山洪是全球范围内最危险的自然灾害之一,具有突发性强、成灾快和破坏力大并且难以短时临近预测的特点.传统山洪预报预警方法主要依赖于基于物理机制的水文-水动力山洪过程模拟,然而这种方法计算复杂耗时较长,难以满足山洪的短时临近预测需求.以浙江临安仁里村为例,在水文-水动力物理模拟所产生的8 378条降雨时序和对应山洪淹没时空序列数据集的基础上,以基于卷积门控循环单元(convolutional gated recurrent unit convGRU)的深度神经网络作为核心,构建山洪时空序列预测代理模型.该模型通过输入过去24小时降雨观测时序和未来6小时的降雨预报时序,可实现未来6小时山洪淹没时空演变过程的快速预测.代理模型在测试集中能可靠地预测未来逐小时的山洪淹没范围、最大淹没深度以及淹没位置,未来6小时预测的可决系数均值为0.96,且预测速度平均比物理模拟快15625倍.这表明该代理模型能够捕捉物理模拟中降雨到山洪的复杂映射关系,实现目标区域山洪的快速预测,为山洪预警及应急响应决策制定提供有力的模型基础.
Surrogate model for spatio-temporal prediction of flash flooding based on deep learning
Mountain flash floods are one of the most dangerous natural disasters globally,characterized by strong suddenness,rapid development,great destructive power,and difficult short-term prediction.Traditional mountain flood prediction and early warning methods mainly rely on hydro-hydraulic mountain flood process simulation based on physical mechanisms.However,this method is computationally complex,time-consuming,and difficult to meet the short-term prediction needs of mountain floods.Taking Renli Village,Lin'an,Zhejiang as an example,based on 8 378 rainfall time series generated from hydro-hydraulic physical simulation and the corresponding mountain flood inundation spatiotemporal sequence dataset,a deep neural network model was constructed using convolutional gated recurrent units(convGRU)as the core to build a spatiotemporal sequence prediction agent model for mountain flash floods.This model can rapidly predict the spatiotemporal evolution process of mountain flood inundation in the next 6 hours by inputting the past 24 hours of rainfall observation time series and the next 6 hours of rainfall forecast time series.The agent model can reliably predict the hourly mountain flood inundation range,the maximum inundation depth,and inundation position in the next 6 hours in the test set,with a mean coefficient of determination of 0.96,and the prediction speed is on average 15 625 times faster than physical simulation.This indicates that the agent model can capture the complex mapping relationship from rainfall to mountain flash floods in physical simulation,achieve rapid prediction of mountain flash floods in the target area,and provide a strong model foundation for mountain flash flood early warning and emergency response decision-making.
deep learningflash floods modelspatio-temporal sequence predictionConvGRUsurrogate model
NETL
NSTL
万方数据
