Application Research Of Machine Learning Technology In Rapid Forecast Of Urban Flood

With the worsening of global warming and the acceleration of urbanization,the condition of local climate and the effects of urban hydrological cycle have undergone significant changes,leading to increasing flood disasters and potential risk losses.In recent years,pluvial flooding have occurred frequently in China,which has a serious impact on the life and property safety of the people and the development of social economy.Urban storm water models and machine learning methods are used to predict pluvial flooding.The former is based on the hydrologic principle of physical mechanism,which can realize the elaborate simulation of pluvial flooding,but its timeliness is insufficient.The latter can realize the rapid prediction of pluvial flooding based on the data-driven method.At present,the most economical and effective method is to combine the advantages of the two models for rapid hydrological prediction.Based on previous research results,this paper systematically summarizes the model principle,parameter selection and modeling process based on IFMS Urban software,and describes the principles of machine learning methods,that is used for regression prediction,dimension reduction analysis and cluster analysis.In this paper,a combined model is developed for rapid prediction of peak flood flow and 2-D surface water,and its parameter selection and prediction accuracy are evaluated.The main contents of this paper are as follows:(1)The paper describes surface runoff yield,the overland flow and pipe flow,two-dimensional surface flood routing and 1D-2D coupled principle of the IFMS Urban model,and summarizes the model parameter selection and the process of modeling.Based on IFMS Urban software platform,a 1D-2D coupled model of Chengdu city is constructed.The model is calibrated and verified by simulation using data from two historical rainfall events.The results show that the coupled model has a good simulation accuracy.(2)A rainfall division method of long sequence rainfall data is proposed.Based on this method,73,40 and 75 rainfall with a duration of 3h,6h and 12h are divided from the rainfall data of Chengdu from 2016 to 2019.The 1D-2D coupled model of Chengdu city is used to simulate these rainfall events.From the simulation results,the total rainfall data of 45 rainfall stations,the peak flow of the given cross-section,the hourly inundated depth and the average inundated depth data of the grid are collected.These data are integrated into a data set for machine learning training.(3)The Ridge Regression,LASSO Regression,ARD Regression and Bayesian Ridge Regression methods are adopted to construct a machine learning model for predicting the peak flow of a specific cross-section.R-Square value is used to evaluate the accuracy of the model,showing that all the four regression methods have high prediction accuracy,among which LASSO regression works best.Comparing the prediction results of the four models on different cross-sections,it is shown that the simulation accuracy will decrease when the upstream of the section has lateral inflow.At the same time,a fully connected neural network(IFMS-FCNN)that can be used for multi-section prediction is constructed.The network structure and parameter selection are compared and analyzed,and the optimal network structure and parameters of the model are determined as follows:The network structure is 45-500-200-2.Dropout is 0.3.Learning rate is 0.005.Activation function is rectified linear unit(ReLU).Optimizer is Adam,and loss function is mean absolute error(MAE).By analyzing the simulation results of IFMS-Lasso model and IFMS-FCNN model for five rainfall events,it is shown that both models have good prediction accuracy.But the average error of cross-section 2 with lateral inflow upstream is higher than that of cross-section 1 without lateral inflow.(4)Principal Components Analysis(PCA),K-Means and LASSO Regression methods are selected to construct a combined models for predicting surface inundation depth and range.The interpretability of PC A method data under different principal component quantities is analyzed,and the results show that the explanatory variance no longer increased significantly after the principal component exceeded 40.Based on the results,the explanatory variance and principal component number of this paper are determined to be 95%and 38 respectively.The group sets of K-Means clustering are determined by reference error sum of squares(SSE)and silhouette coefficient,and the number of K-Means clustering is 4,8 and 13.The LASSO regression model between rainfall information and average regional inundation depth(ARID)is constructed,and the combined models is formed by ARID to realize the connection with the clustering model.Based on the combination model,the surface inundated depth prediction of a rainfall event is carried out.The results show that the model works best,when the number of group sets is 13 and the R-Square value of the model is 0.83.