A flood evolution prediction method, system, device and storage medium

Abstract

The present invention discloses a flood evolution prediction method, system, device and storage medium, wherein the system includes a data acquisition module for acquiring basic data of a target area; a mechanism model module for acquiring a target flood evolution mechanism model; a data construction module for extracting historical rainfall data of a target area in the basic data, and constructing multiple sets of rainfall meteorological boundary conditions based on the historical rainfall data; a sample set module for acquiring multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data to form a sample set; a training prediction module for constructing a graph neural network model, and calling the sample set to train the graph neural network model to obtain a prediction model for predicting flood evolution; and a result output module for outputting prediction results. By integrating the flood mechanism model and the graph neural network technology, the present invention can realize rapid prediction of the flood evolution process in the process of rainfall scenario changes, and is conducive to improving the accuracy of the prediction results.

Description

Technical Field

[0001] The present invention belongs to the technical field of flood prediction, and in particular relates to a flood evolution prediction method, system, device and storage medium. Background Art

[0002] Basin flood evolution prediction is a key technical means for flood prevention and disaster reduction. Traditional flood evolution prediction methods usually rely only on physical mechanism models. However, due to the complexity of flood evolution mechanism models, which involve a large number of nonlinear hydrological processes, especially at the basin scale, it is difficult to complete simulations under large-scale or complex conditions in a short time. In addition, although deep learning models can make rapid predictions, the limited measured data limits the prediction accuracy, thus affecting the timeliness, accuracy and effectiveness of flood warnings. Summary of the invention

[0003] Purpose of the invention: The first purpose of the present invention is to provide a flood evolution prediction method with fast prediction speed and high prediction accuracy.

[0004] A second object of the present invention is to provide a flood evolution prediction system.

[0005] A third object of the present invention is to provide an apparatus.

[0006] A fourth object of the present invention is to provide a storage medium.

[0007] Technical solution: The present invention discloses a flood evolution prediction method, comprising the following steps:

[0008] S1: Obtain basic data of the target area;

[0009] S2: Construct an initial flood evolution mechanism model, define and verify the initial flood evolution mechanism model based on basic data, and obtain the target flood evolution mechanism model;

[0010] S3: extract historical rainfall data of the target area in the basic data, and construct multiple sets of rainfall meteorological boundary conditions based on the historical rainfall data;

[0011] S4: input the constructed rainfall meteorological boundary conditions into the target flood evolution mechanism model, and the target flood evolution mechanism model outputs the water level spatiotemporal distribution data corresponding to the rainfall meteorological boundary conditions. Multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data constitute a sample set;

[0012] S5: Build a graph neural network model and use the sample set to train the graph neural network model to obtain a prediction model for predicting flood evolution;

[0013] S6: The user inputs the rainfall meteorological boundary conditions to be predicted into the target flood evolution mechanism model and the prediction model, the target flood evolution mechanism model outputs the first predicted water level spatiotemporal distribution data, and the prediction model outputs the second predicted water level spatiotemporal distribution data;

[0014] S7: Based on the first predicted water level spatiotemporal distribution data, determine whether the second predicted water level spatiotemporal distribution data is normal. If normal, output the second predicted water level spatiotemporal distribution data as the prediction result; if abnormal, expand the data samples under the rainfall meteorological boundary conditions, and add the expanded data samples to the sample set, and repeat steps S4-S7 until the second predicted water level spatiotemporal distribution data under the rainfall meteorological boundary conditions to be predicted is normal.

[0015] Furthermore, the basic data in step S1 include terrain elevation data, research scope, river data, underlying surface attribute data, meteorological data and calibration data; the meteorological data include rainfall time series data, temperature time series data, relative humidity time series data, wind speed time series data, radiation time series data, atmospheric pressure time series data and snowmelt time series data.

[0016] Furthermore, the initial flood evolution mechanism model in step S2 is inteliway-SSIM;

[0017] The steps to define and verify the initial flood evolution mechanism model are as follows:

[0018] Based on the terrain elevation data, research scope and river data, the research scope of the target basin is divided into grids of different sizes in Intelliway-SSIM, where the land area is a triangular grid and the river channel is a non-triangular grid;

[0019] Assign corresponding attributes to the triangular mesh based on the terrain elevation data and the underlying surface attribute data, and assign corresponding attributes to the non-triangular mesh based on the river data and the underlying surface attribute data;

[0020] Input meteorological data into Intelliway-SSIM and drive Intelliway-SSIM. Intelliway-SSIM outputs simulation values ​​and compares the simulation values ​​with the calibration data. If the error between the two exceeds the preset value, adjust the key parameters of the Intelliway-SSIM model or adjust the grid division of Intelliway-SSIM until the error is less than or equal to the preset value.

[0021] Furthermore, the historical rainfall data in step S3 refers to rainfall time series data;

[0022] The steps of constructing multiple sets of rainfall meteorological boundary conditions in step S3 are as follows:

[0023] The rainfall intensity-duration-rainfall frequency relationship in the target area is calculated based on historical rainfall data;

[0024] Design rainfall frequency and duration, and calculate the corresponding rainfall intensity based on the relationship between rainfall intensity-duration-rainfall frequency;

[0025] The time series data consisting of duration and corresponding rainfall intensity is defined as the rainfall meteorological boundary condition.

[0026] Furthermore, the calculation steps of the rainfall intensity-duration-rainfall frequency relationship are as follows:

[0027] Design rainfall frequency;

[0028] Fitting historical rainfall data to obtain rainfall frequencies of different rainfall intensities and durations;

[0029] By performing regression analysis on the rainfall frequency data of different rainfall intensities and durations, we obtain the relationship between rainfall intensity-duration-rainfall frequency: ,

[0030] , where I refers to rainfall intensity, in mm / h; D refers to rainfall duration, in h; R refers to return period, in year; f refers to rainfall frequency, and a, b, and c are constants.

[0031] Further, the steps of judging whether the second predicted water level spatiotemporal distribution data is normal in step S7 are as follows:

[0032] Calculate the difference between the second predicted water level spatiotemporal distribution data and the first predicted water level spatiotemporal distribution data;

[0033] A threshold is set. If the difference value does not exceed the threshold, the second predicted water level spatiotemporal distribution data is judged to be normal; if the difference value exceeds the threshold, the second predicted water level spatiotemporal distribution data is judged to be abnormal.

[0034] Furthermore, the steps of expanding the data sample in step S7 are as follows:

[0035] The meteorological boundary conditions of the rainfall to be predicted are expanded by random disturbance; the meteorological boundary conditions of the rainfall to be predicted are expanded by random disturbance in three ways: disturbance rainfall time, disturbance rainfall intensity and comprehensive disturbance rainfall time and rainfall intensity;

[0036] The expanded rainfall meteorological boundary conditions are input into the target flood evolution mechanism model to obtain the water level temporal and spatial distribution data corresponding to the expanded rainfall meteorological boundary conditions;

[0037] The expanded rainfall meteorological boundary conditions and the corresponding expanded water level spatiotemporal distribution data constitute the expanded data samples.

[0038] Based on the same inventive concept, the present invention also discloses a flood evolution prediction system, comprising:

[0039] Data acquisition module, to obtain basic data of the target area;

[0040] The mechanism model module is used to construct an initial flood evolution mechanism model, and define and verify the initial flood evolution mechanism model based on the basic data of the data acquisition module to obtain the target flood evolution mechanism model;

[0041] A data construction module is used to extract historical rainfall data of the target area in the basic data and construct multiple sets of rainfall meteorological boundary conditions based on the historical rainfall data;

[0042] The sample set module can input the constructed rainfall meteorological boundary conditions into the target flood evolution mechanism model, and the target flood evolution mechanism model outputs the water level spatiotemporal distribution data corresponding to the rainfall meteorological boundary conditions. Multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data constitute the sample set;

[0043] Train the prediction module, build a graph neural network model, and call the sample set to train the graph neural network model to obtain a prediction model for predicting flood evolution;

[0044] The result output module includes an integration submodule, a prediction judgment submodule, a data expansion submodule, a sample expansion submodule and a circulation submodule;

[0045] Integration submodule, used to integrate the target flood evolution mechanism model and prediction model;

[0046] The prediction and judgment submodule is used to input the rainfall meteorological boundary conditions to be predicted into the integration submodule and drive the integration submodule, the integration submodule outputs the first predicted water level spatiotemporal distribution data corresponding to the target flood evolution mechanism model, the integration submodule outputs the second predicted water level spatiotemporal distribution data corresponding to the prediction model, and then judges whether the second predicted water level spatiotemporal distribution data is normal based on the first predicted water level spatiotemporal distribution data. If normal, the second predicted water level spatiotemporal distribution data is output as the prediction result; if not normal, the command to expand the rainfall meteorological boundary conditions is transmitted to the sample expansion submodule and the sample expansion submodule is driven;

[0047] The sample expansion submodule expands the data samples under the rainfall meteorological boundary conditions, adds the expanded data samples to the sample set of the sample set module, and drives the loop submodule;

[0048] The loop submodule cyclically runs the training prediction module, the integration submodule, and the prediction judgment submodule until the prediction judgment submodule determines that the second prediction water level spatiotemporal distribution data under the rainfall meteorological boundary condition is normal.

[0049] Based on the same inventive concept, the present invention also discloses an electronic device, comprising one or more processors, one or more memories and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the flood evolution prediction method are implemented.

[0050] Based on the same inventive concept, the present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the steps of the flood evolution prediction method.

[0051] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: By integrating the flood mechanism model and graph neural network technology, the present invention can realize the rapid prediction of the flood evolution process in the process of rainfall scenario changes, and is conducive to improving the accuracy of the prediction results. The present invention reduces the excessive reliance on historical data by integrating the advantages of the physical mechanism model and the graph neural network model, and the present invention simultaneously adopts two methods to make up for the lack of measured data, so that the model can still provide accurate prediction results when the data is insufficient, and as the user's use time increases, the accuracy of the prediction results of this method continues to improve. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the method of the present invention;

[0053] Figure 2 The structure diagram of the definition and verification of the initial flood evolution mechanism model of the present invention is as follows Figure 1 ;

[0054] Figure 3 The structure diagram of the definition and verification of the initial flood evolution mechanism model of the present invention is as follows Figure 2 ;

[0055] Figure 4 The structure diagram of the definition and verification of the initial flood evolution mechanism model of the present invention is as follows Figure 3 ;

[0056] Figure 5 It is a structural schematic diagram of the graph neural network model of the present invention;

[0057] Figure 6 A schematic diagram of the temporal and spatial distribution of actual surface water depth output by the present invention;

[0058] Figure 7 A schematic diagram of the temporal and spatial distribution of the predicted surface flood water depth output by the present invention;

[0059] Figure 8 A comparison chart of the actual surface water depth and the predicted surface flood depth output by the present invention;

[0060] Fig. 9 It is a schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0062] Example 1

[0063] The present invention discloses a flood evolution prediction method, such as Figure 1 As shown, the following steps are included:

[0064] S1: Obtain basic data of the target area. Basic data include terrain elevation data, research scope, river data, underlying surface attribute data, meteorological data and calibration data. River data include river location, river cross-sectional shape and river hardening data. Surface attribute data include soil type, vegetation distribution, geological structure and land use type. Calibration data include river water level flow and groundwater level data. Meteorological data include rainfall time series data, temperature time series data, relative humidity time series data, wind speed time series data, radiation time series data, atmospheric pressure time series data and snowmelt time series data.

[0065] S2: Construct an initial flood evolution mechanism model, and define and verify the initial flood evolution mechanism model based on basic data to obtain the target flood evolution mechanism model.

[0066] Construct the inteliway-SSIM as the initial flood evolution mechanism model. The inteliway-SSIM is an existing computing platform, and the inteliway-SSIM is also called the surface water-groundwater coupling model. The model takes into account the complexity of surface hydrology, groundwater hydrology and their interactions, and can fully reflect the various hydrological dynamics in the process of flood evolution, and provide accurate support for flood warning and emergency decision-making. The model includes the following functions: one-dimensional river hydrodynamic process simulation, two-dimensional surface water process simulation, three-dimensional groundwater process simulation, two-way interactive simulation of surface water and groundwater, interactive simulation of river channel and surrounding surface groundwater, boundary interactive simulation and reservoir simulation. In this embodiment, the one-dimensional river hydrodynamic process simulation, two-dimensional surface water process simulation, three-dimensional groundwater process simulation, two-way interactive simulation of surface water and groundwater, interactive simulation of river channel and surrounding surface groundwater, and boundary interactive simulation functions of the model are used.

[0067] Defining and verifying the initial flood evolution mechanism model based on basic data includes the following steps:

[0068] (1) Based on the terrain elevation data, research area and river data, the research area of ​​the target basin is divided into grids of different sizes in Intelliway-SSIM, where the land area is a triangular grid and the river channel is a non-triangular grid. Figures 2 to 4 As shown in the figure, a grid is generated within the research scope, and the edges of the triangular grid are formed according to the river channel position of the river channel data. The terrain elevation data provides the elevation information of each pixel value, and then the DelaunayTriangulation triangulation method is used to divide the triangular grid, and the specific shape and size of the non-triangular grid are set according to the specific shape of the river channel cross section, such as V-type, U-type, etc.; the non-triangular grid of the river channel will also have a horizontal interaction with the triangular grid of the adjacent land area. The interaction here refers to the interaction of the surface water part and the groundwater part, that is, the interaction of water. Specifically, it refers to the horizontal interaction process between various grids, including the surface water production and convergence process, the groundwater runoff process, and the interaction process between groundwater and the river channel and lake body. For example, surface runoff converges into the river channel. When the water volume of the river channel exceeds its carrying capacity, the water body will overflow and form surface flow. The relationship between groundwater and river water is manifested as a hydraulic connection: when the groundwater level is higher than the river level, the groundwater is discharged into the river; conversely, when the river level is higher than the groundwater level, the river water recharges the groundwater system. The land area includes the vegetation layer, the surface layer, the shallow unsaturated soil layer, the deep unsaturated soil layer and the groundwater layer, and the river includes the surface layer, the shallow unsaturated soil layer, the deep unsaturated soil layer and the groundwater layer. According to the actual situation, some areas within the research scope are selected in advance as key areas, and the key areas are set as dense grids, and the non-key areas are set as sparse grids.

[0069] (2) Assign corresponding attributes to the triangular mesh based on the terrain elevation data and the underlying surface attribute data, and assign corresponding attributes to the non-triangular mesh based on the river data and the underlying surface attribute data. Figure 4 As shown, assigning attributes is to assign values ​​to each triangular mesh / non-triangular mesh according to characteristics such as soil type, land use type, vegetation distribution, geological structure, river cross-section shape, and river hardening data.

[0070] (3) Then input the meteorological data into Intelliway-SSIM and drive Intelliway-SSIM. Intelliway-SSIM outputs the simulation value and compares the simulation value with the calibration data. If the error between the two exceeds the preset value, adjust the key parameters of the Intelliway-SSIM model or adjust the grid division of Intelliway-SSIM until the error is less than or equal to the preset value. The key parameters of the model include the permeability coefficient of the river channel, the hydrological process parameters and the groundwater flow parameters; adjusting the grid division refers to adjusting the distribution of the encrypted grid and the sparse grid. Preferably, the adjustment of the key parameters of the model and the adjustment of the grid division are based on the actual situation and empirical values. The key parameters of the model are recognized parameters in the field of hydrology. These parameters have empirical value ranges. For example, the permeability coefficient of the river channel must first be based on the riverbed material (sand, fine sand, clay, etc.) according to the empirical range of this material. For example, the permeability coefficient of clay is between 10 -8 ~10 -5 m / s range, and select the appropriate value through continuous debugging.

[0071] S3: Extract historical rainfall data of the target area from the basic data, and construct multiple sets of rainfall meteorological boundary conditions based on the historical rainfall data, where the historical rainfall data refers to rainfall time series data.

[0072] The relationship between rainfall intensity, duration and rainfall frequency in the target area is calculated based on historical rainfall data. The steps to construct multiple sets of different rainfall meteorological boundary conditions based on historical rainfall data are as follows:

[0073] First, the rainfall frequency is designed, such as once in one year, once in 10 years, once in 50 years, etc. The rainfall frequency refers to the probability of a certain rainfall intensity occurring within a specific time.

[0074] The rainfall frequency analysis is performed on the historical rainfall data, and the extreme value distribution or normal distribution is used to fit the historical rainfall data to obtain the rainfall frequency of different rainfall intensities and durations.

[0075] Regression analysis is performed on the data of rainfall frequency with different rainfall intensities and durations to obtain the relationship between rainfall intensity-duration-rainfall frequency. Preferably, the least squares method is used to perform regression analysis, and the relationship is as follows: ,

[0076] , where I refers to rainfall intensity, in mm / h; D refers to rainfall duration, in h; R refers to recurrence period, in year; f refers to rainfall frequency, and a, b, and c are constants; that is, a, b, and c are calculated based on multiple sets of different rainfall intensities, durations, and rainfall frequencies.

[0077] The rainfall frequency and duration are designed according to the relationship between rainfall intensity, duration and rainfall frequency, and the rainfall intensity under different conditions is calculated. The time series data consisting of the duration and the corresponding rainfall intensity is defined as the rainfall meteorological boundary condition. The corresponding rainfall in each hour, all time points in the duration and their corresponding rainfall constitute the time series data, and all time points in the duration are included in each hour.

[0078] S4: Input the constructed rainfall meteorological boundary conditions into the target flood evolution mechanism model. The target flood evolution mechanism model outputs the water level spatiotemporal distribution data under the corresponding rainfall meteorological boundary conditions. Multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data constitute a sample set.

[0079] Use different rainfall meteorological boundary conditions to drive the target flood evolution mechanism model calculation in turn, and simulate the spatial and temporal distribution data of regional surface water levels. During the simulation process, the target flood evolution mechanism model will calculate the interaction between water bodies in the basin, and the calculation principle is water balance and mass balance. Specifically, it includes: inputting the river data and meteorological data of the target area to obtain the water flow velocity and water level of the river, inputting the land use data, soil data, vegetation distribution data and meteorological data of the target area to calculate the surface water flow, inputting the geological structure data, soil data, land use data and boundary data of the basin to calculate the spatial and temporal changes of groundwater infiltration and horizontal and vertical migration, inputting the head difference between surface water and groundwater to calculate the interactive water volume between surface water and groundwater, inputting the head difference between the river and the water body in the basin to calculate the interactive water volume between the river and nearby surface water and groundwater, thereby obtaining the interactive water volume of the basin to identify the evolution process of the basin flood. The specific formula is as follows:

[0080] a. Calculation of rainfall interception process: The calculation of rainfall interception process includes leaf interception and through-flow precipitation. The leaf interception is subject to the maximum water capacity of vegetation. The impact of The calculation formula is as follows: , where K L is the interception coefficient, which is usually set to 0.2; LAI is the leaf area index;

[0081] The calculation formula of leaf surface interception is: , where t is time, is the leaf interception, P is the rainfall, ET canopy is the evaporation from the leaves, TF is the throughfall precipitation;

[0082] The calculation formula of through-flow precipitation TF is: , , where b is the drainage coefficient (dimensionless), the value of b is usually between 3.0-4.6, and the unit of k is L / t, L refers to the length, and k is usually (mm / min).

[0083] b. Calculation of leaf evaporation process: , where Δ is the saturated vapor pressure curve at temperature t1 ( 0 C) slope, R n is the net radiation of vegetation surface (MJm -2 day -1 ), G is the soil heat flux (MJm -2 day -1 ), ρ a is the atmospheric density (kg / m 3 ), C p is the specific heat of air (MJkg -10 C -1 ), e s and e a are the saturated vapor pressure and actual vapor pressure (kPa), r a is the aerodynamic drag (sm -1 ), γ is the temperature constant (kPa 0 C -1 ), is the area ratio of wet vegetation.

[0084] c. Calculation of vegetation transpiration:

[0085] That is, vegetation extracts water from soil moisture to perform transpiration, and the calculation formula is as follows: , where vFrac is the vegetation coverage ratio, r s For vegetation c. Calculation of vegetation transpiration:

[0086] That is, vegetation extracts water from soil moisture to perform transpiration, and the calculation formula is as follows: , where vFrac is the vegetation coverage ratio, r s is the stomatal resistance of vegetation (sm -1 ).

[0087] d. Calculation of soil surface evapotranspiration process: , ,in, β s Reflects the effect of soil surface saturation on surface evaporation. vFrac is the vegetation cover ratio, θ fl =0.75 θ sat is the field water holding capacity, θ satis the saturated soil moisture content, θ g The moisture content of the soil surface.

[0088] e. Calculation of surface water processes: Based on the Saint-Venant equation of dynamic wave or diffusion wave approximation, the vertical approximation is a completely mixed state. The calculation formula is: , ,in, is the gradient of the surface water head in the direction of maximum slope, d ij is the distance between adjacent grids i and j, is the water depth of the adjacent grid i, z i is the elevation of the adjacent grid i, is the water depth of the adjacent grid j, z j is the elevation of the adjacent grid j, n s is the Manning coefficient. At the same time, the head gradient needs to be calculated between all adjacent grids to determine the water flow value in each direction.

[0089] f. Calculation process of unsaturated soil layer: The change process of unsaturated soil layer can also be called infiltration process. The unsaturated layer is the part above the groundwater level and below the soil surface. Since the groundwater level will be replenished by the seepage water under the soil surface, the groundwater level will change continuously, causing the depth and saturation of the unsaturated layer to change over time. The saturation calculation of the unsaturated layer takes into account the infiltration water from the soil surface, and also takes into account the groundwater rising due to capillary action. The direction of these water flows will change with the interaction of rainfall events and evaporation. The calculation formula is: , , where ReI is the amount of water infiltrating from the surface soil to the unsaturated layer, Re is the amount of water infiltrating from the unsaturated layer soil to the groundwater, ψ3 is the surface soil head (m), K(ψ3) is the hydraulic conductivity (L / t), ψ4 is the groundwater level, ψ5 is the unsaturated layer soil head, K(ψ5) is the unsaturated layer soil hydraulic conductivity; z3 is the surface soil elevation, z4 is the groundwater elevation, and z5 is the unsaturated layer soil elevation. Here, according to the changes in water flow between these layers, the direction of water flow may be reversed, such as groundwater transferring water to the unsaturated layer above through capillary action. Therefore, in actual calculations, this formula will determine the value of water flow direction and hydraulic conductivity based on the real-time head gradient.

[0090] g. Calculation process of groundwater layer: The horizontal flow of groundwater is Darcy flow, which is mainly determined by the hydraulic conductivity and the hydraulic gradient between adjacent grids. The calculation formula is: , where ψ 4i and ψ 4j is the groundwater level of the adjacent grid i and grid j (m), z bi and zbj is the elevation of the adjacent grid i and grid j (m), d ij is the distance between adjacent grids i and j, K(ψ ij ) is the hydraulic conductivity between adjacent grids i and j.

[0091] h. Calculation of surface water-groundwater layer calculation process: The surface water infiltration and groundwater seepage process are mainly determined by the hydraulic difference between surface water and surface soil water. The calculation formula is: , , where ψ2 is the surface water level (m), z is the surface elevation (m), ψ3 is the surface soil head (m), K(ψ3) is the hydraulic conductivity (L / t), K(ψ) is the hydraulic conductivity between surface water and groundwater, and z u is the topsoil elevation (m), d is the distance between surface water and topsoil (m), K s is the saturated hydraulic conductivity (L / t), α and n are Van Genuchten parameters, and the unit of α is L -1 .

[0092] S5: Construct a graph neural network model, and use the sample set to train the graph neural network model to obtain a prediction model for predicting flood evolution. Preferably, the sample set is divided into a training set and a test set according to a preset ratio, and the training set is used to train the graph neural network model to obtain a prediction model, and then the test set is used to test the trained prediction model, so as to facilitate testing the accuracy of the prediction model.

[0093] The triangular mesh of the land is defined as the node of the graph neural network model. The feature vector of the node includes the meteorological characteristics and terrain characteristics of the triangular mesh. The meteorological characteristics refer to the meteorological data, and the terrain characteristics refer to the terrain elevation data. The river flow and surface water flow are defined as the edges of the graph neural network model. The edges are the connections between nodes. In the graph neural network model, the edges are the relationship between the river flow and the surface water flow, and the topological relationship between the river flow and the surface water flow, that is, the connection and interaction between them in space and time. Based on the spatiotemporal distribution of water levels output by the target flood evolution mechanism model, the river flow and surface water flow, such as flow direction, flow, water level, and flow velocity, can be calculated. Figure 5 As shown, Figure 5 Here X1-X6 refers to nodes, and h1-h6 refers to feature vectors of nodes.

[0094] The functions of information transmission and node update of the graph neural network model are as follows:

[0095] Information transfer: In the information transfer mechanism, nodes are updated according to the characteristics of adjacent nodes after each iteration. The updated node characteristics contain more information about its adjacent nodes. With multiple iterations, the node characteristics will be gradually enriched, which can better express the relationship between nodes in the topology graph. Information transfer function: ,in It represents the aggregation of the information received by node v at layer k to the adjacent nodes. represents the aggregation of the neighboring nodes of node v, refers to the message passing function, and They refer to the topographic and meteorological characteristics of nodes u and v at the k-1 layer, respectively. It refers to the characteristics of the edge (u,v).

[0096] Node update: The node update mechanism helps to fuse node information to a deeper level, allowing the model to make better predictions on unseen data. This update mechanism can help the model identify more complex patterns and trends, thereby improving the generalization ability of the model. The node update function is as follows: ,in, refers to the feature representation of node v at layer k, is the update function to calculate the representation of the node at the kth layer, It refers to the information received by node v at layer k.

[0097] Preferably, the obtained prediction model is verified and optimized, including but not limited to the following verification optimization strategies: improving the prediction accuracy of the prediction model through the loss function, such as selecting the NSE function as the loss function to measure the prediction error; improving the prediction accuracy of the prediction model through the regularization strategy, regularization refers to reducing the impact of overfitting on the model accuracy by adding penalty terms; improving the accuracy and generalization performance of the prediction model through hyperparameter tuning, hyperparameter tuning refers to presetting parameters and then finding the best parameter combination through grid search, random search or cross-validation.

[0098] S6: The user inputs the rainfall meteorological boundary conditions to be predicted into the target flood evolution mechanism model and the prediction model, and the target flood evolution mechanism model and the prediction model output the first predicted water level spatiotemporal distribution data and the second predicted water level spatiotemporal distribution data respectively.

[0099] S7: Based on the first predicted water level spatiotemporal distribution data, determine whether the second predicted water level spatiotemporal distribution data is normal. If normal, output the second predicted water level spatiotemporal distribution data as the prediction result; if abnormal, expand the data sample under the rainfall meteorological boundary condition, and add the expanded data sample to the sample set, and repeat steps S4-S7 until the second predicted water level spatiotemporal distribution data under the rainfall meteorological boundary condition to be predicted is normal. Preferably, in actual use, according to user needs, the relevant data of flood evolution can be counted and output based on the second predicted water level spatiotemporal distribution data, including but not limited to flood depth, flood diffusion range, flooding time, etc. Figure 7 As shown, the output is the time series data of flood depth and inundation time, that is, a schematic diagram of the predicted spatiotemporal distribution of surface flood depth. The time series data can be used to draw a spatiotemporal distribution diagram of the water level, thereby obtaining the flood diffusion range. Figure 6 Schematic diagram of the temporal and spatial distribution of actual surface water depth. Figure 8 A comparison chart of actual surface water depth and predicted surface flood depth.

[0100] The second predicted water level spatiotemporal distribution data is compared with the first predicted water level spatiotemporal distribution data to calculate the difference between the two, and a threshold is set. If the difference does not exceed the threshold, the second predicted water level spatiotemporal distribution data is judged to be normal; if the difference exceeds the threshold, the second predicted water level spatiotemporal distribution data is judged to be abnormal. The preferred threshold is 30%.

[0101] The difference value is calculated using the mean absolute percentage error, and the calculation formula is as follows: , where n is the sample size, Refers to the spatiotemporal distribution data of the second predicted water level of the ith i Refers to the spatiotemporal distribution data of the first predicted water level of the ith time. In actual use, users can choose different mathematical formulas to calculate the difference value according to actual needs and actual conditions, such as using absolute error, mean square error, logarithmic error, etc. to calculate the difference value.

[0102] The rainfall meteorological boundary conditions to be predicted are expanded by random disturbance, and the expanded rainfall meteorological boundary conditions are input into the target flood evolution mechanism model to obtain the water level spatiotemporal distribution data corresponding to the expanded rainfall meteorological boundary conditions, and the expanded rainfall meteorological boundary conditions and the corresponding expanded spatiotemporal distribution data of water level constitute a data sample, and the data sample is added to the sample set, and the graph neural network model is retrained in step S4 until the difference value under the rainfall meteorological boundary conditions to be predicted does not exceed the threshold.

[0103] There are three ways to randomly perturb and expand the boundary conditions of the rainfall meteorology to be predicted:

[0104] 1. Disturbance of rainfall timing

[0105] The temporal characteristics of rainfall events (e.g., start time, duration) may affect hydrological processes, especially in flood simulations. To capture the spatiotemporal variability of rainfall, the rainfall timing can be perturbed in the following ways:

[0106] Disturbance rainfall starts at:

[0107] By randomly perturbing the start time of rainfall, we can simulate different start times of rainfall events. Specifically, we can randomly select the start time of rainfall within a fixed time window. For example, if the start time of rainfall is T start , can be obtained by adding a random perturbation value ∆T start , so that the starting time after the disturbance is T start +∆T start .

[0108] Random perturbations can be implemented by sampling from a uniform or normal distribution, for example:

[0109] ∆T start ∈(−ΔT max ,ΔT max ), where ΔT max It is the maximum time range of the disturbance, which can usually be set based on historical data or expert experience.

[0110] Disturbance rainfall duration:

[0111] The duration of the rainfall event (i.e. how long it rains) is also a key factor. The duration of the rainfall can be randomly perturbed to simulate scenarios with different rainfall persistence.

[0112] If the duration of a rainfall event is T duration , then through the random perturbation ∆T duration To get a new duration: T duration =T duration +ΔT duration Where, ∆T duration You can sample from a distribution, such as a normal distribution or a uniform distribution.

[0113] Disturbance rainfall interval:

[0114] The intervals between multiple rainfall events are disturbed to simulate different rainfall intervals. The perturbation method is similar, using random distribution to generate different intervals.

[0115] 2. Disturbance of rainfall intensity

[0116] Rainfall intensity is an important factor affecting basin runoff and hydrological processes. Disturbing rainfall intensity helps simulate the impact of changes in rainfall intensity on hydrological processes. The following methods can be used for disturbance:

[0117] Spatial distribution of disturbance rainfall intensity:

[0118] Rainfall intensity often has spatial variability, and the spatial distribution of rainfall intensity can be perturbed. For example, suppose an existing rainfall intensity map (based on meteorological observations or historical data) can be used to generate a new rainfall intensity field by adding noise.

[0119] The rainfall intensity is disturbed in the following way:

[0120] I perturbed =I original +ε

[0121] Among them, I perturbed Refers to the rainfall intensity after the disturbance, I original refers to the original rainfall intensity, ε is random noise that can be sampled from a normal distribution or a uniform distribution, and represents the range of rainfall intensity.

[0122] The size (standard deviation) of the random noise can be adjusted according to the range of rainfall intensity to ensure that the rainfall intensity after disturbance remains within a reasonable range.

[0123] Temporal variation of disturbance rainfall intensity:

[0124] The rainfall intensity usually varies over time, and different rainfall patterns can be simulated by perturbing the rainfall intensity over time. For example, if the rainfall intensity increases or decreases linearly over a period of time, you can add random perturbations to the trend:

[0125] I(t2)=I0(t2)+ΔI(t2)

[0126] Where t2 refers to time, I(t2) is the intensity after perturbation, I0(t2) is the original intensity, and ΔI(t2) is the perturbation of the intensity change over time, which can be generated by random distribution.

[0127] Extreme values ​​of disturbance rainfall intensity:

[0128] In rainfall events, extreme rainfall intensity (such as heavy rainfall in a short period of time) has a greater impact on floods. Extreme rainfall events can be simulated to disturb the rainfall intensity at certain time points to make it more extreme. This can be done by: I(t2)=I0(t2)+βI0(t2)

[0129] Among them, β is an amplification factor, which is usually greater than 1 and is used to simulate extreme weather conditions.

[0130] 3. Comprehensive perturbation strategy

[0131] In addition to perturbing rainfall time and rainfall intensity separately, you can also combine the two to perform a joint perturbation. This can help simulate more complex rainfall event scenarios:

[0132] Comprehensive perturbations: Simultaneously perturb the start time, duration, and intensity of rainfall to generate multiple different rainfall scenarios for simulation. For example, noise or random perturbations can be added to the start time and intensity of each rainfall event.

[0133] The flood evolution prediction method disclosed in the present invention determines whether the prediction result of the prediction model is normal by comparing the prediction results of the prediction model with the prediction results of the target flood evolution mechanism model. If the prediction result is abnormal, the number of samples in the sample set will be expanded, and the prediction model will be retrained to improve the prediction accuracy of the prediction model. When the sample set in step S3 is initially constructed based on historical rainfall data, the samples in the initially constructed sample set will not be comprehensive due to the objective data volume limitation of the historical rainfall data and the limitation of the subjective rainfall meteorological boundary conditions designed by the designers. As the user uses the prediction model and the target flood evolution mechanism model for a longer time, the samples of different rainfall meteorological boundary conditions will be continuously expanded, and the number and types of samples in the sample set will be increased. In particular, when facing new and unseen rainfall meteorological boundary conditions, the prediction model can make more accurate predictions.

[0134] In summary, the flood evolution prediction method disclosed in the present invention has two ways to compensate for the limited measured data. The first way is to calculate the relationship between rainfall intensity, duration and rainfall frequency based on historical rainfall data, and then construct multiple groups of different rainfall meteorological boundary conditions based on rainfall intensity, duration and rainfall frequency; the second way is that during the user's use, the user inputs different rainfall meteorological boundary conditions into the prediction model and the target flood evolution mechanism model according to the use requirements. If the second predicted water level spatiotemporal distribution data under the rainfall meteorological boundary condition is abnormal, the rainfall meteorological boundary condition will be expanded by random disturbance, and the expanded rainfall meteorological boundary condition to be predicted will be input into the target flood evolution mechanism model, the number of samples in the sample set will be increased, and the prediction model will be retrained using the increased sample set.

[0135] The present invention can improve the initiative and foresight of flood warnings. Traditional flood management models usually rely on passive responses after floods occur, lack foresight and initiative, and are often unable to timely predict the occurrence and evolution of floods, especially when dealing with complex climate changes and extreme weather events, resulting in delayed decision-making responses. By integrating flood mechanism models and graph neural network technology, the present invention can achieve rapid prediction of flood evolution processes during changes in rainfall scenarios. The model can warn of floods in advance and provide timely information on the spatiotemporal distribution of water levels at different time intervals, thereby effectively enhancing the flood response capabilities of cities and river basins and improving the foresight of flood control decisions.

[0136] The present invention is conducive to enhancing the accuracy and real-time performance of flood prediction. Existing flood prediction methods often have large prediction errors when faced with lack of or incomplete data, affecting the accuracy and timeliness of early warnings. In particular, traditional hydraulic models often require a large amount of historical data for training and verification when faced with complex hydrological processes, resulting in prediction results that are limited by data quality and quantity. By integrating the Intelliway-SSIM (surface water-groundwater coupling model) and graph neural network technology, the present invention can improve the accuracy and timeliness of predictions in the real-time prediction process, solving the problem that traditional models cannot quickly respond to and process early warnings under complex hydrological conditions, ensuring that flood prediction results are more accurate and early warning information is more timely.

[0137] In the prior art, many flood evolution prediction models are based on traditional physical mechanism models. These models are computationally complex and time-consuming when dealing with complex hydrological processes. Especially in flood prediction of large-scale river basins, it is difficult to provide effective support in a short period of time. This application realizes large-scale flood evolution prediction in a short period of time by integrating graph neural network technology. This not only improves the real-time decision-making, but also makes flood simulation under complex conditions more efficient, and can guide decision-making in real time and quickly respond to sudden flood risks.

[0138] Traditional machine learning models rely on a large amount of historical observation data for training. However, when the amount of historical data is insufficient or the data is not representative enough, the prediction results are likely to deviate from the actual situation, affecting the accuracy of flood warnings. This invention reduces excessive reliance on historical data by integrating the advantages of physical mechanism models and graph neural network technology, allowing the model to provide accurate prediction results even when data is insufficient.

[0139] Example 2

[0140] The present invention discloses a flood evolution prediction system, such as Fig. 9 As shown, it includes a data acquisition module, a mechanism model module, a data construction module, a sample set module, a training prediction module and a result output module.

[0141] The data acquisition module acquires the basic data of the target area. The basic data include terrain elevation data, research scope, river data, underlying surface attribute data, meteorological data and calibration data. The river data include river location, river cross-sectional shape and river hardening data. The underlying surface attribute data include soil type, vegetation distribution, geological structure and land use type. The calibration data includes river water level flow and groundwater level data. The meteorological data include rainfall time series data, temperature time series data, relative humidity time series data, wind speed time series data, radiation time series data, atmospheric pressure time series data and snowmelt time series data.

[0142] The mechanism model module is used to construct an initial flood evolution mechanism model, and define and verify the initial flood evolution mechanism model based on the basic data of the data acquisition module to obtain the target flood evolution mechanism model.

[0143] Construct the inteliway-SSIM as the initial flood evolution mechanism model. The inteliway-SSIM is an existing computing platform, and the inteliway-SSIM is also called the surface water-groundwater coupling model. The model takes into account the complexity of surface hydrology, groundwater hydrology and their interactions, and can fully reflect the various hydrological dynamics in the process of flood evolution, and provide accurate support for flood warning and emergency decision-making. The model includes the following functions: one-dimensional river hydrodynamic process simulation, two-dimensional surface water process simulation, three-dimensional groundwater process simulation, two-way interactive simulation of surface water and groundwater, interactive simulation of river channel and surrounding surface groundwater, boundary interactive simulation and reservoir simulation. In this embodiment, the one-dimensional river hydrodynamic process simulation, two-dimensional surface water process simulation, three-dimensional groundwater process simulation, two-way interactive simulation of surface water and groundwater, interactive simulation of river channel and surrounding surface groundwater, and boundary interactive simulation functions of the model are used.

[0144] Defining and verifying the initial flood evolution mechanism model based on basic data includes the following steps:

[0145] (1) Based on the terrain elevation data, research area and river data, the research area of ​​the target basin is divided into grids of different sizes in Intelliway-SSIM, where the land area is a triangular grid and the river channel is a non-triangular grid. Figures 2 to 4As shown in the figure, a grid is generated within the research scope, and the edges of the triangular grid are formed according to the river channel position of the river channel data. The terrain elevation data provides the elevation information of each pixel value, and then the DelaunayTriangulation triangulation method is used to divide the triangular grid, and the specific shape and size of the non-triangular grid are set according to the specific shape of the river channel cross section, such as V-type, U-type, etc.; the non-triangular grid of the river channel will also have a horizontal interaction with the triangular grid of the adjacent land area. The interaction here refers to the interaction of the surface water part and the groundwater part, that is, the interaction of water. Specifically, it refers to the horizontal interaction process between various grids, including the surface water production and convergence process, the groundwater runoff process, and the interaction process between groundwater and the river channel and lake body. For example, surface runoff converges into the river channel. When the water volume of the river channel exceeds its carrying capacity, the water body will overflow and form surface flow. The relationship between groundwater and river water is manifested as a hydraulic connection: when the groundwater level is higher than the river level, the groundwater is discharged into the river; conversely, when the river level is higher than the groundwater level, the river water recharges the groundwater system. The land area includes the vegetation layer, the surface layer, the shallow unsaturated soil layer, the deep unsaturated soil layer and the groundwater layer, and the river includes the surface layer, the shallow unsaturated soil layer, the deep unsaturated soil layer and the groundwater layer. According to the actual situation, some areas within the research scope are selected in advance as key areas, and the key areas are set as dense grids, and the non-key areas are set as sparse grids.

[0146] (2) Assign corresponding attributes to the triangular mesh based on the terrain elevation data and the underlying surface attribute data, and assign corresponding attributes to the non-triangular mesh based on the river data and the underlying surface attribute data. Figure 4 As shown, assigning attributes is to assign values ​​to each triangular mesh / non-triangular mesh according to characteristics such as soil type, land use type, vegetation distribution, geological structure, river cross-section shape, and river hardening data.

[0147] (3) Then input the meteorological data into Intelliway-SSIM and drive Intelliway-SSIM. Intelliway-SSIM outputs the simulation value and compares the simulation value with the calibration data. If the error between the two exceeds the preset value, adjust the key parameters of the Intelliway-SSIM model or adjust the grid division of Intelliway-SSIM until the error is less than or equal to the preset value. The key parameters of the model include the permeability coefficient of the river channel, the hydrological process parameters and the groundwater flow parameters; adjusting the grid division refers to adjusting the distribution of the encrypted grid and the sparse grid. Preferably, the adjustment of the key parameters of the model and the adjustment of the grid division are based on the actual situation and empirical values. The key parameters of the model are recognized parameters in the field of hydrology. These parameters have empirical value ranges. For example, the permeability coefficient of the river channel must first be based on the riverbed material (sand, fine sand, clay, etc.) according to the empirical range of this material. For example, the permeability coefficient of clay is between 10 -8 ~10 -5 m / s range, and select the appropriate value through continuous debugging.

[0148] The data construction module is used to extract the historical rainfall data of the target area in the basic data and construct multiple sets of rainfall meteorological boundary conditions based on the historical rainfall data. The historical rainfall data refers to the rainfall time series data.

[0149] The relationship between rainfall intensity, duration and rainfall frequency in the target area is calculated based on historical rainfall data. The steps to construct multiple sets of different rainfall meteorological boundary conditions based on historical rainfall data are as follows:

[0150] First, the rainfall frequency is designed, such as once in one year, once in 10 years, once in 50 years, etc. The rainfall frequency refers to the probability of a certain rainfall intensity occurring within a specific time.

[0151] The rainfall frequency analysis is performed on the historical rainfall data, and the extreme value distribution or normal distribution is used to fit the historical rainfall data to obtain the rainfall frequency of different rainfall intensities and durations.

[0152] Regression analysis is performed on the data of rainfall frequency with different rainfall intensities and durations to obtain the relationship between rainfall intensity-duration-rainfall frequency. Preferably, the least squares method is used to perform regression analysis, and the relationship is as follows: , , where I refers to rainfall intensity, in mm / h; D refers to rainfall duration, in h; R refers to recurrence period, in year; f refers to rainfall frequency, and a, b, and c are constants; that is, a, b, and c are calculated based on multiple sets of different rainfall intensities, durations, and rainfall frequencies.

[0153] The rainfall frequency and duration are designed according to the relationship between rainfall intensity, duration and rainfall frequency, and the rainfall intensity under different conditions is calculated. The time series data consisting of the duration and the corresponding rainfall intensity is defined as the rainfall meteorological boundary condition. The corresponding rainfall in each hour, all time points in the duration and their corresponding rainfall constitute the time series data, and all time points in the duration are included in each hour.

[0154] The sample set module can input the constructed rainfall meteorological boundary conditions into the target flood evolution mechanism model. The target flood evolution mechanism model outputs the water level spatiotemporal distribution data under the corresponding rainfall meteorological boundary conditions. Multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data constitute the sample set.

[0155] Different rainfall meteorological boundary conditions are used to drive the target flood evolution mechanism model calculation in turn, and the temporal and spatial distribution data of regional surface water levels are simulated. During the simulation process, the target flood evolution mechanism model will calculate the interaction between water bodies in the basin, and the calculation principle is water balance and mass balance. Specifically include: inputting the river data and meteorological data of the target area to obtain the water flow velocity and water level of the river, inputting the land use data, soil data, vegetation distribution data and meteorological data of the target area to calculate the surface water flow, inputting the geological structure data, soil data, land use data and boundary data of the basin to calculate the spatiotemporal changes of groundwater infiltration and horizontal and vertical migration, inputting the head difference between surface water and groundwater to calculate the interactive water volume between surface water and groundwater, inputting the head difference between the river and the water body in the basin to calculate the interactive water volume between the river and nearby surface water and groundwater, thereby obtaining the interactive water volume of the basin to identify the evolution process of basin floods. The specific formula is as follows: a. Calculation of rainfall interception process: The calculation of rainfall interception process includes leaf interception part and through-flow precipitation; among which the leaf interception part is affected by the maximum capacity of vegetation water. The impact of The calculation formula is as follows: , where K L is the interception coefficient, which is usually set to 0.2; LAI is the leaf area index;

[0156] The calculation formula of leaf surface interception is: , where t is time, is the leaf interception, P is the rainfall, ET canopy is the evaporation from the leaves, TF is the throughfall precipitation;

[0157] The calculation formula of through-flow precipitation TF is: , , where b is the drainage coefficient (dimensionless), the value of b is usually between 3.0-4.6, and the unit of k is L / t, L refers to the length, and k is usually (mm / min).

[0158] b. Calculation of leaf evaporation process: , where Δ is the saturated vapor pressure curve at temperature t1 ( 0 C) slope, R n is the net radiation of vegetation surface (MJm -2 day -1 ), G is the soil heat flux (MJm -2 day -1 ), ρ a is the atmospheric density (kg / m 3 ), C p is the specific heat of air (MJkg -10 C -1 ), e s and e a are the saturated vapor pressure and actual vapor pressure (kPa), r a is the aerodynamic drag (sm -1 ), γ is the temperature constant (kPa 0 C -1 ), is the area ratio of wet vegetation.

[0159] c. Calculation of vegetation transpiration:

[0160] That is, vegetation extracts water from soil moisture to perform transpiration, and the calculation formula is as follows: , where vFrac is the vegetation coverage ratio, r s For vegetation c. Calculation of vegetation transpiration:

[0161] That is, vegetation extracts water from soil moisture to perform transpiration, and the calculation formula is as follows: , where vFrac is the vegetation coverage ratio, r s is the stomatal resistance of vegetation (sm -1 ).

[0162] d. Calculation of soil surface evapotranspiration process: , ,in, β s Reflects the effect of soil surface saturation on surface evaporation. vFrac is the vegetation cover ratio, θ fl =0.75 θ sat is the field water holding capacity, θ sat is the saturated soil moisture content, θ g The moisture content of the soil surface.

[0163] e. Calculation of surface water processes: Based on the Saint-Venant equation of dynamic wave or diffusion wave approximation, the vertical approximation is a completely mixed state. The calculation formula is: , ,in, is the gradient of the surface water head in the direction of maximum slope, d ij is the distance between adjacent grids i and j, is the water depth of the adjacent grid i, z i is the elevation of the adjacent grid i, is the water depth of the adjacent grid j, z j is the elevation of the adjacent grid j, n s is the Manning coefficient. At the same time, the head gradient needs to be calculated between all adjacent grids to determine the water flow value in each direction.

[0164] f. Calculation process of unsaturated soil layer: The change process of unsaturated soil layer can also be called infiltration process. The unsaturated layer is the part above the groundwater level and below the soil surface. Since the groundwater level will be replenished by the seepage water under the soil surface, the groundwater level will change continuously, causing the depth and saturation of the unsaturated layer to change over time. The saturation calculation of the unsaturated layer takes into account the infiltration water from the soil surface, and also takes into account the groundwater rising due to capillary action. The direction of these water flows will change with the interaction of rainfall events and evaporation. The calculation formula is: , , where ReI is the amount of water infiltrating from the surface soil to the unsaturated layer, Re is the amount of water infiltrating from the unsaturated layer soil to the groundwater, ψ3 is the surface soil head (m), K(ψ3) is the hydraulic conductivity (L / t), ψ4 is the groundwater level, ψ5 is the unsaturated layer soil head, K(ψ5) is the unsaturated layer soil hydraulic conductivity; z3 is the surface soil elevation, z4 is the groundwater elevation, and z5 is the unsaturated layer soil elevation. Here, according to the changes in water flow between these layers, the direction of water flow may be reversed, such as groundwater transferring water to the unsaturated layer above through capillary action. Therefore, in actual calculations, this formula will determine the value of water flow direction and hydraulic conductivity based on the real-time head gradient.

[0165] g. Calculation process of groundwater layer: The horizontal flow of groundwater is Darcy flow, which is mainly determined by the hydraulic conductivity and the hydraulic gradient between adjacent grids. The calculation formula is: , where ψ 4i and ψ 4j is the groundwater level of the adjacent grid i and grid j (m), z bi and z bj is the elevation of the adjacent grid i and grid j (m), d ijis the distance between adjacent grids i and j, K(ψ ij ) is the hydraulic conductivity between adjacent grids i and j.

[0166] h. Calculation of surface water-groundwater layer calculation process: The surface water infiltration and groundwater seepage process are mainly determined by the hydraulic difference between surface water and surface soil water. The calculation formula is: , , where ψ2 is the surface water level (m), z is the surface elevation (m), ψ3 is the surface soil head (m), K(ψ3) is the hydraulic conductivity (L / t), K(ψ) is the hydraulic conductivity between surface water and groundwater, and z u is the topsoil elevation (m), d is the distance between surface water and topsoil (m), K s is the saturated hydraulic conductivity (L / t), α and n are Van Genuchten parameters, and the unit of α is L -1 .

[0167] The prediction module is trained to construct a graph neural network model, and the sample set is called to train the graph neural network model to obtain a prediction model for predicting flood evolution. Preferably, the sample set is divided into a training set and a test set according to a preset ratio, and the training set is used to train the graph neural network model to obtain a prediction model, and then the test set is used to test the trained prediction model, so as to facilitate the accuracy of the prediction model.

[0168] The triangular mesh of the land area is defined as the node of the graph neural network model. The feature vector of the node includes the meteorological characteristics and terrain characteristics of the triangular mesh. The meteorological characteristics refer to the meteorological data, and the terrain characteristics refer to the terrain elevation data. The river flow and surface water flow are defined as the edges of the graph neural network model. The edges are the connections between nodes. In this graph neural network model, the edges are the relationship between the river flow and the surface water flow, and the topological relationship between the river flow and the surface water flow, that is, the connection and interaction between them in space and time. Based on the spatiotemporal distribution of water levels output by the target flood evolution mechanism model, the river flow and surface water flow, such as flow direction, flow, water level, and flow velocity, can be calculated. Figure 5 As shown, Figure 5 Here X1-X6 refers to nodes, and h1-h6 refers to feature vectors of nodes.

[0169] The functions of information transmission and node update of the graph neural network model are as follows:

[0170] Information transfer: In the information transfer mechanism, nodes are updated according to the characteristics of adjacent nodes after each iteration. The updated node characteristics contain more information about its adjacent nodes. With multiple iterations, the node characteristics will be gradually enriched, which can better express the relationship between nodes in the topology graph. Information transfer function: ,in It represents the aggregation of the information received by node v at layer k to the adjacent nodes. represents the aggregation of the neighboring nodes of node v, refers to the message passing function, and They refer to the topographic and meteorological characteristics of nodes u and v at the k-1 layer, respectively. It refers to the characteristics of the edge (u,v).

[0171] Node update: The node update mechanism helps to fuse node information to a deeper level, allowing the model to make better predictions on unseen data. This update mechanism can help the model identify more complex patterns and trends, thereby improving the generalization ability of the model. The node update function is as follows: ,in, refers to the feature representation of node v at layer k, is the update function to calculate the representation of the node at the kth layer, It refers to the information received by node v at layer k.

[0172] Preferably, the obtained prediction model is verified and optimized, including but not limited to the following verification optimization strategies: improving the prediction accuracy of the prediction model through the loss function, such as selecting the NSE function as the loss function to measure the prediction error; improving the prediction accuracy of the prediction model through the regularization strategy, regularization refers to reducing the impact of overfitting on the model accuracy by adding penalty terms; improving the accuracy and generalization performance of the prediction model through hyperparameter tuning, hyperparameter tuning refers to presetting parameters and then finding the best parameter combination through grid search, random search or cross-validation.

[0173] The result output module includes an integration submodule, a prediction judgment submodule, a data expansion submodule, a sample expansion submodule and a circulation submodule.

[0174] The prediction and judgment submodule is used to input the rainfall meteorological boundary conditions to be predicted into the integration submodule and drive the integration submodule. The integration submodule outputs the first predicted water level spatiotemporal distribution data corresponding to the target flood evolution mechanism model. The integration submodule outputs the second predicted water level spatiotemporal distribution data corresponding to the prediction model, and then judges whether the second predicted water level spatiotemporal distribution data is normal based on the first predicted water level spatiotemporal distribution data. If normal, the second predicted water level spatiotemporal distribution data is output as the prediction result; if abnormal, the command to expand the rainfall meteorological boundary conditions is transmitted to the sample expansion submodule and the sample expansion submodule is driven.

[0175] The second predicted water level spatiotemporal distribution data is compared with the first predicted water level spatiotemporal distribution data to calculate the difference between the two, and a threshold is set. If the difference does not exceed the threshold, the second predicted water level spatiotemporal distribution data is judged to be normal; if the difference exceeds the threshold, the second predicted water level spatiotemporal distribution data is judged to be abnormal. The preferred threshold is 30%.

[0176] The mean percentage error uses the mean absolute percentage error to calculate the difference value, and the calculation formula is as follows: , where n is the sample size, Refers to the spatiotemporal distribution data of the second predicted water level of the ith i Refers to the spatiotemporal distribution data of the first predicted water level of the ith time. In actual use, users can choose different mathematical formulas to calculate the difference value according to actual needs and actual conditions, such as using absolute error, mean square error, logarithmic error, etc. to calculate the difference value.

[0177] Preferably, in actual use, according to user needs, the relevant data of flood evolution can be counted and output based on the second predicted water level spatiotemporal distribution data, including but not limited to flood depth, flood diffusion range, flooding time, etc. Figure 7 As shown, the output is the time series data of flood depth and inundation time, that is, a schematic diagram of the predicted spatiotemporal distribution of surface flood depth. The time series data can be used to draw a spatiotemporal distribution diagram of the water level, thereby obtaining the flood diffusion range. Figure 6 Schematic diagram of the temporal and spatial distribution of actual surface water depth. Figure 8 A comparison chart of actual surface water depth and predicted surface flood depth.

[0178] The sample expansion submodule expands the data samples under the rainfall meteorological boundary condition, adds the expanded data samples to the sample set of the sample set module, and drives the circulation submodule.

[0179] The rainfall meteorological boundary conditions to be predicted are expanded by random disturbance, and the expanded rainfall meteorological boundary conditions are input into the target flood evolution mechanism model to obtain the water level spatiotemporal distribution data corresponding to the expanded rainfall meteorological boundary conditions. The expanded rainfall meteorological boundary conditions and the corresponding expanded spatiotemporal distribution data of water level constitute a data sample, and the data sample is added to the sample set.

[0180] There are three ways to randomly perturb the meteorological boundary conditions of the rainfall to be predicted: 1. Perturbation of rainfall time

[0181] The temporal characteristics of rainfall events (e.g., start time, duration) may affect hydrological processes, especially in flood simulations. To capture the spatiotemporal variability of rainfall, the rainfall timing can be perturbed in the following ways:

[0182] Disturbance rainfall starts at:

[0183] By randomly perturbing the start time of rainfall, we can simulate different start times of rainfall events. Specifically, we can randomly select the start time of rainfall within a fixed time window. For example, if the start time of rainfall is T start , can be obtained by adding a random perturbation value ∆T start , so that the starting time after the disturbance is T start +∆T start .

[0184] Random perturbations can be implemented by sampling from a uniform or normal distribution, for example:

[0185] ∆T start ∈(−ΔT max ,ΔT max ), where ΔT max It is the maximum time range of the disturbance, which can usually be set based on historical data or expert experience.

[0186] Disturbance rainfall duration:

[0187] The duration of the rainfall event (i.e. how long it rains) is also a key factor. The duration of the rainfall can be randomly perturbed to simulate scenarios with different rainfall persistence.

[0188] If the duration of a rainfall event is T duration , then through the random perturbation ∆T duration To get a new duration: T duration =T duration +ΔT duration Where, ∆T duration You can sample from a distribution, such as a normal distribution or a uniform distribution.

[0189] Disturbance rainfall interval:

[0190] The intervals between multiple rainfall events are disturbed to simulate different rainfall intervals. The perturbation method is similar, using random distribution to generate different intervals.

[0191] 2. Disturbance of rainfall intensity

[0192] Rainfall intensity is an important factor affecting basin runoff and hydrological processes. Disturbing rainfall intensity helps simulate the impact of changes in rainfall intensity on hydrological processes. The following methods can be used for disturbance:

[0193] Spatial distribution of disturbance rainfall intensity:

[0194] Rainfall intensity often has spatial variability, and the spatial distribution of rainfall intensity can be perturbed. For example, suppose an existing rainfall intensity map (based on meteorological observations or historical data) can be used to generate a new rainfall intensity field by adding noise.

[0195] The rainfall intensity is disturbed in the following way:

[0196] I perturbed =I original +ε

[0197] Among them, I perturbed Refers to the rainfall intensity after the disturbance, I original refers to the original rainfall intensity, ε is random noise that can be sampled from a normal distribution or a uniform distribution, and represents the range of rainfall intensity.

[0198] The size (standard deviation) of the random noise can be adjusted according to the range of rainfall intensity to ensure that the rainfall intensity after disturbance remains within a reasonable range.

[0199] Temporal variation of disturbance rainfall intensity:

[0200] The rainfall intensity usually varies over time, and different rainfall patterns can be simulated by perturbing the rainfall intensity over time. For example, if the rainfall intensity increases or decreases linearly over a period of time, you can add random perturbations to the trend:

[0201] I(t2)=I0(t2)+ΔI(t2)

[0202] Where t2 refers to time, I(t2) is the intensity after perturbation, I0(t2) is the original intensity, and ΔI(t2) is the perturbation of the intensity change over time, which can be generated by random distribution.

[0203] Extreme values ​​of disturbance rainfall intensity:

[0204] In rainfall events, extreme rainfall intensity (such as heavy rainfall in a short period of time) has a greater impact on floods. Extreme rainfall events can be simulated to disturb the rainfall intensity at certain time points to make it more extreme. This can be done by: I(t2)=I0(t2)+βI0(t2)

[0205] Among them, β is an amplification factor, which is usually greater than 1 and is used to simulate extreme weather conditions.

[0206] 3. Comprehensive perturbation strategy

[0207] In addition to perturbing rainfall time and rainfall intensity separately, you can also combine the two to perform a joint perturbation. This can help simulate more complex rainfall event scenarios:

[0208] Comprehensive perturbations: Simultaneously perturb the start time, duration, and intensity of rainfall to generate multiple different rainfall scenarios for simulation. For example, noise or random perturbations can be added to the start time and intensity of each rainfall event.

[0209] The loop submodule cyclically runs the training prediction module, the integration submodule, and the prediction judgment submodule until the prediction judgment submodule determines that the second prediction water level spatiotemporal distribution data under the rainfall meteorological boundary condition is normal.

[0210] Example 3

[0211] An electronic device disclosed in the present invention includes one or more processors, one or more memories and one or more programs. The programs are stored in the memories and are configured to be executed by the processors. When the programs are loaded into the processors, the steps of the flood evolution prediction method of Example 1 are implemented.

[0212] Example 4

[0213] The present invention discloses a computer-readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the flood evolution prediction method of embodiment 1.

Claims

1. A flood evolution prediction method, characterized by: The following steps are included: S1: Obtain basic data of the target area; S2: Construct an initial flood evolution mechanism model, define and verify the initial flood evolution mechanism model based on basic data, and obtain the target flood evolution mechanism model; S3: Extract historical rainfall data of the target area in the basic data, and calculate the rainfall intensity-duration-rainfall frequency relationship in the target area based on the historical rainfall data; design the rainfall frequency and duration, and calculate the corresponding rainfall intensity based on the rainfall intensity-duration-rainfall frequency relationship; define the time series data consisting of the duration and the corresponding rainfall intensity as the rainfall meteorological boundary conditions; S4: input the constructed rainfall meteorological boundary conditions into the target flood evolution mechanism model, and the target flood evolution mechanism model outputs the water level spatiotemporal distribution data corresponding to the rainfall meteorological boundary conditions. Multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data constitute a sample set; S5: Build a graph neural network model and use the sample set to train the graph neural network model to obtain a prediction model for predicting flood evolution; S6: The user inputs the rainfall meteorological boundary conditions to be predicted into the target flood evolution mechanism model and the prediction model, the target flood evolution mechanism model outputs the first predicted water level spatiotemporal distribution data, and the prediction model outputs the second predicted water level spatiotemporal distribution data; S7: Based on the first predicted water level spatiotemporal distribution data, determine whether the second predicted water level spatiotemporal distribution data is normal. If normal, output the second predicted water level spatiotemporal distribution data as the prediction result; if abnormal, expand the data samples under the rainfall meteorological boundary conditions, and add the expanded data samples to the sample set, and repeat steps S4-S7 until the second predicted water level spatiotemporal distribution data under the rainfall meteorological boundary conditions to be predicted is normal.

2. The flood evolution prediction method according to claim 1, characterized in that: The basic data described in step S1 include terrain elevation data, research scope, river data, underlying surface attribute data, meteorological data and calibration data; the meteorological data include rainfall time series data, temperature time series data, relative humidity time series data, wind speed time series data, radiation time series data, atmospheric pressure time series data and snowmelt time series data.

3. The flood evolution prediction method according to claim 1, characterized in that: The initial flood evolution mechanism model in step S2 is Intelliway-SSIM; The steps to define and verify the initial flood evolution mechanism model are as follows: Based on the terrain elevation data, research scope and river data, the research scope of the target basin is divided into grids of different sizes in Intelliway-SSIM, where the land area is a triangular grid and the river channel is a non-triangular grid; Assign corresponding attributes to the triangular mesh based on the terrain elevation data and the underlying surface attribute data, and assign corresponding attributes to the non-triangular mesh based on the river data and the underlying surface attribute data; Input meteorological data into Intelliway-SSIM and drive Intelliway-SSIM. Intelliway-SSIM outputs simulation values ​​and compares the simulation values ​​with the calibration data. If the error between the two exceeds the preset value, adjust the key parameters of the Intelliway-SSIM model or adjust the grid division of Intelliway-SSIM until the error is less than or equal to the preset value.

4. The flood evolution prediction method according to claim 1, characterized in that: The calculation steps of the rainfall intensity-duration-rainfall frequency relationship are as follows: Design rainfall frequency; Fitting historical rainfall data to obtain rainfall frequencies of different rainfall intensities and durations; By performing regression analysis on the rainfall frequency data of different rainfall intensities and durations, we obtain the relationship between rainfall intensity-duration-rainfall frequency: ， ； Where I refers to rainfall intensity, in mm / h; D refers to rainfall duration, in h; R refers to recurrence period, in year; f refers to rainfall frequency, and a, b, and c are constants.

5. The flood evolution prediction method according to claim 1, characterized in that: The steps for judging whether the second predicted water level spatiotemporal distribution data is normal in step S7 are as follows: Calculate the difference between the second predicted water level spatiotemporal distribution data and the first predicted water level spatiotemporal distribution data; A threshold is set. If the difference value does not exceed the threshold, the second predicted water level spatiotemporal distribution data is judged to be normal; if the difference value exceeds the threshold, the second predicted water level spatiotemporal distribution data is judged to be abnormal.

6. The flood evolution prediction method according to claim 1, characterized in that: The steps of expanding the data sample in step S7 are as follows: The meteorological boundary conditions of the rainfall to be predicted are expanded by random disturbance; the meteorological boundary conditions of the rainfall to be predicted are expanded by random disturbance in three ways: disturbance rainfall time, disturbance rainfall intensity and comprehensive disturbance rainfall time and rainfall intensity; The expanded rainfall meteorological boundary conditions are input into the target flood evolution mechanism model to obtain the water level temporal and spatial distribution data corresponding to the expanded rainfall meteorological boundary conditions; The expanded rainfall meteorological boundary conditions and the corresponding expanded water level spatiotemporal distribution data constitute the expanded data samples.

7. A flood evolution prediction system, characterized by: include, Data acquisition module, to obtain basic data of the target area; The mechanism model module is used to construct an initial flood evolution mechanism model, and define and verify the initial flood evolution mechanism model based on the basic data of the data acquisition module to obtain the target flood evolution mechanism model; The data construction module is used to extract the historical rainfall data of the target area in the basic data, and calculate the rainfall intensity-duration-rainfall frequency relationship in the target area based on the historical rainfall data; design the rainfall frequency and duration, and calculate the corresponding rainfall intensity based on the rainfall intensity-duration-rainfall frequency relationship; define the time series data composed of the duration and the corresponding rainfall intensity as the rainfall meteorological boundary conditions; The sample set module can input the constructed rainfall meteorological boundary conditions into the target flood evolution mechanism model, and the target flood evolution mechanism model outputs the water level spatiotemporal distribution data corresponding to the rainfall meteorological boundary conditions. Multiple sets of rainfall meteorological boundary conditions and their corresponding water level spatiotemporal distribution data constitute the sample set; Train the prediction module, build a graph neural network model, and call the sample set to train the graph neural network model to obtain a prediction model for predicting flood evolution; The result output module includes an integration submodule, a prediction judgment submodule, a data expansion submodule, a sample expansion submodule and a circulation submodule; Integration submodule, used to integrate the target flood evolution mechanism model and prediction model; The prediction and judgment submodule is used to input the rainfall meteorological boundary conditions to be predicted into the integration submodule and drive the integration submodule, the integration submodule outputs the first predicted water level spatiotemporal distribution data corresponding to the target flood evolution mechanism model, the integration submodule outputs the second predicted water level spatiotemporal distribution data corresponding to the prediction model, and then judges whether the second predicted water level spatiotemporal distribution data is normal based on the first predicted water level spatiotemporal distribution data. If normal, the second predicted water level spatiotemporal distribution data is output as the prediction result; if not normal, the command to expand the rainfall meteorological boundary conditions is transmitted to the sample expansion submodule and the sample expansion submodule is driven; The sample expansion submodule expands the data samples under the rainfall meteorological boundary conditions, adds the expanded data samples to the sample set of the sample set module, and drives the loop submodule; The loop submodule cyclically runs the training prediction module, the integration submodule, and the prediction judgment submodule until the prediction judgment submodule determines that the second prediction water level spatiotemporal distribution data under the rainfall meteorological boundary condition is normal.

8. An electronic device, characterized in that: The method comprises one or more processors, one or more memories and one or more programs, wherein the programs are stored in the memories and are configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the flood evolution prediction method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the flood evolution prediction method according to any one of claims 1 to 6.

Images