灾害链的精准高效预测,为防范和减轻内涝灾害的应急决策制定提供科学支持。
Traditional numerical simulation methods for urban flooding are mainly based on the principles of fluid mechanics and hydrology for inundation modeling.While this method offers high accuracy,it is computationally time-consuming and struggles to meet the timeliness requirements of urban flooding forecasts and early warning.Taking Lingxi Town in Cangnan County as an example,this paper,based on a dataset of 8000 rainfall time series corresponding to urban flooding generated by a simulation model,constructs a data-driven multi-step advance prediction surrogate model for urban rainstorm flooding by coupling a long short-term memory(LSTM)network and a convolutional neural network(CNN).The surrogate model,by inputting the measured rainfall of the past 6 h and the forecast rainfall time series for the next 6 h,achieves spatiotemporal prediction of urban flooding in the next 6 h.Results show that the surrogate model achieves a coefficient of determination(R 2)of 0.9574 on the regression line between predicted and labeled values in the test set.In the case of Typhoon Fitow(the 23rd typhoon of 2013),the surrogate model completed a precise 24 h flooding prediction in just 10 s.This model enables accurate and efficient prediction of the rainstorm-waterlogging disaster chain,providing scientific support for emergency decision-making in preventing and mitigating waterlogging disasters.
作者
岳子怡
王俊彦
王乃玉
许月萍
YUE Ziyi;WANG Junyan;WANG Naiyu;XU Yueping(College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310058,China)
出处
《自然灾害学报》
北大核心
2026年第1期33-45,共13页
Journal of Natural Disasters
基金
浙江省科学技术厅尖兵领雁研发攻关计划项目(2024C03255)。
关键词
城市内涝
深度学习
多步提前预测
时空预测
卷积神经网络
长短期记忆神经网络
urban waterlogging
deep learning
multi-step ahead prediction
spatiotemporal forecasting
convolutional neural networks
long short-term memory neural networks
