Urban inland inundation situation deduction method based on fusion of mechanism knowledge and deep learning

Abstract

The invention discloses an urban inland inundation situation deduction method based on fusion of mechanism knowledge and deep learning, and the method comprises the steps: collecting data, such as weather, terrain and historical disaster situations, and constructing a standardized database in combination with an inland inundation production and convergence mechanism; constructing an urban inland inundation numerical twinborn model, and optimizing hyper-parameters to improve prediction precision by adopting finite element calculation and a multi-scale coupling method based on hydrological and hydrodynamic models; constructing a causal association network by using the numerical twinborn data, and identifying a core affected area; designing a situation deduction algorithm of a space-time diagram neural network, capturing rainfall waterlogging dynamic characteristics, and realizing advanced early warning and accurate deduction; based on the waterlogging intensity and the exposure and vulnerability of the disaster-bearing body, a risk assessment system is constructed, the disaster diffusion trend and the disaster risk are analyzed, and support is provided for flood prevention decision and emergency management.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent information processing and relates to a method for deducing urban waterlogging situations based on the integration of mechanism knowledge and deep learning. Background Art

[0002] With the acceleration of urbanization and the frequent occurrence of extreme climate events, urban waterlogging has gradually become a major challenge for urban management. Existing waterlogging monitoring and prevention methods still have many deficiencies and cannot meet the needs of modern urban disaster prevention and mitigation. Traditional monitoring methods mainly rely on manual inspections and vehicle inspections, which have obvious lags. The untimely discovery of waterlogging points leads to low efficiency of early warning and disaster relief. In addition, although fixed-point monitoring based on sensors can provide local waterlogging information, the monitoring range is limited and the number of equipment and deployment density are high, resulting in huge construction and maintenance costs. At the same time, the existing waterlogging early warning mechanism is difficult to quickly and comprehensively analyze and count urban disasters, and fails to achieve real-time updating and refined transmission of waterlogging information, further increasing the passive risk of the public in disasters. Therefore, at this stage, there is an urgent need for an urban waterlogging monitoring and prevention technology that can perceive in real time, accurately predict and efficiently warn, so as to improve disaster response capabilities and reduce the damage to life and property caused by waterlogging.

[0003] With the in-depth advancement of smart city construction, the development of video sensors, Internet of Things technology and cloud computing platforms has provided new opportunities for urban waterlogging monitoring and early warning. In particular, the introduction of artificial intelligence technology has laid the foundation for the construction of a digital waterlogging monitoring platform that combines real-time perception with situation deduction. This study proposes a method for urban waterlogging situation deduction based on the fusion of mechanism knowledge and deep learning. By integrating multi-source data and mechanism models, a numerical twin model integrating meteorology, hydrology and hydrodynamics is constructed, and the prediction accuracy is improved by multi-scale coupling and finite element calculation. At the same time, the causal association inference technology is used to construct a complex network model, deeply analyze the dynamic relationship between urban waterlogging points, and combine the spatiotemporal graph neural network to realize the refined deduction and advance warning of waterlogging situation. This method can dynamically update disaster information, accurately analyze the occurrence and expansion process of waterlogging, and provide strong technical support for urban flood control decision-making and disaster emergency management. Summary of the invention

[0004] In order to overcome the above-mentioned defects, the present invention proposes a method for deducing urban waterlogging situation based on the fusion of mechanism knowledge and deep learning. The specific steps of the present invention are as follows:

[0005] S1, Multi-source data fusion and waterlogging mechanism analysis, by obtaining multi-source urban waterlogging related data, including meteorological data (rainfall intensity, rainfall distribution), hydrological data (topography, runoff characteristics) and disaster data (historical waterlogging points, disaster-affected areas, etc.), combined with the knowledge of waterlogging runoff, analyze the main controlling factors of rainfall to waterlogging and their sensitivity to waterlogging disasters. Based on the above analysis, a standard database of waterlogging process covering multi-dimensional factors is constructed to provide data support for subsequent model construction and situation deduction;

[0006] S2, construction of numerical twin model of urban waterlogging. Aiming at the complex dynamic characteristics of urban waterlogging, meteorological model, hydrological model and hydrodynamic model are constructed respectively. The multi-angle coupling of different models is realized through finite element calculation technology and multi-scale numerical calculation method, and the numerical twin model of urban rainfall waterlogging is constructed. The model hyperparameters are efficiently determined through historical waterlogging data and fast fitting technology, and the prediction accuracy and reliability of the model are verified by actual disaster data to ensure that the numerical twin model can accurately simulate the dynamic process of urban waterlogging.

[0007] S3, based on the construction of a complex network model of flood-prone areas based on causal association inference, using the waterlogging runoff data generated by the numerical twin model, combined with local meteorological characteristics, to build a standardized waterlogging process database, through multi-source time series data analysis, design a multi-cascade causal relationship inference method, explore the causal relationship between urban waterlogging points, vulnerable points and difficult waterlogging points, convert discrete spatiotemporal signals into a complex association network with time and space information, identify the core waterlogging disaster areas and priority monitoring areas based on the network topology structure, and provide structured support for the dynamic deduction of waterlogging situation;

[0008] S4, Design of waterlogging situation deduction algorithm based on spatiotemporal graph neural network. Taking the complex network of flood-prone areas as the research object, the waterlogging situation deduction algorithm based on spatiotemporal graph neural network (ST-GNN) is designed. The deep learning model is used to capture the spatiotemporal dynamic characteristics in a refined manner. The algorithm can provide advance warning of the precipitation and waterlogging process in flood-prone areas, and accurately deduce the waterlogging situation in the affected areas, providing highly timely dynamic analysis of waterlogging.

[0009] S5, urban waterlogging risk assessment and comprehensive early warning, based on the integrated assessment framework of waterlogging intensity, exposure and vulnerability of disaster-bearing bodies, designs a risk assessment method for urban waterlogging. Through comprehensive estimates of the degree of occurrence, spread trend and risks of waterlogging disasters in the affected areas, it achieves a further extension from single waterlogging degree prediction to overall waterlogging situation and risk assessment, providing scientific and comprehensive technical support for urban flood control decision-making and emergency management.

[0010] The technical solution features and improvements of the present invention are:

[0011] For step S2, the present invention constructs meteorological, hydrological and hydrodynamic numerical models based on the results of the main controlling factor analysis and in combination with multi-source urban waterlogging and environmental information data, respectively, establishes a numerical twin model of urban rainfall waterlogging through finite element calculation technology and multi-scale and multi-angle numerical calculation coupling method, and constructs multi-dimensional and multi-scale numerical simulation models of meteorology, hydrology and hydrodynamics based on the main controlling factor analysis of waterlogging disasters and urban rainfall and flood numerical simulation theory;

[0012] Based on the drainage pipe network system data, drainage metering data and river data in the urban area, a one-dimensional hydrodynamic model of the urban drainage system is constructed. The fluid movement law in the pipe network is complex, the pipe network structure and properties are different, the internal connections and branches are intricate, and the water flow movement of the entire urban drainage system is a series of extremely complex. The improved Saint-Venant equations and Preissmann narrow gap hypothesis are used as the hydrodynamic model framework of the one-dimensional drainage system to calculate the non-steady flow problems of the pipe network and river network (the remaining factors can be included in the framework according to the analysis results of the main control factors). At the same time, the local head loss and the water exchange between the pipe network and the surface are considered in the calculation process. The one-dimensional river channel and pipe network control equations are as follows:

[0013] Continuity equation:

[0014]

[0015] Momentum equation:

[0016]

[0017]

[0018] Among them, q r is the lateral inflow of the pipe network, that is, the lateral coupling exchange water between the river and the urban surface, q m is the amount of water exchanged between the pipe network and the ground surface through vertical coupling; A is the cross-sectional area of ​​water flow; S 0 , S f and S L are the bottom slope source term, along-the-way resistance loss and local resistance loss respectively; f b is the pressure caused by the change of the cross-sectional width of the river or pipe network; ξ is the cross-sectional width, b is the corresponding cross-sectional width, and h is the water depth;

[0019] According to the characteristics of the drainage network, the staggered grid and semi-implicit discretization method are used for calculation. The flow rate and cross-sectional shape of the pipe segment are defined at the center of the pipe segment, and the water head and buried pipe elevation are defined at the node of the pipe segment. The convection adopts the first-order upwind scheme to ensure the stability of the model and avoid numerical oscillation. Therefore, the continuity equation is discretized at the one-dimensional network node (rainwater outlet, drainage outlet), and the motion equation is discretized on the one-dimensional pipe segment (pipe network, river network);

[0020] The momentum equation is converted into i Discrete the solution at the position and use semi-implicit discretization to get:

[0021]

[0022] Where Δx i For pipe segment E i length, Δt is the time step; Q i For pipe segment E i Traffic at different times; Q m , Q n For node N m , N n The flow rate, ν m , ν n For node N m 、N n Flow velocity at, h l 、h m For node N l 、N m The water level at S 0i , S fi , S Li They are the bottom slope source resistance loss, along-the-way resistance loss, and local resistance loss. For the flow at the node, the first-order upwind interpolation format is used to increase the stability of the discrete:

[0023]

[0024] Arranging the above formulas yields:

[0025]

[0026] Among them, the coefficient a i = -θ·max{v l ,0}, c i = -θ·max{v m , 0}, F is the sum of the pressures caused by various loss values ​​and changes in the pipe area, θ is the implicit coefficient, and the discrete format is stable when θ>0.5; the continuity equation is discretized at the node Nm to obtain:

[0027]

[0028] Among them, Am is the area of ​​the drainage outlet or rainwater well connected to the pipe network, Qmh is the flow rate at the artificial outlet or drainage outlet, and Q j =(Q m +Q n ) / 2, Q i =(Q l +Q m) / 2 is the flow rate of pipe segment Ej, and the flow rate of Ei can be calculated through the flow rate at the node using the central interpolation method;

[0029] When a rainstorm and waterlogging disaster occurs in a city, the surface water will be discharged from the area through the urban drainage system at the beginning of the rainfall. However, as time goes by and the rainwater accumulates, when the rainfall far exceeds the maximum drainage capacity of the drainage system, the rainwater will overflow the drainage system, gather on the urban surface and produce surface runoff. The flow of surface water belongs to the category of two-dimensional hydrodynamics, and the water flow movement is three-dimensional. However, for large-scale simulation of urban rainstorm and waterlogging, the vertical scale is much smaller than the plane scale. Therefore, it is considered to use the plane two-dimensional shallow water equation to establish the model. At the same time, the control equation is assumed and simplified as follows: it is assumed that the pressure along the water depth direction satisfies the static pressure distribution, and the Coriolis force, wind stress and turbulence terms are ignored. In order to meet the rainfall infiltration conditions of urban stormwater problems, the rainfall term and the infiltration term are added to the source term of the continuity equation. At the same time, the water flow exchange term between the surface and the underground pipe network is also considered;

[0030] Continuity equation:

[0031]

[0032] The momentum equations in the x and y directions are:

[0033]

[0034] Where, h represents water depth; u and v represent flow velocities in x and y directions, respectively; q represents inflow; qj represents the flow exchange with the one-dimensional network; Sx and Sy represent source terms in x and y directions, respectively, including bottom slope source term S0, bottom friction source term Sf, and other source terms Se; n is roughness; z is elevation, α is runoff coefficient, re represents rainfall intensity, f represents regional infiltration, and q ex =Q ex / A ex Represents the flow rate exchanged between the surface per unit area and the one-dimensional pipe network, Q ex is the total amount of water exchanged between the surface and the drainage network, A ex is the area of ​​the drain or wellhead. The discretization method and idea of ​​the two-dimensional surface runoff model are similar to those of the one-dimensional mathematical model and will not be described in detail.

[0035] The selection of the flow generation model method for urban rainstorm waterlogging will affect the size of the surface runoff, so the selection of the flow generation model has a great influence on the accuracy of the model. The deformed unsaturated-saturated Darcy formula and Darcy's law are used, combined with water balance and physical constitutive relations, to apply to the surface flow generation and underground seepage process of the city. The permeability of urban land has a great influence on the flow generation model. Therefore, in the calculation and simulation process, the permeability of urban land is considered separately, and the balance equation of surface runoff and underground seepage is obtained by combining the law of conservation of matter, the principle of water balance and hydrological methods. At the same time, considering that underground pipe networks are laid in urban underground spaces, most urban rainfall flows to the outside of the domain through drainage pipe networks, and it is estimated that the drainage volume of underground pipe networks will be included in the calculation;

[0036] Impervious Area:

[0037] Q j，l =PEHR (10)

[0038] Permeable area:

[0039] Q j，s =PEHqR (11)

[0040] Among them, Q is surface runoff, P is rainfall, E is evaporation, H is plant interception, q is water infiltration in the permeable area, and R is the drainage volume of the underground drainage network;

[0041] The runoff generation mechanism in urban areas is mostly over-seepage runoff. In urban rainfall, as rainfall continues, the amount of surface water continues to increase, and the soil gradually changes from unsaturated to saturated state. The infiltration rate of the soil also increases with the increase of saturation. This belongs to the range of rainfall infiltration. An important difference between saturated soil and unsaturated soil is reflected in the difference in the permeability coefficient of the soil layer. In saturated soil, the permeability coefficient is a constant with a higher value, while the permeability coefficient of unsaturated soil is a function of the saturation of the soil. According to Darcy's seepage law and the continuity equation of soil water flow, the two-dimensional saturated-unsaturated soil term seepage control equation is obtained:

[0042]

[0043] The coupling between the river channel and the surface is mainly manifested as water interaction in the horizontal direction, which can be specifically divided into two forms: forward coupling and lateral coupling. Forward coupling describes the direct impact of river flow on surface water bodies, while lateral coupling reflects the water exchange between the river channel and the surface area. To simulate this coupling relationship, commonly used methods include the water balance method, the weir flow formula method, and the dynamic interchange calculation method based on boundary conditions. Through these methods, the water flow interaction process between the river channel and the surface can be accurately simulated, and a solid calculation support can be provided for the dynamic coupling of the waterlogging model; the coupling between the surface and the underground drainage network is reflected in the dynamic process of vertical water exchange. When urban waterlogging occurs, ground water usually flows into the underground drainage system through drainage wells or rainwater inlets. However, when the rainfall intensity exceeds the designed drainage capacity of the drainage network, the rainwater in the network may overflow, resulting in flooding on the surface. The coupling calculation of the surface and the network uses the weir flow formula. According to the relationship between the node water level of the drainage network at the current time step and the corresponding surface water level, it is determined whether the water flow is a node overflow or a node inflow. In the case of node inflow, it is further subdivided into two modes: free inflow and flooded inflow. The flow exchange is quantitatively calculated through the weir flow or orifice outflow formula. This refined processing can dynamically reflect the complex interactive behavior of the surface and the drainage network during rainfall waterlogging;

[0044] Aiming at the coupling relationship between underground drainage pipe network and river channel, a one-dimensional pipe network river channel open and full flow alternating model is proposed. This model can handle the water volume exchange inside the pipe network and between the pipe network and the river channel. Since this model can well simulate the dynamic coupling relationship in practical applications, it can simplify the further processing of the coupling between the river channel and the pipe network without special requirements.

[0045] In the numerical simulation model of rainstorm waterlogging, parameter calibration is one of the key steps in the construction of numerical twin models. However, traditional parameter calibration mostly relies on manual operation, which has problems such as low efficiency, poor model performance, large variance and frequent updating. Therefore, an automatic history fitting method based on Ensemble Kalman Filter (EnKF) is introduced to quickly calibrate parameters. The model verification includes the following steps: using a ring pipe network to verify the one-dimensional model, using a triangular retaining structure dam break flood evolution example to verify the accuracy of the two-dimensional plane water flow simulation, and combining real historical waterlogging data to comprehensively verify the entire numerical twin model to ensure the effectiveness and reliability of the model under complex conditions.

[0046] For step S3, the present invention systematically studies the complexity of the mechanism of urban area waterlogging caused by rainfall. In practical applications, the time series data of a single waterlogging point usually includes multi-dimensional information such as rainfall intensity, water accumulation depth, drainage speed, and surface slope. These data are of great value for predicting the trend of urban waterlogging. However, the isolation of the data of a single waterlogging point limits its effectiveness in predicting the overall urban waterlogging situation. This is because there is a significant strong correlation between flood-prone areas. The horizontal correlation between flood points is a key factor in achieving accurate prediction of waterlogging trends. Based on the study of the characteristics of flood-prone areas, it is found that their waterlogging correlation meets the applicable scenarios of weighted complex network models. Therefore, the present invention takes the hierarchical Bayesian method of variational inference as the theoretical basis, designs a multi-source time series causal relationship inference method for the rainfall-waterlogging process, and mines the waterlogging correlation between urban flood-prone points; through this method, the area with a strong causal relationship between waterlogging and waterlogging is judged by the edge, and the edge weight is proportional to the correlation strength. The discrete multi-dimensional time domain signal in the flood-prone area is converted into a complex correlation network containing time and space information. At the same time, based on the complex network theory, the important nodes in the network structure are mined. These nodes have a key impact on the development of urban waterlogging trends and can be identified as the core area of ​​waterlogging, which requires priority monitoring, governance and prevention and control. The information mining based on the complex network provides an important decision-making basis for urban waterlogging prevention and control.

[0047] Node degree: the degree of node i is C i Represents the number of edges directly connected to the node in the network G, which is calculated as:

[0048]

[0049] Among them, A ij It is an element in the adjacency matrix, indicating whether nodes i and j are connected.

[0050] Betweenness centrality: Betweenness centrality measures the importance of a node in the shortest path. For each pair of nodes in a connected graph, the total number of shortest paths from node s to node t is denoted by σ st , where the number of shortest paths passing through node i is denoted as σ st (i), the betweenness centrality BC of node i i Defined as:

[0051]

[0052] Betweenness centrality reflects the role of a node as a "bridge" in the network. The higher its value, the stronger the node's control over information flow;

[0053] The PageRank algorithm estimates the importance of a node by calculating the number and quality of links between nodes in the network. Its basic assumption is that more important nodes are referenced by more other nodes. The PageRank value is calculated recursively using the following formula:

[0054]

[0055] Among them, PR(i) is the PageRank value of node i; L(i) is the set of all nodes pointing to node i; d(j) is the out-degree of node j; α is the random jump parameter, usually set to 0.85.

[0056] For step S4, the present invention uses an end-to-end spatiotemporal graph neural network to predict the future attributes of nodes in multivariate time series data. The spatiotemporal graph neural network framework can extract and integrate complex spatiotemporal dependencies by integrating graph neural networks and various time series prediction methods, and make accurate predictions on the future degree and trend changes of urban waterlogging points. Assuming that the spatiotemporal series data of urban waterlogging points X = {x t ∈R N×F |t=0,1,...,T}, where N is the number of flood-prone points in the entire urban area, and F is the multi-source time series information feature of the flood-prone points. For this type of data, the space-time graph can represent the relationship between different nodes within a certain space-time range. The space-time graph can be expressed as G t =(V, E t , A t ), where V is the set of flood-prone nodes, E t and A t They represent the association set and association matrix at time t respectively, and use the GraphSAGE algorithm and the GRU algorithm to process spatial and temporal information respectively;

[0057] GraphSAGE includes sampling and aggregation. First, it uses the connection information between nodes to sample neighbor nodes, and then continuously fuses the information of adjacent nodes through multi-layer aggregation functions, and uses the fused information to predict node attributes. The detailed steps are as follows:

[0058] Step-1 Neighborhood sampling: Randomly sample neighbor nodes. The number of neighbors sampled at each level is no more than S. k indivual;

[0059] Step-2 Generate the embedded attributes of the target node: first aggregate the time series features of the second-order neighbors to generate the embedded representation of the first-order neighbors, and then aggregate the first-order embedded representation to generate the representation vector of the target node. The aggregation function in this project uses the average aggregation method, that is, first average the representation vectors of the neighbors by dimension, and then perform a nonlinear transformation on the obtained average value;

[0060]

[0061] Step-3: Input the embedding representation vector of the target node into the fully connected network to obtain the predicted attribute value of the target node.

[0062] GraphSAGE is responsible for mining the complex spatial correlation information of the flood-prone point association network, and GRU is used to decouple and process the temporal information of different flood-prone points. The factor neural architecture is used to integrate the temporal and spatial computing processes. In the factor neural architecture, the spatial learning network and the temporal learning network modules are stacked in parallel or serially. The detailed process is as follows: Assume that f(A, X t ) represents the output of the spatial GraphSAGE algorithm at time step t, and then f(A, X t ) is fed into the GRU to obtain the hidden state at time step t:

[0063]

[0064] Finally, a multi-layer perceptron (MLP) was used to utilize the embedding vectors of flood-prone nodes output by the GraphSAGE+GRU model to issue an early warning of whether precipitation in flood-prone areas would cause flooding or to simulate the flood situation in disaster-stricken areas. The simulation parameters included but were not limited to: changes in the depth of flood points, changes in the amount of water accumulated at flood points, the area of ​​water accumulated at flood points, water infiltration, drainage speed, and other flood-sensitive indicators.

[0065] The urban waterlogging situation deduction method based on the fusion of mechanism knowledge and deep learning of the present invention realizes the accurate prediction and dynamic deduction of urban waterlogging situation through the coordinated application of multi-source data fusion, complex network modeling and deep learning algorithm. By integrating the multi-dimensional data in the rainfall-waterlogging process into a complex correlation network with spatiotemporal characteristics, and combining it with a dynamic deduction algorithm, an efficient and reliable urban waterlogging situation assessment and early warning system is constructed, providing scientific technical support for urban waterlogging prevention and control and emergency management. The method has the following advantages:

[0066] (1) The present invention constructs a multi-dimensional and multi-scale standardized waterlogging process database by integrating waterlogging runoff data, meteorological characteristics and drainage system information, providing comprehensive basic data support for dynamic deduction;

[0067] (2) The present invention adopts the hierarchical Bayesian method of variational inference and designs a causal relationship inference algorithm for the rainfall-to-waterlogging process. The lateral correlation between waterlogging points is quantified as the edge weight of the complex network, which significantly improves the accuracy of waterlogging relationship mining in flood-prone areas.

[0068] (3) The present invention is combined with a spatiotemporal graph neural network, which can dynamically capture the evolution of the urban waterlogging situation, achieve advanced warning of rainfall-induced waterlogging and fine-grained dynamic analysis of the core disaster-stricken areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flow chart of the urban waterlogging situation deduction method based on the integration of mechanism knowledge and deep learning in the present invention.

[0070] Figure 2 This is a flow chart of multi-source urban waterlogging disaster information acquisition, main controlling factors and sensitivity analysis in the present invention.

[0071] Figure 3 The flowchart of constructing the numerical twin model of urban waterlogging coupled with the mechanism knowledge in the present invention. DETAILED DESCRIPTION

[0072] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0073] A method for constructing a panoramic view of urban flooding based on panoramic stitching and spatial projection, such as Figure 1 As shown in FIG. 1 , a flowchart of constructing a panoramic view of flooding based on panoramic stitching and spatial projection of the present invention is shown. The method includes:

[0074] S1, obtain multi-source data such as meteorology (meteorological observations in the past 50 years, climate model predictions, short-term precipitation forecasts, etc.), hydrology (urban drainage pipe networks, drainage facilities, drainage engineering scheduling information, historical waterlogging monitoring data), geography (digital high-precision geographic information, land use, soil type), socio-economics (population density distribution, resident migration patterns, disaster resistance and bearing capacity), urban management (transportation facilities, medical facilities, building information) and satellite remote sensing; statistically analyze local meteorological laws and rainfall-induced waterlogging characteristics, compare with other urban waterlogging events at home and abroad to study the common and heterogeneous factors of local waterlogging, and systematically analyze the characteristics and causes of real historical waterlogging events in the city; obtain multi-source information data such as meteorology, hydrology, geography, geology and humanities to build a historical database of local real rainfall-induced waterlogging events. Use popular machine learning algorithms, such as random forest, principal component analysis (PCA) and Lasso regression, and take the degree of waterlogging as the target to analyze the main controlling factors and screen the importance of variables for multi-source data. Taking random forest as an example, assuming that there are N types of waterlogging event-related data that can be obtained and used, after normalization or scalarization of each type of signal, a waterlogging historical event feature database B = {b 1 , b 2 , ...b N, O}, where O is the flood disaster degree variable. A random forest model is constructed on set B to fit the flood disaster degree variable. The relative rank of the feature of the decision node in the tree is used to evaluate the relative importance of the feature relative to the target variable. By averaging the prediction ability of several trees in the random forest to reduce the estimation variance, the quantitative results of factor importance can be obtained. Combining the calculation results of various main control factor analysis methods, the main control factors of local urban rainstorm flooding are analyzed and screened out; assuming that there is a function form between the waterlogging factor set and the flood disaster degree variable: O = f(b 1 , b 2 , ...b N ), perform Sobol sampling on each main control factor to obtain the total variance D of the function f, and decompose it into several sub-variances to obtain:

[0075]

[0076] Where D i is the variance of the O value generated by the ith factor, D ij is the variance of the O value produced by the combined effect of the i-th and j-th factors, and so on. After normalizing the above formula, we can get:

[0077]

[0078] The following sensitivity indexes of the main controlling factors on the degree of waterlogging can be obtained:

[0079] First-order sensitivity index: Second-order sensitivity index: Global sensitivity index: D ～i Divide factor b i The influence of other parameters.

[0080] The Sobol sensitivity analysis method is used to analyze and calculate the main controlling factors of the city's historical waterlogging data, and to screen the main controlling factors that have a greater impact on changes in the severity of waterlogging. In addition to providing qualitative and quantitative directions for urban waterlogging prevention, resistance and control, it also provides theoretical support for the establishment of a numerical twin model of urban waterlogging.

[0081] S2, based on the results of the main controlling factor analysis, combined with multi-source urban waterlogging and environmental information data, respectively build meteorological, hydrological and hydrodynamic numerical models, and establish a numerical twin model of urban rainfall waterlogging through finite element calculation technology and multi-scale and multi-angle numerical calculation coupling method. Based on the main controlling factor analysis of waterlogging disasters and the urban rainfall flood numerical simulation theory, respectively build multi-dimensional and multi-scale numerical simulation models of meteorology, hydrology and hydrodynamics;

[0082] Based on the drainage pipe network system data, drainage metering data and river data in the urban area, a one-dimensional hydrodynamic model of the urban drainage system is constructed. The fluid movement law in the pipe network is complex, the pipe network structure and properties are different, the internal connections and branches are intricate, and the water flow movement of the entire urban drainage system is a series of extremely complex. The improved Saint-Venant equations and Preissmann narrow gap hypothesis are used as the hydrodynamic model framework of the one-dimensional drainage system to calculate the non-steady flow problems of the pipe network and river network (the remaining factors can be included in the framework according to the analysis results of the main control factors). At the same time, the local head loss and the water exchange between the pipe network and the surface are considered in the calculation process. The one-dimensional river channel and pipe network control equations are as follows:

[0083] Continuity equation:

[0084]

[0085] Momentum equation:

[0086]

[0087]

[0088] Among them, q r is the lateral inflow of the pipe network, that is, the lateral coupling exchange water between the river and the urban surface, q m is the amount of water exchanged between the pipe network and the ground surface through vertical coupling; A is the cross-sectional area of ​​water flow; S 0 , S f and S L are the bottom slope source term, along-the-way resistance loss and local resistance loss respectively; f b is the pressure caused by the change of the cross-sectional width of the river or pipe network; ξ is the cross-sectional width, b is the corresponding cross-sectional width, and h is the water depth;

[0089] According to the characteristics of the drainage network, the staggered grid and semi-implicit discretization method are used for calculation. The flow rate and cross-sectional shape of the pipe segment are defined at the center of the pipe segment, and the water head and buried pipe elevation are defined at the node of the pipe segment. The convection adopts the first-order upwind scheme to ensure the stability of the model and avoid numerical oscillation. Therefore, the continuity equation is discretized at the one-dimensional network node (rainwater outlet, drainage outlet), and the motion equation is discretized on the one-dimensional pipe segment (pipe network, river network);

[0090] The momentum equation is converted into i Discrete the solution at the position and use semi-implicit discretization to get:

[0091]

[0092] Where Δx i For pipe segment E i length, Δt is the time step; Qi For pipe segment E i Traffic at different times; Q m , Q n For node N m , N n The flow rate, ν m , ν n For node N m 、N n Flow velocity at, h l 、h m For node N l 、N m The water level at S 0i , S fi , S Li They are bottom slope source resistance loss, along-the-way resistance loss, and local resistance loss;

[0093] For the flow at the node, the first-order upwind interpolation format is used to increase the stability of the discretization:

[0094]

[0095] Arranging the above formulas yields:

[0096]

[0097] Among them, the coefficient a i = -θ·max{v l ,0}, c i = -θ·max{v m , 0}, F is the sum of pressures caused by various loss values ​​and changes in pipe area, θ is the implicit coefficient, and the discrete format is stable when θ>0.5;

[0098] Discretize the continuity equation at node Nm to obtain:

[0099]

[0100] Among them, Am is the area of ​​the drainage outlet or rainwater well connected to the pipe network, Qmh is the flow rate at the artificial outlet or drainage outlet, and Q j =(Q m +Q n ) / 2, Q i =(Q l +Q m ) / 2 is the flow rate of pipe segment Ej, and the flow rate of Ei can be calculated through the flow rate at the node using the central interpolation method;

[0101] When a rainstorm and waterlogging disaster occurs in a city, the surface water will be discharged from the area through the urban drainage system at the beginning of the rainfall. However, as time goes by and the rainwater accumulates, when the rainfall far exceeds the maximum drainage capacity of the drainage system, the rainwater will overflow the drainage system, gather on the urban surface and produce surface runoff. The flow of surface water belongs to the category of two-dimensional hydrodynamics, and the water flow movement is three-dimensional. However, for large-scale simulation of urban rainstorm and waterlogging, the vertical scale is much smaller than the plane scale. Therefore, it is considered to use the plane two-dimensional shallow water equation to establish the model. At the same time, the control equation is assumed and simplified as follows: it is assumed that the pressure along the water depth direction satisfies the static pressure distribution, and the Coriolis force, wind stress and turbulence terms are ignored. In order to meet the rainfall infiltration conditions of urban stormwater problems, the rainfall term and the infiltration term are added to the source term of the continuity equation. At the same time, the water flow exchange term between the surface and the underground pipe network is also considered;

[0102] Continuity equation:

[0103]

[0104] The momentum equations in the x and y directions are:

[0105]

[0106] Where, h represents water depth; u and v represent flow velocities in x and y directions, respectively; q represents inflow; qj represents the flow exchange with the one-dimensional network; Sx and Sy represent source terms in x and y directions, respectively, including bottom slope source term S0, bottom friction source term Sf, and other source terms Se; n is roughness; z is elevation, α is runoff coefficient, re represents rainfall intensity, f represents regional infiltration, and q ex =Q ex / A ex Represents the flow rate exchanged between the surface per unit area and the one-dimensional pipe network, Q ex is the total amount of water exchanged between the surface and the drainage network, A ex is the area of ​​the drain or wellhead. The discretization method and idea of ​​the two-dimensional surface runoff model are similar to those of the one-dimensional mathematical model and will not be described in detail.

[0107] The selection of the flow generation model method for urban rainstorm waterlogging will affect the size of the surface runoff, so the selection of the flow generation model has a great influence on the accuracy of the model. The deformed unsaturated-saturated Darcy formula and Darcy's law are used, combined with water balance and physical constitutive relations, to apply to the surface flow generation and underground seepage process of the city. The permeability of urban land has a great influence on the flow generation model. Therefore, in the calculation and simulation process, the permeability of urban land is considered separately, and the balance equation of surface runoff and underground seepage is obtained by combining the law of conservation of matter, the principle of water balance and hydrological methods. At the same time, considering that underground pipe networks are laid in urban underground spaces, most urban rainfall flows to the outside of the domain through drainage pipe networks, and it is estimated that the drainage volume of underground pipe networks will be included in the calculation;

[0108] Impervious Area:

[0109] Q j,l =PEHR (29)

[0110] Permeable area:

[0111] Q j，s =PEHqR (30)

[0112] Among them, Q is surface runoff, P is rainfall, E is evaporation, H is plant interception, q is water infiltration in the permeable area, and R is the drainage volume of the underground drainage network;

[0113] The runoff generation mechanism in urban areas is mostly over-seepage. In urban rainfall, as rainfall continues, the amount of surface water continues to increase, and the soil gradually changes from unsaturated to saturated state. The infiltration rate of the soil also increases with the increase of saturation. This belongs to the range of rainfall infiltration. An important difference between saturated soil and unsaturated soil is reflected in the difference in the permeability coefficient of the soil layer. In saturated soil, the permeability coefficient is a constant with a higher value, while the permeability coefficient of unsaturated soil is a function of the soil saturation. According to Darcy's seepage law and the continuity equation of soil water flow motion, the two-dimensional saturated-unsaturated soil seepage control equation is obtained:

[0114]

[0115] The coupling between the river channel and the surface is mainly manifested as the water interaction in the horizontal direction, which can be specifically divided into two forms: forward coupling and lateral coupling. Forward coupling describes the direct impact of river flow on surface water bodies, while lateral coupling reflects the water exchange between the river channel and the surface area. To simulate this coupling relationship, commonly used methods include water balance method, weir flow formula method and dynamic interchange calculation method based on boundary conditions. Through these methods, the water flow interaction process between the river channel and the surface can be accurately simulated, and solid calculation support can be provided for the dynamic coupling of the waterlogging model. The coupling between the surface and the underground drainage network is reflected in the vertical water volume. The dynamic process of exchange. When urban waterlogging occurs, ground water usually flows into the underground drainage system through drainage wells or rainwater inlets. However, when the rainfall intensity exceeds the designed drainage capacity of the drainage network, the rainwater in the network may overflow, resulting in flooding on the surface. The coupling calculation of the surface and the network uses the weir flow formula. According to the relationship between the node water level of the drainage network and the corresponding surface water level at the current time step, it is determined whether the water flow is node overflow or node inflow. In the case of node inflow, it is further subdivided into free inflow and submerged inflow. The flow exchange is quantitatively calculated through the weir flow or orifice outflow formula. This refined processing can dynamically reflect the complex interactive behavior of the surface and the drainage network during rainfall waterlogging;

[0116] Aiming at the coupling relationship between underground drainage pipe network and river channel, a one-dimensional pipe network-river channel open full flow alternating model is proposed. This model can effectively handle the water exchange inside the pipe network and between the pipe network and the river channel, and accurately simulate the dynamic coupling characteristics in practical applications. In the absence of special requirements, this model is sufficient to meet the actual needs and can simplify the further processing of the coupling between the river channel and the pipe network. In the construction of the numerical simulation model of rainstorm waterlogging, parameter calibration is an important basis for the numerical twin model. However, the traditional parameter calibration method mainly relies on manual operation, which has problems such as low efficiency, unstable model performance, large parameter variance and frequent adjustment. In order to solve these limitations, the present invention introduces an automatic history fitting method based on ensemble Kalman filtering to achieve fast and efficient parameter calibration. The model verification process includes the following steps: first, the one-dimensional model is verified for accuracy using a ring pipe network; second, the reliability of the two-dimensional plane water flow simulation is verified using a triangular water retaining building dam break flood evolution example; finally, the entire numerical twin model is comprehensively tested in combination with real historical waterlogging data.

[0117] S3, systematically studied the complexity of the mechanism of urban waterlogging caused by rainfall. In practical applications, the time series data of a single waterlogging point usually includes multi-dimensional information such as rainfall intensity, water depth, drainage speed, and surface slope. These data are of great value for predicting the trend of urban waterlogging. However, the isolation of data from a single waterlogging point limits its effectiveness in predicting the overall situation of urban waterlogging. This is because there is a significant strong correlation between flood-prone areas. The horizontal correlation between flood points is a key factor in achieving accurate prediction of waterlogging trends. Based on the study of the characteristics of flood-prone areas, it is found that their waterlogging correlation meets the applicable scenarios of weighted complex network models. Therefore, the present invention takes the hierarchical Bayesian method of variational inference as the theoretical basis, designs a multi-source time series causal relationship inference method for the rainfall-waterlogging process, and mines the waterlogging correlation between urban flood-prone points; through this method, the area with a strong causal relationship between waterlogging and waterlogging is judged by the edge, and the edge weight is proportional to the correlation strength. The discrete multi-dimensional time domain signal in the flood-prone area is converted into a complex correlation network containing time and space information. At the same time, based on the complex network theory, the important nodes in the network structure are mined. These nodes have a key impact on the development of urban waterlogging trends and can be identified as the core area of ​​waterlogging, which requires priority monitoring, governance and prevention and control. The information mining based on the complex network provides an important decision-making basis for urban waterlogging prevention and control.

[0118] Node degree: the degree of node i is C i Represents the number of edges directly connected to the node in the network G, which is calculated as:

[0119]

[0120] Among them, A ijIt is an element in the adjacency matrix, indicating whether nodes i and j are connected.

[0121] Betweenness centrality: Betweenness centrality measures the importance of a node in the shortest path. For each pair of nodes in a connected graph, the total number of shortest paths from node s to node t is denoted by σ st , where the number of shortest paths passing through node i is denoted as σ st (i), the betweenness centrality BC of node i i Defined as:

[0122]

[0123] Betweenness centrality reflects the role of a node as a "bridge" in the network. The higher its value, the stronger the node's control over information flow;

[0124] The PageRank algorithm estimates the importance of a node by calculating the number and quality of links between nodes in the network. Its basic assumption is that more important nodes are referenced by more other nodes. The PageRank value is calculated recursively using the following formula:

[0125]

[0126] Among them, PR(i) is the PageRank value of node i; L(i) is the set of all nodes pointing to node i; d(j) is the out-degree of node j; α is the random jump parameter, usually set to 0.85.

[0127] S4, an end-to-end spatiotemporal graph neural network is used to predict the future attributes of nodes in multivariate time series data. The spatiotemporal graph neural network framework can extract and integrate complex spatiotemporal dependencies by integrating graph neural networks and various time series prediction methods, and make accurate predictions on the future degree and trend changes of urban waterlogging points. Assuming that the spatiotemporal series data of urban waterlogging points X = {x t ∈R N×F |t=0,1,...,T}, where N is the number of flood-prone points in the entire urban area, and F is the multi-source time series information characteristics of flood-prone points. For this type of data, the space-time graph can represent the relationship between different nodes within a certain space-time range. The space-time graph can be expressed as G t =(V, E t , A t ), where V is the set of flood-prone nodes, E t and A t They represent the association set and association matrix at time t respectively, and use the GraphSAGE algorithm and the GRU algorithm to process spatial and temporal information respectively;

[0128] GraphSAGE includes sampling and aggregation. First, it uses the connection information between nodes to sample neighbor nodes, and then continuously fuses the information of adjacent nodes through multi-layer aggregation functions, and uses the fused information to predict node attributes. The detailed steps are as follows:

[0129] Step-1 Neighborhood sampling: Randomly sample neighbor nodes. The number of neighbors sampled at each level is no more than S. k indivual;

[0130] Step-2 Generate the embedded attributes of the target node: first aggregate the time series features of the second-order neighbors to generate the embedded representation of the first-order neighbors, and then aggregate the first-order embedded representation to generate the representation vector of the target node. The aggregation function in this project uses the average aggregation method, that is, first average the representation vectors of the neighbors by dimension, and then perform a nonlinear transformation on the obtained average value;

[0131]

[0132] Step-3: Input the embedding representation vector of the target node into the fully connected network to obtain the predicted attribute value of the target node.

[0133] GraphSAGE is responsible for mining the complex spatial correlation information of the flood-prone point association network, and GRU is used to decouple and process the temporal information of different flood-prone points. The factor neural architecture is used to integrate the temporal and spatial computing processes. In the factor neural architecture, the spatial learning network and the temporal learning network modules are stacked in parallel or serially. The detailed process is as follows: Assume that f(A, X t ) represents the output of the spatial GraphSAGE algorithm at time step t, and then f(A, X t ) is fed into the GRU to obtain the hidden state at time step t:

[0134]

[0135] Finally, a multi-layer perceptron (MLP) was used to utilize the embedding vectors of flood-prone nodes output by the GraphSAGE+GRU model to issue an early warning of whether precipitation in flood-prone areas would cause flooding or to simulate the flood situation in disaster-stricken areas. The simulation parameters included but were not limited to: changes in the depth of flood points, changes in the amount of water accumulated at flood points, the area of ​​water accumulated at flood points, water infiltration, drainage speed, and other flood-sensitive indicators.

[0136] S5. Whether urban waterlogging will cause disaster depends not only on the objective waterlogging intensity and situation, but also on the vulnerability and exposure of the disaster-bearing body. Therefore, the conceptual mathematical expression of waterlogging disaster risk is as follows:

[0137] R=f(H,V,E) (37)

[0138] Among them, R represents natural disaster risk; H represents disaster events; E represents the exposure of the disaster-prone body; V represents the vulnerability of the disaster-prone body. The disaster event H is the primary condition for the formation of risk. Urban waterlogging disaster events refer to the frequency, duration, scope and intensity of natural disaster events (rainstorm waterlogging) that pose a threat to the city in a certain urban area. It can be expressed by the following expression:

[0139] H=f(P,I,R,T) (38)

[0140] Among them, H is the disaster event, P is the probability of urban waterlogging, I is the intensity of the disaster event, R is the impact range of waterlogging, and T is the duration of waterlogging;

[0141] Exposure refers to the number and value of disaster-prone bodies such as social systems, economic systems, and natural ecosystems affected by heavy rain and waterlogging. Exposure is a necessary condition for the occurrence of urban waterlogging risks. Only when a disaster-prone body is submerged by waterlogging can it be damaged and cause economic losses or casualties. The expression of urban waterlogging disaster exposure is as follows:

[0142] E=f(M,V,T,R) (39)

[0143] Among them, E is the exposure of the hazard-bearing body, M is the number of hazard-bearing bodies affected by waterlogging, V is the value of the hazard-bearing body affected by waterlogging, T represents the duration of waterlogging, and R is the area affected by waterlogging;

[0144] The vulnerability of a disaster-prone body refers to the characteristics of the disaster-prone body in the risk area that it is easy to be destroyed and lacks the ability to restore to its original state when it suffers from waterlogging disasters. It generally includes three aspects: sensitivity, response ability, and resilience. Sensitivity refers to the attribute of whether the disaster-prone body itself is easily damaged by waterlogging; response ability refers to the ability and level of human disaster prevention and mitigation during the occurrence of waterlogging; resilience refers to the ability of the disaster-prone body to return to normal or higher levels after the occurrence of waterlogging disasters. The greater the vulnerability, the worse the area's ability to resist and recover from disasters;

[0145] Combining the disaster risk assessment system with the waterlogging situation simulation algorithm can not only deduce the development dynamics and trends of waterlogging in urban areas, but also understand the disaster situations that different flood-prone areas may face in the future, and transform from waterlogging prediction to disaster prediction.

[0146] In summary, the present invention proposes a method for deducing the urban waterlogging situation based on the fusion of mechanism knowledge and deep learning. By combining multi-source data fusion, numerical twin model construction, causal correlation analysis and spatiotemporal graph neural network technology, the accurate deduction and real-time prediction of the development trend of urban waterlogging are realized. This method constructs a numerical model of waterlogging covering multiple fields such as meteorology, hydrology and hydrodynamics, mines the correlation structure of waterlogged areas through complex network analysis, and uses deep learning technology to design advance warning and refined situation deduction algorithms, and at the same time establishes a three-in-one waterlogging risk assessment framework. The present invention can dynamically perceive the occurrence and spread of urban waterlogging, provide scientific technical support for the accurate prediction, real-time monitoring and emergency response of waterlogging disasters, and significantly improve the city's disaster prevention and mitigation capabilities and emergency management efficiency.

[0147] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for predicting urban waterlogging based on the fusion of mechanism knowledge and deep learning, its characteristics and The specific steps are as follows: S1, Multi-source data fusion and waterlogging mechanism analysis, by obtaining multi-source urban waterlogging related data, including meteorological data (rainfall intensity, rainfall distribution), hydrological data (topography, runoff characteristics) and disaster data (historical waterlogging points, disaster-affected areas, etc.), combined with the knowledge of waterlogging runoff, analyze the main controlling factors of rainfall to waterlogging and their sensitivity to waterlogging disasters. Based on the above analysis, a standard database of waterlogging process covering multi-dimensional factors is constructed to provide data support for subsequent model construction and situation deduction; S2, construction of numerical twin model of urban waterlogging. Aiming at the complex dynamic characteristics of urban waterlogging, meteorological model, hydrological model and hydrodynamic model are constructed respectively. The multi-angle coupling of different models is realized through finite element calculation technology and multi-scale numerical calculation method, and the numerical twin model of urban rainfall waterlogging is constructed. The model hyperparameters are efficiently determined through historical waterlogging data and fast fitting technology, and the prediction accuracy and reliability of the model are verified by actual disaster data to ensure that the numerical twin model can accurately simulate the dynamic process of urban waterlogging. S3, based on the construction of a complex network model of flood-prone areas based on causal association inference, using the waterlogging runoff data generated by the numerical twin model, combined with local meteorological characteristics, to build a standardized waterlogging process database, through multi-source time series data analysis, design a multi-cascade causal relationship inference method, explore the causal relationship between urban waterlogging points, vulnerable points and difficult waterlogging points, convert discrete spatiotemporal signals into a complex association network with time and space information, identify the core waterlogging disaster areas and priority monitoring areas based on the network topology structure, and provide structured support for the dynamic deduction of waterlogging situation; S4, Design of waterlogging situation deduction algorithm based on spatiotemporal graph neural network. Taking the complex network of flood-prone areas as the research object, the waterlogging situation deduction algorithm based on spatiotemporal graph neural network (ST-GNN) is designed. The deep learning model is used to capture the spatiotemporal dynamic characteristics in a refined manner. The algorithm can provide advance warning of the precipitation and waterlogging process in flood-prone areas, and accurately deduce the waterlogging situation in the affected areas, providing highly timely dynamic analysis of waterlogging. S5, urban waterlogging risk assessment and comprehensive early warning, based on the integrated assessment framework of waterlogging intensity, exposure and vulnerability of disaster-bearing bodies, designs a risk assessment method for urban waterlogging. Through comprehensive estimates of the degree of occurrence, spread trend and risks of waterlogging disasters in the affected areas, it achieves a further extension from single waterlogging degree prediction to overall waterlogging situation and risk assessment, providing scientific and comprehensive technical support for urban flood control decision-making and emergency management.

2. According to claim 1, a method for deducing urban waterlogging situation based on the integration of mechanism knowledge and deep learning is characterized in that: For step S2, the present invention, based on the analysis results of the main controlling factors, combines multi-source urban waterlogging and environmental information data, respectively constructs meteorological, hydrological and hydrodynamic numerical models, establishes a numerical twin model of urban rainfall waterlogging through finite element calculation technology and multi-scale multi-angle numerical calculation coupling method, and constructs multi-dimensional and multi-scale numerical simulation models of meteorology, hydrology and hydrodynamics based on the main controlling factor analysis of waterlogging disasters and urban rain and flood numerical simulation theory; based on the drainage pipe network system data, drainage metering data and urban river data in the region, constructs a one-dimensional urban drainage system hydrodynamic model. The fluid movement law in the pipe network is complex, the pipe network structure and attributes are different, the internal connections and branches are intricate, and the water flow movement of the entire city's drainage system is a series of extremely complex. The improved Saint-Venant equations and Preissmann narrow gap hypothesis are used as the hydrodynamic model framework of the one-dimensional drainage system to calculate the non-steady flow problems of the pipe network and river network (the remaining factors can be included in the framework according to the main controlling factor analysis results). At the same time, the local head loss and the water exchange between the pipe network and the surface are considered in the calculation process. The one-dimensional river channel and pipe network control equations are as follows, the continuity equation: Momentum equation: Among them, q r is the lateral inflow of the pipe network, that is, the lateral coupling exchange water between the river and the urban surface, q m is the amount of water exchanged between the pipe network and the ground surface through vertical coupling; A is the cross-sectional area of ​​water flow; S0, S f and S L are the bottom slope source term, along-the-way resistance loss and local resistance loss respectively; g b is the pressure caused by the change of the cross-sectional width of the river or pipe network; ξ is the cross-sectional width, b is the corresponding cross-sectional width, and h is the water depth; According to the characteristics of the drainage network, the staggered grid and semi-implicit discretization method are used for calculation. The flow rate and cross-sectional shape of the pipe segment are defined at the center of the pipe segment, and the water head and buried pipe elevation are defined at the node of the pipe segment. The convection adopts the first-order upwind scheme to ensure the stability of the model and avoid numerical oscillation. Therefore, the continuity equation is discretized at the node of the one-dimensional network segment (rainwater inlet, drainage outlet), and the motion equation is discretized on the one-dimensional pipe segment (pipe network, river network); the momentum equation is discretized and solved at the pipe segment Ei, and the semi-implicit discretization is used to obtain: Among them, Δxi is the length of pipe section Ei, Δt is the time step; Qi is the flow rate of pipe section Ei at different times; Qm, Qn are the flow rates of nodes Nm, Nn, νm, νn are the flow velocities at nodes Nm, Nn, hl, hm are the water level values ​​at nodes Nl, Nm; S0i, Sfi, SLi are the bottom slope source resistance loss, along-the-way resistance loss, and local resistance loss, respectively; the first-order upwind interpolation format is used for the flow at the node to increase the stability of the discrete: Arranging the above formulas yields: Among them, the coefficient a i = -θ·max{v l ,0}, c i = -θ·max{v m , 0}, F is the sum of the pressures caused by various loss values ​​and changes in the pipe area, θ is the implicit coefficient, and the discrete format is stable when θ>0.5; the continuity equation is discretized at the node Nm to obtain: Among them, Am is the area of ​​the drainage outlet or rainwater well connected to the pipe network, Qmh is the flow rate at the artificial outlet or drainage outlet, and Q j =(Q m +Q n ) / 2, Q i =(Q l +Q m ) / 2 is the flow of pipe section Ej, and Ei can be calculated by the flow at the node using the central interpolation method; when a rainstorm disaster occurs in a city, the surface water will be discharged from the area through the urban drainage system at the beginning of the rainfall, but as time goes by and the rainwater accumulates, when the rainfall far exceeds the maximum drainage capacity of the drainage system, the rainwater will overflow the drainage system, gather on the urban surface and produce surface runoff. The flow of surface water belongs to the category of two-dimensional hydrodynamics, and the water flow movement is three-dimensional. However, for large-scale urban rainstorm waterlogging simulation, the vertical scale is much smaller than the plane scale. Therefore, it is considered to use the plane two-dimensional shallow water equation to establish the model. At the same time, the control equation is assumed and simplified as follows: it is assumed that the pressure along the water depth direction satisfies the static pressure distribution, and the Coriolis force, wind stress and turbulence terms are ignored. In order to meet the rainfall infiltration conditions of urban rainstorm problems, the rainfall term and the infiltration term are added to the source term of the continuity equation, and the water flow exchange term between the surface and the underground pipe network is also considered; Continuity equation: The momentum equations in the x and y directions are: Where, h represents water depth; u and v represent flow velocities in x and y directions, respectively; q represents inflow; qj represents the flow exchange with the one-dimensional network; Sx and Sy represent source terms in x and y directions, respectively, including bottom slope source term S0, bottom friction source term Sf, and other source terms Se; n is roughness; z is elevation, α is runoff coefficient, re represents rainfall intensity, f represents regional infiltration, and q ex =Q ex / A ex Represents the flow rate exchanged between the surface per unit area and the one-dimensional pipe network, Q ex is the total amount of water exchanged between the surface and the drainage network, A ex is the area of ​​the drainage outlet or wellhead. The discrete method and idea of ​​the two-dimensional surface runoff model are similar to those of the one-dimensional mathematical model and will not be repeated here. The selection of the runoff model method for urban rainstorm waterlogging will affect the size of the surface runoff. Therefore, the selection of the runoff model has a great influence on the accuracy of the model. The deformed unsaturated-saturated Darcy formula and Darcy's law are used in combination with water balance and physical constitutive relations to apply to the urban surface runoff and underground seepage process. The permeability of urban land has a great influence on the runoff model. Therefore, in the calculation and simulation process, the permeability of urban land is considered separately. The balance equation of surface runoff and underground seepage is obtained by combining the law of conservation of matter, the principle of water balance and the hydrological method. At the same time, taking into account the laying of underground pipe networks in urban underground space, most of the urban rainfall flows to the outside of the domain through the drainage pipe network, and it is estimated that the drainage volume of the underground pipe network will be included in the calculation; impervious area: Q j，l =P-E-H-R (10) Permeable area: Q j，s =P-E-H-q-R (11) Among them, Q is the surface runoff, P is the rainfall, E is the evaporation, H is the plant interception, q is the infiltration water in the permeable area, and R is the drainage of the underground drainage network; the flow generation mechanism in urban areas is mostly over-infiltration flow generation. In urban rainfall, as the rainfall continues, the surface water volume continues to increase, and the soil gradually changes from unsaturated to saturated state. The infiltration rate of the soil also increases with the increase of saturation. This belongs to the range of rainfall infiltration. An important difference between saturated soil and unsaturated soil is reflected in the difference in the permeability coefficient of the soil layer. In saturated soil, the permeability coefficient is a constant with a higher value, while the permeability coefficient of unsaturated soil is a function of the soil saturation. According to Darcy's seepage law and the continuity equation of soil water flow motion, the two-dimensional saturated-unsaturated soil term seepage control equation is obtained: The coupling between the river channel and the surface is mainly manifested as water interaction in the horizontal direction, which can be specifically divided into two forms: forward coupling and lateral coupling. Forward coupling describes the direct impact of river flow on surface water bodies, while lateral coupling reflects the water exchange between the river channel and the surface area. To simulate this coupling relationship, commonly used methods include the water balance method, the weir flow formula method, and the dynamic interchange calculation method based on boundary conditions. Through these methods, the water flow interaction process between the river channel and the surface can be accurately simulated, and a solid calculation support can be provided for the dynamic coupling of the waterlogging model; the coupling between the surface and the underground drainage network is reflected in the dynamic process of vertical water exchange. When urban waterlogging occurs, ground water usually flows into the underground drainage system through drainage wells or rainwater inlets. However, when the rainfall intensity exceeds the designed drainage capacity of the drainage network, the rainwater in the network may overflow, resulting in flooding on the surface. The coupling calculation of the surface and the network uses the weir flow formula. According to the relationship between the node water level of the drainage network at the current time step and the corresponding surface water level, it is determined whether the water flow is a node overflow or a node inflow. In the case of node inflow, it is further subdivided into two modes: free inflow and flooded inflow. The flow exchange is quantitatively calculated through the weir flow or orifice outflow formula. This refined processing can dynamically reflect the complex interactive behavior of the surface and the drainage network during rainfall waterlogging; Aiming at the coupling relationship between underground drainage pipe network and river channel, a one-dimensional pipe network river channel open and full flow alternating model is proposed. This model can handle the water exchange inside the pipe network and between the pipe network and the river channel. Since the model can well simulate the dynamic coupling relationship in practical applications, the further processing of the coupling between the river channel and the pipe network can be simplified without special requirements. In the numerical simulation model of rainstorm waterlogging, parameter calibration is one of the key steps in the construction of numerical twin model. However, traditional parameter calibration mostly relies on manual operation, which has problems such as low efficiency, poor model performance, large variance and frequent updating. Therefore, an automatic history fitting method based on Ensemble Kalman Filter (EnKF) is introduced for rapid parameter calibration. The model verification includes the following steps: the one-dimensional model is verified by using a ring pipe network, the accuracy of the two-dimensional plane water flow simulation is verified by using a triangular retaining structure dam break flood evolution example, and the entire numerical twin model is fully verified in combination with real historical waterlogging data to ensure the effectiveness and reliability of the model under complex conditions.

3. According to claim 1, a method for deducing urban waterlogging situation based on the integration of mechanism knowledge and deep learning is characterized in that: For step S3, the present invention has systematically studied the complexity of the mechanism of urban area waterlogging caused by rainfall. In practical applications, the time series data of a single waterlogging point usually includes multi-dimensional information such as rainfall intensity, water accumulation depth, drainage speed, surface slope, etc. These data are of great value for predicting the trend of urban waterlogging. However, the isolation of the data of a single waterlogging point limits its effectiveness in predicting the overall urban waterlogging situation. This is because there is a significant strong correlation between flood-prone areas. The horizontal correlation between flood points is a key factor in achieving accurate prediction of waterlogging trends. Based on the study of the characteristics of flood-prone areas, it is found that their waterlogging correlation meets the applicable scenarios of weighted complex network models. Therefore, the present invention takes the hierarchical Bayesian method of variational inference as the theoretical basis, designs a multi-source time series causal relationship inference method for the rainfall-waterlogging process, and mines the waterlogging correlation between urban flood-prone points; through this method, the area with a strong causal relationship between waterlogging and waterlogging is judged by the edge, and the edge weight is proportional to the correlation strength. The discrete multi-dimensional time domain signal in the flood-prone area is converted into a complex correlation network containing time and space information. At the same time, based on the complex network theory, the important nodes in the network structure are mined. These nodes have a key impact on the development of urban waterlogging trends and can be identified as the core area of ​​waterlogging, which requires priority monitoring, governance and prevention and control. The information mining based on the complex network provides an important decision-making basis for urban waterlogging prevention and control. Node degree: The degree Ci of node i represents the number of edges directly connected to the node in the network G. The specific calculation is: Among them, Aij is an element in the adjacency matrix, indicating whether nodes i and j are connected. Betweenness centrality: Betweenness centrality measures the importance of a node in the shortest path. For each pair of nodes in a connected graph, the total number of shortest paths from node s to node t is denoted by σ st , where the number of shortest paths passing through node i is denoted as σ st (i), the betweenness centrality BC of node i i Defined as: Betweenness centrality reflects the role of a node as a "bridge" in the network. The higher its value, the stronger the node's control over information flow. The PageRank algorithm estimates the importance of a node by calculating the number and quality of links between nodes in the network. Its basic assumption is that more important nodes will be cited by more other nodes. The PageRank value is calculated recursively using the following formula: Among them, PR(i) is the PageRank value of node i; L(i) is the set of all nodes pointing to node i; d(j) is the out-degree of node j; α is the random jump parameter, usually set to 0.

85.

4. According to claim 1, a method for deducing urban waterlogging situation based on the integration of mechanism knowledge and deep learning is characterized in that: For step S5, the present invention uses an end-to-end spatiotemporal graph neural network to predict the future attributes of nodes in multivariate time series data. The spatiotemporal graph neural network framework can extract and integrate complex spatiotemporal dependencies by integrating graph neural networks and various time series prediction methods, and make accurate predictions on the future degree and trend changes of urban waterlogging points. Assuming that the spatiotemporal series data of urban waterlogging points X = {x t ∈R N×F |t=0,1,...,T}, where N is the number of flood-prone points in the entire urban area, and F is the multi-source time series information characteristics of flood-prone points. For this type of data, the space-time graph can represent the relationship between different nodes within a certain space-time range. The space-time graph can be expressed as G t =(V, E t , A t ), where V is the set of flood-prone nodes, E t and A t They represent the association set and association matrix at time t respectively. The GraphSAGE algorithm and the GRU algorithm are combined to process spatial and temporal information respectively. GraphSAGE includes sampling and aggregation. First, the connection information between nodes is used to sample neighbor nodes. Then, the information of adjacent nodes is continuously fused together through multi-layer aggregation functions. The fused information is used to predict node attributes. The detailed steps are as follows: Step-1 Neighborhood sampling: Randomly sample neighbor nodes, and the number of neighbors sampled at each level shall not exceed Sk; Step-2 Generate the embedded attributes of the target node: first aggregate the time series features of the second-order neighbors to generate the embedded representation of the first-order neighbors, and then aggregate the first-order embedded representation to generate the representation vector of the target node. The aggregation function in this project uses the average aggregation method, that is, first average the representation vectors of the neighbors by dimension, and then perform a nonlinear transformation on the obtained average value; Step-3: Input the embedding representation vector of the target node into the fully connected network to obtain the predicted attribute value of the target node. GraphSAGE is responsible for mining the complex spatial correlation information of the flood-prone point association network, and GRU is used to decouple and process the temporal information of different flood-prone points. The factor neural architecture is used to integrate the temporal and spatial computing processes. In the factor neural architecture, the spatial learning network and the temporal learning network modules are stacked in parallel or serially. The detailed process is as follows: Assume that f(A, X t ) represents the output of the spatial GraphSAGE algorithm at time step t, and then f(A, X t ) is fed into the GRU to obtain the hidden state at time step t: f(X, A) = σ(AXW0), u t =σ(W u [f(A,X t ),h t-1 ]+b u ), r t =σ(W r [f(A,X t ),h t-1 ]+b r ), c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c ), h t =u t *h t-1 +(1-u t )*c t (17) Finally, a multi-layer perceptron (MLP) was used to utilize the embedding vectors of flood-prone nodes output by the GraphSAGE+GRU model to issue an early warning of whether precipitation in flood-prone areas would cause flooding or to simulate the flood situation in disaster-stricken areas. The simulation parameters included but were not limited to: changes in the depth of flood points, changes in the amount of water accumulated at flood points, the area of ​​water accumulated at flood points, water infiltration, drainage speed, and other flood-sensitive indicators.

Images