考虑时空特征的城市内涝智能预报模型研究

Study on urban waterlogging intelligent forecast model considering temporal and spatial characteristics

针对传统城市内涝预报模型计算耗时长、实测内涝样本少、内涝特征因子欠考虑等问题,通过耦合SWMM模型和LISFLOOD-FP模型搭建了城市内涝机理模型,利用不同重现期下的设计暴雨进行数值模拟并生成内涝样本;基于样本和内涝特征因子构建了三维时空矩阵,实现对内涝特征因子数据的有序组织;在此基础上,将卷积神经网络(CNN)与长短时记忆网络(LSTM)进行耦合,构建了一种考虑多时空特征的城市内涝智能预报模型(CNN-LSTM);最后以三维时空矩阵为驱动,对该智能模型进行训练,选取广州市天河区的实测样本对其性能进行评估。结果表明:CNN-LSTM模型可以快速预报淹没水位和淹没范围,易涝控制点水位过程模拟的纳什效率系数在0.9以上,各个时刻淹没面积的平均匹配率达到92.2%,相对于机理模型的模拟效率提高了近70倍。该智能模型具有良好的预报精度和效率,可有效支撑城市防灾减灾工作。

Given the problems of the traditional urban waterlogging forecast model,such as being time-consuming,few measured waterlogging samples,and insufficient consideration of waterlogging characteristic factors,an urban waterlogging mechanism model was first built by coupling the SWMM model and the LISFLOOD-FP model.The mechanism model was used to numerically simulate the designed rainstorm in different recurrence periods to generate waterlogging samples.Based on samples and waterlogging characteristic factors,a three-dimensional spatio-temporal matrix was constructed to realize the orderly organization of waterlogging characteristic factor data.Based on the above,a convolutional neural network(CNN)was coupled with long short term memory network(LSTM),and an urban waterlogging intelligent forecast model considering multi-temporal characteristics(CNN-LSTM)was constructed.The intelligent model was trained by a three-dimensional space-time matrix using measured samples from Tianhe District,Guangzhou City.The results show that the CNN-LSTM model can quickly predict the inundation depth and inundation range.The Nash coefficient of waterlogging control point water level simulation was above 0.9,and the average matching rate of the inundated area at every moment reached 92.2%.Compared with the mechanism model,the simulation efficiency was improved by nearly 70 times.The intelligent model had good forecasting accuracy and efficiency,and could effectively support the work of urban waterlogging prevention and disaster reduction.