Rapid urban flood simulation and prediction method based on grid point rainfall data

Abstract

The invention discloses an urban flood rapid simulation and prediction method based on grid point rainfall data, and the method comprises the steps: building a hydrological hydrodynamic model through urban basic data, simulating and capturing flood response characteristics under different terrains and land utilization types, representing the spatial distribution characteristics of urban rainfall based on refined grid point rainfall data, and carrying out the rapid simulation and prediction of urban flood. The method comprises the following steps: constructing grid point data by combining monitoring station data and design rainfall data and adopting a Gaussian attenuation model and an interpolation calculation method, extracting spatial rainfall distribution characteristics in the grid rainfall data by utilizing a convolutional neural network (CNN) model, and automatically identifying rainfall intensity, rainfall range and spatial association relationship characteristics through convolution and pooling operation. And finally, a flood rapid simulation prediction model is obtained. The problems that in the prior art, the flood simulation precision is insufficient due to uneven rainfall space-time distribution, and the simulation efficiency of a traditional physical numerical model is low are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of numerical simulation, and in particular relates to a method for rapid simulation and prediction of urban floods based on grid point rainfall data. Background Art

[0002] Rainfall, a major contributing factor to urban flooding, is influenced by the highly irregular nature of local microclimates and storm clouds within urban areas. This leads to uneven spatial distribution, variable rainfall frequency, and difficulty capturing the center of heavy rain, making it difficult to accurately simulate flooding. In recent years, with the iterative updating of meteorological data, combined with live observations from ground, radar, and satellite observations, and the assimilation of GRAPES-MESO numerical model products, gridded, high-resolution rainfall data has been increasingly adopted across various fields, providing more accurate data support for urban flood simulations.

[0003] Furthermore, urban flood numerical models require repeated iterations to solve complex physical equations during flood simulation, making them incapable of meeting the timeliness requirements of urban flood emergency forecasting. In recent years, AI technology has made significant progress in fields such as natural language processing, machine translation, and big data mining due to its universal applicability and efficiency. Early AI applications in flood forecasting relied primarily on simple machine learning algorithms and statistical models to learn and predict historical flood data. However, with the rapid development of deep learning, this technology has achieved significant breakthroughs in handling complex nonlinear relationships, capturing characteristic factors of time series, and automatically extracting characteristic parameters, providing new approaches for rapid simulation and forecasting of urban flood disasters. Furthermore, due to the difficulty of collecting multiple high-resolution gridded rainfall data, the primary approach to addressing the dataset shortage for rapid flood simulation and forecasting models is to interpolate designed rainfall data and existing meteorological station monitoring data into gridded rainfall data, thereby generating flood data. Rapidly simulating and forecasting flood disaster processes based solely on gridded data to support urban flood warning and forecasting remains an urgent challenge. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for rapid simulation and prediction of urban floods based on grid point rainfall data, which solves the problems in the prior art of insufficient flood simulation accuracy due to uneven temporal and spatial distribution of rainfall, and low simulation efficiency of traditional physical numerical models.

[0005] The technical solution adopted in the present invention is a rapid simulation and prediction method for urban floods based on grid rainfall data. A hydrological and hydrodynamic model is constructed using urban basic data to simulate and capture flood response characteristics under different terrains and land use types. The spatial distribution characteristics of urban rainfall are characterized based on refined grid rainfall data. In combination with monitoring station data and designed rainfall data, a Gaussian attenuation model and an interpolation calculation method are used to construct grid data. A convolutional neural network (CNN) model is used to extract spatial rainfall distribution characteristics from grid rainfall data. Rainfall intensity, rainfall range and spatial correlation relationship characteristics are automatically identified through convolution and pooling operations. The flood process corresponding to the grid rainfall data is then trained and learned through a Transformer model. The complex spatiotemporal relationship between rainfall and flood process is captured through a self-attention mechanism, and finally a rapid simulation and prediction model for floods is obtained.

[0006] The present invention is also characterized in that: Please follow the steps below to implement it: Step 1: Collect basic urban data, build a high-precision hydrological and hydrodynamic model, and complete parameter setting for the hydrological and hydrodynamic model.

[0007] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; Step 6: The spatial features of the grid rainfall are extracted from the designed grid rainfall data constructed in step 4 and the multi-source fusion actual grid rainfall through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through the filter, while the pooling layer downsamples the local area. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model. Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0008] Step 1 is implemented as follows: First, the terrain elevation data and land use type data in the study area were processed into .asc raster files in GIS. Different infiltration parameters and Manning parameters were assigned to different land use type data. A one-dimensional pipe network .inp file was constructed using the pipe network data of the study area. The prepared input file was imported into the hydrological and hydrodynamic model. The parameters and reliability of the hydrological and hydrodynamic model were calibrated and verified based on historical measured rainfall and flood data.

[0009] Secondly, for spatially distributed grid rainfall, the grids in the spatial grid rainfall are numbered as follows: R 1. R 2. R 3.…… R n In order to make the spatial distribution and grid size of the surface and rainfall consistent during the hydrodynamic model simulation, the numbered rainfall grid is resampled to be consistent with the surface terrain grid using the nearest neighbor interpolation method, and the rainfall process is assigned to each rainfall number using a .csv file: (1) Where, is the location of the target grid cell, The target grid The value of ( x , y ) is the pixel location in the original grid that is closest to the center of the target pixel. and for( x , y ) rounded to the nearest integer grid point.

[0010] Step 2 is implemented as follows: Use Euclidean distance to calculate the distance between different weather stations, taking a certain weather station as the midpoint M i, calculate the slope and perpendicular bisector between each station, and divide the control area of ​​each meteorological station into S 1. S 2. S 3.…… S n .

[0011] Step 3 is implemented as follows: The control area of ​​the weather station site is divided using Thiessen polygons, and each control area is taken as the center of the rainstorm. The Gaussian attenuation model is used to calculate the rainfall in the adjacent control area. The specific division method is as follows: (2) Where, represents the rainfall intensity at a distance d from the center of the rainstorm; d is the Euclidean distance between two sites, P i is the rainfall intensity at the center of the rainstorm, and σ is the standard deviation parameter that controls the attenuation.

[0012] Step 4 is implemented as follows: The designed rainfall data of the controlled area of ​​the meteorological station is used to interpolate the grid rainfall process, and the Kriging interpolation method is used as follows: (3) (4) (5) Where, Represents the unknown point S An estimated value of 0; is the weight coefficient; For location Known samples at n is the number of known sample points, C is the covariance between two known samples; d is the covariance between the known sample and the predicted sample; λ is the weight coefficient vector, μ is the Lagrange multiplier, 1 is the vector of all ones; γ (h) is the semivariance of the difference of the random variables between two locations with a distance h between them, Var is the variogram, which is used to describe spatial variability.

[0013] The flood hydraulic elements in step 5 include water accumulation depth, water accumulation range, water accumulation volume, risk level and pipe network load.

[0014] Step 6 is implemented as follows: A convolutional neural network (CNN) algorithm is used to extract the characteristic parameters of grid rainfall. The network structure consists of an input layer, a convolution layer, an activation layer, and a pooling layer, which are connected in sequence. The input layer receives rainfall grid data at different times. The convolution layer performs a sliding operation on the input data using multiple convolution kernels, automatically extracting spatial feature information such as rainfall intensity distribution and concentrated areas within the local area. The pooling layer further processes the rainfall feature matrix output by the convolution layer and downsamples by calculating the maximum or average value of the local area, thereby effectively compressing the feature dimension and reducing the amount of calculation. The specific calculation method is as follows: (6) (7) (8)

[0015] Save the grid rainfall feature parameter information automatically extracted by the CNN model into an .h5 file, and use this feature information as model input in step 7.

[0016] Step 7 is implemented as follows: The Transformer model is used to construct a rapid simulation and forecast model for urban flooding. The spatial characteristic parameter information of the grid rainfall obtained in step 6 and the flood data during the rainfall process of different grids obtained in step 5 are used as the input of the Transformer model. The model can include an encoder and a decoder, mainly composed of modules such as multi-head attention layer, feedforward neural network, residual connection and layer normalization. It has the advantages of high efficiency, flexibility and easy expansion. The calculation method is: (9) (10) Where: For multi-head output; Concat is the connection function; Head t For the t Output of the head; head i For the i Output of the head; W O is the concatenation matrix, which is used to concatenate the output sub-vectors of multiple heads to obtain a complete output vector; W Q , W K , W K is a learnable parameter matrix that maps the vector at each position in the input sequence to Q 、 K 、V ; (11) (12) Where: The output of the current fully connected layer, H Represents the output value of the previous layer; W 1 , W 2 is the learnable parameter matrix; b 1, b 2 is the learnable bias vector, , , , ; is the activation function.

[0017] The beneficial effect of this invention is that the rapid simulation and prediction method for urban flooding based on grid-based rainfall data can accurately depict the spatially non-uniform distribution characteristics of rainfall within urban areas, making it particularly suitable for responding to waterlogging disasters caused by localized sudden heavy rainfall. By fully combining refined grid rainfall data with the ability of deep learning models to automatically extract complex spatial rainfall patterns, flood forecast results are more accurate in key indicators such as water depth and waterlogging range. At the same time, the prediction calculation time is significantly shortened, providing valuable lead time for urban flood control and emergency response, effectively protecting people's lives and property, and possessing excellent engineering practicality and promotional value. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 1 is a flow chart of a method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to the present invention; Figure 2 This is a schematic diagram of spatial rainfall grid division of the present invention; Figure 3 This is a schematic diagram of rainfall intensity distribution based on meteorological stations in the present invention; Figure 4 This is a schematic diagram of the present invention interpolating rainfall data from a weather station into grid point data; Figure 5 This is a comparison chart of the accuracy between the predicted values ​​of the urban flood rapid simulation prediction model based on grid point rainfall data and the simulated values ​​of the hydrological and hydrodynamic model; Figure 6 This is a comparison chart of the computational efficiency between the predicted values ​​of the urban flood rapid simulation prediction model based on grid point rainfall data and the simulated values ​​of the hydrological and hydrodynamic model. DETAILED DESCRIPTION

[0019] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The present invention proposes a rapid simulation and prediction method for urban floods based on grid rainfall data. The method constructs a hydrological and hydrodynamic model with urban basic data to simulate and capture flood response characteristics under different terrains and land use types. The method characterizes the spatial distribution characteristics of urban rainfall based on refined grid rainfall data. In combination with monitoring station data and designed rainfall data, the method adopts a Gaussian attenuation model and an interpolation calculation method to construct grid data. The method uses a convolutional neural network (CNN) model to extract spatial rainfall distribution characteristics from grid rainfall data. The method automatically identifies rainfall intensity, rainfall range, and spatial correlation relationship characteristics through convolution and pooling operations. The method then trains and learns the flood process corresponding to the grid rainfall data through a Transformer model. The method captures the complex spatiotemporal relationship between rainfall and flood process through a self-attention mechanism, ultimately obtaining a rapid simulation and prediction model for floods.

[0021] The present invention is based on the rapid simulation and prediction method of urban floods based on grid point rainfall data, combined with Figure 1 , specifically follow the steps below: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0022] Step 1 is implemented as follows: First, the terrain elevation data and land use type data in the study area were processed into .asc raster files in GIS. Different infiltration parameters and Manning parameters were assigned to different land use type data. A one-dimensional pipe network .inp file was constructed using the pipe network data of the study area. The prepared input file was imported into the hydrological and hydrodynamic model. The parameters and reliability of the hydrological and hydrodynamic model were calibrated and verified based on historical measured rainfall and flood data.

[0023] Secondly, for spatially distributed grid rainfall, the grids in the spatial grid rainfall are numbered as follows: R 1. R 2. R 3.…… R n In order to make the spatial distribution and grid size of the surface and rainfall consistent during the hydrodynamic model simulation, the numbered rainfall grid is resampled to be consistent with the surface terrain grid using the nearest neighbor interpolation method, and the rainfall process is assigned to each rainfall number using a .csv file: (1) Where, is the location of the target grid cell, The target grid The value of ( x , y ) is the pixel location in the original grid that is closest to the center of the target pixel. and for( x , y ) rounded to the nearest integer grid point.

[0024] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 2 is implemented as follows: Use Euclidean distance to calculate the distance between different weather stations, taking a certain weather station as the midpoint M i , calculate the slope and perpendicular bisector between each station, and divide the control area of ​​each meteorological station into S 1. S 2. S 3.…… S n .

[0025] Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 3 is implemented as follows: The control area of ​​the weather station site is divided using Thiessen polygons, and each control area is taken as the center of the rainstorm. The Gaussian attenuation model is used to calculate the rainfall in the adjacent control area. The specific division method is as follows: (2) Where, represents the rainfall intensity at a distance d from the center of the rainstorm; d is the Euclidean distance between two sites, P i is the rainfall intensity at the center of the rainstorm, and σ is the standard deviation parameter that controls the attenuation.

[0026] Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 4 is implemented as follows: The designed rainfall data of the controlled area of ​​the meteorological station is used to interpolate the grid rainfall process, and the Kriging interpolation method is used as follows: (3) (4) (5) Where, Represents the unknown point S An estimated value of 0; is the weight coefficient; For location Known samples at n is the number of known sample points, C is the covariance between two known samples; d is the covariance between the known sample and the predicted sample; λ is the weight coefficient vector, μ is the Lagrange multiplier, 1 is the vector of all ones; γ (h) is the semivariance of the difference of the random variables between two locations with a distance h between them, Var Variogram is a function used to describe spatial variability.

[0027] Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; The flood hydraulic elements in step 5 include water accumulation depth, water accumulation range, water accumulation volume, risk level and pipe network load.

[0028] Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 6 is implemented as follows: A convolutional neural network (CNN) algorithm is used to extract the characteristic parameters of grid rainfall. The network structure consists of an input layer, a convolution layer, an activation layer, and a pooling layer, which are connected in sequence. The input layer receives rainfall grid data at different times. The convolution layer performs a sliding operation on the input data using multiple convolution kernels, automatically extracting spatial feature information such as rainfall intensity distribution and concentrated areas within the local area. The pooling layer further processes the rainfall feature matrix output by the convolution layer and downsamples by calculating the maximum or average value of the local area, thereby effectively compressing the feature dimension and reducing the amount of calculation. The specific calculation method is as follows: (6) (7) (8)

[0029] Save the grid rainfall feature parameter information automatically extracted by the CNN model into an .h5 file, and use this feature information as model input in step 7.

[0030] Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 7 is implemented as follows: The Transformer model is used to construct a rapid simulation and forecast model for urban flooding. The spatial characteristic parameter information of the grid rainfall obtained in step 6 and the flood data during the rainfall process of different grids obtained in step 5 are used as the input of the Transformer model. The model can include an encoder and a decoder, mainly composed of modules such as multi-head attention layer, feedforward neural network, residual connection and layer normalization. It has the advantages of high efficiency, flexibility and easy expansion. The calculation method is: (9) (10) Where: For multi-head output; Concat is the connection function; Head t For the t Output of the head; head i For the i Output of the head; W O is the concatenation matrix, which is used to concatenate the output sub-vectors of multiple heads to obtain a complete output vector; W Q , W K ,W K is a learnable parameter matrix that maps the vector at each position in the input sequence to Q 、 K 、 V ; (11) (12) Where: The output of the current fully connected layer, H Represents the output value of the previous layer; W 1 , W 2 is the learnable parameter matrix; b 1, b 2 is the learnable bias vector, , , , ; is the activation function.

[0031] The above method ultimately generates a rapid urban flood simulation and prediction model based on grid-based rainfall data. By inputting grid-based rainfall data from meteorological forecasts for different forecast periods, the model can quickly predict flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, along with water conservancy elements such as risk level and pipe network load. This output provides data support for subsequent risk assessments, visualization, and urban flood control scheduling and decision-making.

[0032] Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0033] Example 1 The rapid simulation and prediction method for urban flooding based on grid point rainfall data of the present invention is specifically implemented according to the following steps: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0034] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0035] Example 2 The rapid simulation and prediction method for urban flooding based on grid point rainfall data of the present invention is specifically implemented according to the following steps: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0036] Step 1 is implemented as follows: First, the terrain elevation data and land use type data in the study area were processed into .asc raster files in GIS. Different infiltration parameters and Manning parameters were assigned to different land use type data. A one-dimensional pipe network .inp file was constructed using the pipe network data of the study area. The prepared input file was imported into the hydrological and hydrodynamic model. The parameters and reliability of the hydrological and hydrodynamic model were calibrated and verified based on historical measured rainfall and flood data.

[0037] Secondly, for spatially distributed grid rainfall, the grids in the spatial grid rainfall are numbered as follows: R 1. R 2. R 3.…… R n In order to make the spatial distribution and grid size of the surface and rainfall consistent during the hydrodynamic model simulation, the numbered rainfall grid is resampled to be consistent with the surface terrain grid using the nearest neighbor interpolation method, and the rainfall process is assigned to each rainfall number using a .csv file: (1) Where, is the location of the target grid cell, The target grid The value of ( x , y ) is the pixel location in the original grid that is closest to the center of the target pixel. and for( x , y ) rounded to the nearest integer grid point.

[0038] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0039] Example 3 The rapid simulation and prediction method for urban flooding based on grid point rainfall data of the present invention is specifically implemented according to the following steps: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0040] Step 1 is implemented as follows: First, the terrain elevation data and land use type data in the study area were processed into .asc raster files in GIS. Different infiltration parameters and Manning parameters were assigned to different land use type data. A one-dimensional pipe network .inp file was constructed using the pipe network data of the study area. The prepared input file was imported into the hydrological and hydrodynamic model. The parameters and reliability of the hydrological and hydrodynamic model were calibrated and verified based on historical measured rainfall and flood data.

[0041] Secondly, for spatially distributed grid rainfall, the grids in the spatial grid rainfall are numbered as follows: R 1. R 2. R 3.…… R nIn order to make the spatial distribution and grid size of the surface and rainfall consistent during the hydrodynamic model simulation, the numbered rainfall grid is resampled to be consistent with the surface terrain grid using the nearest neighbor interpolation method, and the rainfall process is assigned to each rainfall number using a .csv file: (1) Where, is the location of the target grid cell, The target grid The value of ( x , y ) is the pixel location in the original grid that is closest to the center of the target pixel. and for( x , y ) rounded to the nearest integer grid point.

[0042] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 2 is implemented as follows: Use Euclidean distance to calculate the distance between different weather stations, taking a certain weather station as the midpoint M i , calculate the slope and perpendicular bisector between each station, and divide the control area of ​​each meteorological station into S 1. S 2. S 3.…… S n .

[0043] Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; The flood hydraulic elements in step 5 include water accumulation depth, water accumulation range, water accumulation volume, risk level and pipe network load.

[0044] Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0045] Example 4 The rapid simulation and prediction method for urban flooding based on grid point rainfall data of the present invention is specifically implemented according to the following steps: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0046] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 3 is implemented as follows: The control area of ​​the weather station site is divided using Thiessen polygons, and each control area is taken as the center of the rainstorm. The Gaussian attenuation model is used to calculate the rainfall in the adjacent control area. The specific division method is as follows: (2) Where, represents the rainfall intensity at a distance d from the center of the rainstorm; d is the Euclidean distance between two sites, P i is the rainfall intensity at the center of the rainstorm, and σ is the standard deviation parameter that controls the attenuation.

[0047] Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 4 is implemented as follows: The designed rainfall data of the controlled area of ​​the meteorological station is used to interpolate the grid rainfall process, and the Kriging interpolation method is used as follows: (3) (4) (5) Where, Represents the unknown point S An estimated value of 0; is the weight coefficient; For location Known samples at n is the number of known sample points, C is the covariance between two known samples; d is the covariance between the known sample and the predicted sample; λ is the weight coefficient vector, μ is the Lagrange multiplier, 1 is the vector of all ones; γ (h) is the semivariance of the difference of the random variables between two locations with a distance h between them, Var Variogram is a function used to describe spatial variability.

[0048] Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0049] Example 5 The rapid simulation and prediction method for urban flooding based on grid point rainfall data of the present invention is specifically implemented according to the following steps: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0050] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 4 is implemented as follows: The designed rainfall data of the controlled area of ​​the meteorological station is used to interpolate the grid rainfall process, and the Kriging interpolation method is used as follows: (3) (4) (5) Where, Represents the unknown point S An estimated value of 0; is the weight coefficient; For location Known samples at n is the number of known sample points, Cis the covariance between two known samples; d is the covariance between the known sample and the predicted sample; λ is the weight coefficient vector, μ is the Lagrange multiplier, 1 is the vector of all ones; γ (h) is the semivariance of the difference of the random variables between two locations with a distance h between them, Var Variogram is a function used to describe spatial variability.

[0051] Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; The flood hydraulic elements in step 5 include water accumulation depth, water accumulation range, water accumulation volume, risk level and pipe network load.

[0052] Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 6 is implemented as follows: A convolutional neural network (CNN) algorithm is used to extract the characteristic parameters of grid rainfall. The network structure consists of an input layer, a convolution layer, an activation layer, and a pooling layer, which are connected in sequence. The input layer receives rainfall grid data at different times. The convolution layer performs a sliding operation on the input data using multiple convolution kernels, automatically extracting spatial feature information such as rainfall intensity distribution and concentrated areas within the local area. The pooling layer further processes the rainfall feature matrix output by the convolution layer and downsamples by calculating the maximum or average value of the local area, thereby effectively compressing the feature dimension and reducing the amount of calculation. The specific calculation method is as follows: (6) (7) (8)

[0053] Save the grid rainfall feature parameter information automatically extracted by the CNN model into an .h5 file, and use this feature information as model input in step 7.

[0054] Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0055] Example 6 The rapid simulation and prediction method for urban flooding based on grid point rainfall data of the present invention is specifically implemented according to the following steps: Step 1: Collect high-precision refined grid rainfall data, digital elevation data, land use types, multi-source fusion real-time grid rainfall, pipe network and other urban basic data, build a high-precision hydrological and hydrodynamic model, and complete the parameter setting of the hydrological and hydrodynamic model to provide a reliable model foundation for subsequent simulations.

[0056] Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Combine Figure 2 、 Figure 3 ,Step 3, using the meteorological station site control area divided in step 1, assuming that each control area is the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, according to the principle that rainfall intensity gradually decays from the center position to the adjacent control area with distance, calculate the rainfall intensity in different areas; Step 3 is implemented as follows: The control area of ​​the weather station site is divided using Thiessen polygons, and each control area is taken as the center of the rainstorm. The Gaussian attenuation model is used to calculate the rainfall in the adjacent control area. The specific division method is as follows: (2) Where, represents the rainfall intensity at a distance d from the center of the rainstorm; d is the Euclidean distance between two sites, P i is the rainfall intensity at the center of the rainstorm, and σ is the standard deviation parameter that controls the attenuation.

[0057] Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods (such as Kriging interpolation and inverse distance weighting) are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 4 is implemented as follows: The designed rainfall data of the controlled area of ​​the meteorological station is used to interpolate the grid rainfall process, and the Kriging interpolation method is used as follows: (3) (4) (5) Where, Represents the unknown point S An estimated value of 0; is the weight coefficient; For location Known samples at n is the number of known sample points, C is the covariance between two known samples; d is the covariance between the known sample and the predicted sample; λ is the weight coefficient vector, μ is the Lagrange multiplier, 1 is the vector of all ones; γ (h) is the semivariance of the difference of the random variables between two locations with a distance h between them, Var Variogram is a function used to describe spatial variability.

[0058] Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; The flood hydraulic elements in step 5 include water accumulation depth, water accumulation range, water accumulation volume, risk level and pipe network load.

[0059] Step 6: The spatial features of the grid rainfall data and the multi-source fusion grid rainfall constructed in step 4 are extracted through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through filters, while the pooling layer downsamples the local area to reduce the feature space dimension. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model for subsequent model input. Step 6 is implemented as follows: A convolutional neural network (CNN) algorithm is used to extract the characteristic parameters of grid rainfall. The network structure consists of an input layer, a convolution layer, an activation layer, and a pooling layer, which are connected in sequence. The input layer receives rainfall grid data at different times. The convolution layer performs a sliding operation on the input data using multiple convolution kernels, automatically extracting spatial feature information such as rainfall intensity distribution and concentrated areas within the local area. The pooling layer further processes the rainfall feature matrix output by the convolution layer and downsamples by calculating the maximum or average value of the local area, thereby effectively compressing the feature dimension and reducing the amount of calculation. The specific calculation method is as follows: (6) (7) (8)

[0060] Save the grid rainfall feature parameter information automatically extracted by the CNN model into an .h5 file, and use this feature information as model input in step 7.

[0061] Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 7 is implemented as follows: The Transformer model is used to construct a rapid simulation and forecast model for urban flooding. The spatial characteristic parameter information of the grid rainfall obtained in step 6 and the flood data during the rainfall process of different grids obtained in step 5 are used as the input of the Transformer model. The model can include an encoder and a decoder, mainly composed of modules such as multi-head attention layer, feedforward neural network, residual connection and layer normalization. It has the advantages of high efficiency, flexibility and easy expansion. The calculation method is: (9) (10) Where: For multi-head output; Concat is the connection function; Head t For the t Output of the head; head i For the i Output of the head; W O is the concatenation matrix, which is used to concatenate the output sub-vectors of multiple heads to obtain a complete output vector; W Q , W K , W K is a learnable parameter matrix that maps the vector at each position in the input sequence to Q 、 K 、 V ; (11) (12) Where: The output of the current fully connected layer, HRepresents the output value of the previous layer; W 1 , W 2 is the learnable parameter matrix; b 1, b 2 is the learnable bias vector, , , , ; is the activation function.

[0062] The above method ultimately generates a rapid urban flood simulation and prediction model based on grid-based rainfall data. By inputting grid-based rainfall data from meteorological forecasts for different forecast periods, the model can quickly predict flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, along with water conservancy elements such as risk level and pipe network load. This output provides data support for subsequent risk assessments, visualization, and urban flood control scheduling and decision-making.

[0063] Step 8: Continuously optimize the model parameters in steps 6 and 7 to ultimately obtain a rapid urban flood simulation and forecasting model. By inputting meteorological forecast grid rainfall data for different forecast periods, this model can quickly forecast flood outcomes for each grid in the study area, as well as waterlogging depth, area, and volume at flood-prone locations, and water conservancy elements such as risk level and pipe network load, providing sufficient lead time for urban flood control emergencies.

[0064] Figure 4 This is a schematic diagram of interpolating the rainfall data from the weather station into grid point data. Figure 4 It can be seen that by performing spatial interpolation processing on discrete meteorological station data, the rainfall intensity on different grids is obtained, which can intuitively reflect the spatial distribution characteristics of rainfall and make up for the deficiency that traditional surface rainfall cannot depict the differences in local heavy rainfall.

[0065] Figure 5 This is a comparison chart of the accuracy between the predicted values ​​of the urban flood rapid simulation prediction model based on grid rainfall data and the simulated values ​​of the hydrological and hydrodynamic model. Figure 5 It can be seen that by comparing the errors between the simulated values ​​of the hydrological and hydrodynamic model and the predicted values ​​of the CNN-Transformer model for the waterlogged area, waterlogged depth and waterlogged area under different rainfall return periods, the average error is within 15%.

[0066] Figure 6 This is a comparison chart of the computational efficiency between the predicted values ​​of the urban flood rapid simulation prediction model based on grid rainfall data and the simulated values ​​of the hydrological and hydrodynamic model; Figure 6It can be seen that by calculating the average time consumption of the hydrological and hydrodynamic model and the CNN-Transformer model under 10 rainfall events, it is found that the CNN-Transformer model can speed up the process by 100-200 times compared with the hydrological and hydrodynamic model.

Claims

1. A rapid simulation and prediction method for urban flooding based on grid rainfall data, characterized by: A hydrological and hydrodynamic model is constructed based on urban basic data to simulate and capture the flood response characteristics under different terrain and land use types. The spatial distribution characteristics of urban rainfall are characterized based on refined grid rainfall data. Combined with monitoring station data and designed rainfall data, the Gaussian attenuation model and interpolation calculation method are used to construct grid data. The convolutional neural network (CNN) model is used to extract the spatial rainfall distribution characteristics in the grid rainfall data. The rainfall intensity, rainfall range and spatial correlation relationship characteristics are automatically identified through convolution and pooling operations. The flood process corresponding to the grid rainfall data is then trained and learned through the Transformer model. The complex spatiotemporal relationship between rainfall and flood process is captured through the self-attention mechanism, and finally a rapid flood simulation and prediction model is obtained.

2. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 1 is characterized in that: Please follow the steps below to implement it: Step 1: Collect basic urban data, build a high-precision hydrological and hydrodynamic model, and complete parameter setting for the hydrological and hydrodynamic model; Step 2: Based on the meteorological station rainfall data and the designed rainfall data generated by the local rainstorm formula, the meteorological stations in the study area are connected by Tyson polygons, and the perpendicular bisectors of each connecting line are drawn. The study area is divided into multiple polygons with the perpendicular bisectors as the boundaries. Each of the multiple polygons is the area controlled by the meteorological station, representing the rainfall control area of ​​the meteorological station; Step 3: Using the meteorological station control areas divided in step 1, assuming each control area as the center of the rainstorm, and using the Gaussian attenuation model to simulate the spatial variation of rainfall intensity, the rainfall intensity in different areas is calculated based on the principle that rainfall intensity gradually decays from the center to the adjacent control area with distance; Step 4: Based on the design rainfall data of the control area, different spatial interpolation methods are used to generate a gridded rainfall distribution process to make the rainfall distribution more refined; Step 5: Use the constructed hydrological and hydrodynamic model to simulate the flood process under rainfall in different grids, generate flood hydraulic elements, and create a data results database; Step 6: The spatial features of the grid rainfall are extracted from the designed grid rainfall data constructed in step 4 and the multi-source fusion actual grid rainfall through the convolutional neural network (CNN) model. The convolution layer extracts spatial feature information through the filter, while the pooling layer downsamples the local area. The generated rainfall spatial feature parameters are saved as an .h5 file through the CNN model. Step 7: Based on the Transformer model, the .h5 file of rainfall spatial characteristic parameters generated in step 6 and the flood data generated in step 5 are input to construct an urban flood rapid simulation and forecasting model and save it as a .h5 file; Step 8: Continuously optimize the model parameters in steps 6 and 7, and finally obtain a rapid simulation and forecasting model for urban floods. The model inputs the meteorological forecast grid rainfall data for different forecast periods in the future.

3. The method for rapid simulation and prediction of urban flooding based on grid rainfall data according to claim 2, characterized in that: The step 1 is specifically implemented according to the following steps: First, terrain elevation data and land use type data in the study area were processed into .asc raster files in GIS. Different infiltration parameters and Manning parameters were assigned to different land use types. A one-dimensional pipe network .inp file was constructed using the pipe network data of the study area. This input file was imported into the hydrological and hydrodynamic model. The parameters and reliability of the hydrological and hydrodynamic model were calibrated and verified based on historical rainfall and flood data. Secondly, for spatially distributed grid rainfall, the grids in the spatial grid rainfall are numbered as follows: R 1. R 2. R 3.…… R n In order to make the spatial distribution and grid size of the surface and rainfall consistent during the hydrodynamic model simulation, the numbered rainfall grid is resampled to be consistent with the surface terrain grid using the nearest neighbor interpolation method, and the rainfall process is assigned to each rainfall number using a .csv file: （1） Where, is the location of the target grid cell, The target grid The value of ( x , y ) is the pixel position closest to the target pixel center in the original grid, and for( x , y ) rounded to the nearest integer grid point.

4. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 3 is characterized in that: The step 2 is specifically implemented according to the following steps: Use Euclidean distance to calculate the distance between different weather stations, taking a certain weather station as the midpoint M i , calculate the slope and perpendicular bisector between each station, and divide the control area of ​​each meteorological station into S 1. S 2. S 3.…… S n .

5. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 4 is characterized in that: The step 3 is specifically implemented according to the following steps: The control area of ​​the weather station site is divided using Thiessen polygons, and each control area is taken as the center of the rainstorm. The Gaussian attenuation model is used to calculate the rainfall in the adjacent control area. The specific division method is as follows: （2） Where, represents the rainfall intensity at a distance d from the center of the rainstorm; d is the Euclidean distance between two sites, P i is the rainfall intensity at the center of the rainstorm, and σ is the standard deviation parameter that controls the attenuation.

6. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 5, characterized in that: The step 4 is specifically implemented according to the following steps: The designed rainfall data of the controlled area of ​​the meteorological station is used to interpolate the grid rainfall process, and the Kriging interpolation method is used as follows: （3） （4） （5） Where, Represents the unknown point S An estimated value of 0; is the weight coefficient; For location Known samples at n is the number of known sample points, C is the covariance between two known samples; d is the covariance between the known sample and the predicted sample; λ is the weight coefficient vector, μ is the Lagrange multiplier, 1 is the vector of all ones; γ (h) is the semivariance of the difference of the random variables between two locations with a distance h between them, Var is the variogram, which is used to describe spatial variability.

7. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 6, characterized in that: The flood hydraulic elements in step 5 include water accumulation depth, water accumulation range, water accumulation volume, risk level and pipe network load.

8. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 7, characterized in that: The step 6 is specifically implemented according to the following steps: The convolutional neural network (CNN) algorithm is used to extract the characteristic parameters of grid rainfall. The network structure consists of an input layer, a convolution layer, an activation layer, and a pooling layer connected in sequence. The input layer receives rainfall grid data at different times. The convolution layer performs a sliding operation on the input data through multiple convolution kernels to automatically extract spatial feature information such as rainfall intensity distribution and concentrated areas in the local area. The pooling layer further processes the rainfall feature matrix output by the convolution layer and downsamples by calculating the maximum or average value of the local area, thereby effectively compressing the feature dimension and reducing the amount of calculation. The specific calculation method is as follows: （6） （7） （8） Save the grid rainfall feature parameter information automatically extracted by the CNN model into an .h5 file, and use this feature information as model input in step 7.

9. The method for rapid simulation and prediction of urban flooding based on grid point rainfall data according to claim 8, characterized in that: The step 7 is specifically implemented according to the following steps: The Transformer model is used to construct a rapid simulation and forecast model for urban floods. The grid rainfall spatial characteristic parameter information obtained in step 6 and the flood data during rainfall in different grids obtained in step 5 are used as inputs to the Transformer model. The model can include an encoder and a decoder. The calculation method is: （9） （10） Where: For multi-head output; Concat is the connection function; Head t For the t Output of the head; head i For the i Output of the head; W O is the concatenation matrix, which is used to concatenate the output sub-vectors of multiple heads to obtain a complete output vector; W Q , W K , W K is a learnable parameter matrix that maps the vector at each position in the input sequence to Q 、 K 、 V ; （11） （12） Where: The output of the current fully connected layer, H Represents the output value of the previous layer; W 1 , W 2 is the learnable parameter matrix; b 1, b 2 is the learnable bias vector, , , , ; is the activation function.

Images