Urban hydrological multi-physical field decoupling supervised AI flood real-time simulation method

Abstract

The application provides an AI flood real-time simulation method for urban hydrology multi-physical field decoupling supervision, constructs an intelligent physical prediction model for urban flood, separates a one-dimensional drainage pipe network submodule from a two-dimensional surface flood module by hierarchical decoupling of an urban hydrology and water power system, designs a multi-scale physical constraint loss function for different links, realizes multi-scale physical constraint fusion at a microscopic, mesoscopic and macroscopic level, embeds a residual constraint of Saint-Venant equation in a neural network agent of a drainage system at a microscopic level, embeds a residual constraint of a shallow water equation in a surface flood evolution system at a mesoscopic level, realizes water balance constraint of the whole system at a macroscopic level, and realizes multi-scale physical field supervision constraint by decoupling the urban hydrology and water power.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hydraulic engineering, and particularly relates to an AI real-time simulation method for urban hydrology multi-physical field decoupling supervision of flood. BACKGROUND

[0002] With the acceleration of global urbanization, the population and infrastructure of cities are highly concentrated, and the economic losses and social impacts caused by urban flood disasters are becoming increasingly serious. According to statistics, the annual economic losses caused by urban floods worldwide have increased several times in the past few decades, posing a great threat to the life and property safety of urban residents. This phenomenon is mainly attributed to the changes in underlying surface caused by urbanization and the increase in extreme rainfall events caused by climate change. During the process of urbanization, a large number of impervious surfaces appear, which change the natural hydrological cycle, making it difficult for rainwater to naturally penetrate, and the surface runoff rapidly increases; at the same time, climate change makes the intensity and frequency of heavy rain significantly increase, further exacerbating the risk of urban floods.

[0003] In the field of urban flood risk management, flood simulation models are an important tool to ensure urban safety. However, the existing single driving models (physically driven numerical models, data-driven surrogate models) still have many limitations and are difficult to meet the actual application requirements.

[0004] Physically driven numerical models are commonly used means for current urban flood simulation, such as multi-module structured models, which can comprehensively simulate urban hydrodynamics by coupling hydrological runoff modules, one-dimensional Navier-Stokes equation modules, and two-dimensional shallow water equation modules, etc. However, the calculation cost is extremely high. In particular, two-dimensional high-resolution hydrodynamic models, although they have high simulation accuracy for surface flood dynamics, their running time is long and cannot meet the real-time requirements. In actual flood emergency management, accurate flood prediction information is often needed within a short time to guide decision-making, and the high calculation cost of such models makes it difficult for them to be widely used in practice.

[0005] To solve the timeliness problem, data-driven surrogate models, such as Response Surface Models (RSM) and Low-Fidelity Models (LFM), have become the main strategy to improve computational efficiency. As an end-to-end model, RSM greatly reduces the computational cost by approximating the behavior of high-fidelity models through statistical or empirical formulas. However, RSM has difficulty in ensuring convergence under multi-module coupling conditions, and due to the black-box nature of the model process, domain experts have doubts about its credibility. In addition, although LFM retains physical information to some extent, it may lose key information when dealing with complex urban drainage systems, especially when the resolution is adjusted, the coupling between the surface module and the drainage system may fail, leading to numerical instability or even simulation failure. In addition, the common defects of these data-driven surrogate models are the lack of explainability and physical consistency. In practical application scenarios, urban flood problems not only involve complex water flow movement, but also closely related to various factors such as urban topography, drainage pipe network layout, land use type, etc. The lack of explainability makes the model unable to provide clear physical mechanism explanation for decision-makers when facing actual problems, and it is difficult to effectively analyze and adjust the model results according to the actual situation. Real-world urban hydrological and hydrodynamic systems follow strict physical laws such as mass conservation and momentum conservation. However, the above-mentioned surrogate models often fail to incorporate these physical laws into the model construction process. Taking mass conservation as an example, in the drainage pipe network, the water inflow of a node should be equal to the sum of the water outflow of the node and the change of the node's storage water. However, the surrogate model may not accurately reflect this relationship, leading to unreasonable simulation results in the physical layer. The lack of physical consistency makes the model's output deviate greatly from the actual situation when simulating extreme rainfall or special drainage conditions, and cannot provide reliable basis for urban flood control and disaster reduction.

[0006] Existing research shows that for the behavior of complex urban hydrological and hydrodynamic systems such as urban flooding, by decoupling the system, especially emphasizing the attention to dynamic information of the drainage system, the accuracy and explainability of the model can be improved to some extent, but the existing methods only stay at the level of "physical supervision information", without deeply integrating physical mechanisms (such as the core constraints of Saint-Venant equation and shallow water equation), which cannot explain the causal relationship of the flood process from the essence, and the adaptability to complex urban environments (such as differences in pipe network layout and topographic changes) is insufficient.

[0007] An article by Hancheng Ren, Bo Pang, Gang Zhao, Haijun Yu, Peinan Tian, and Chenran Xie entitled "Incorporating dynamic drainage supervision into deep learning for accurate real-time flood simulation in urban areas" published in Water Research in November 2024 proposes that urban flooding has become a common problem faced by cities worldwide. Urban flood dynamics differ significantly from those of natural catchments, mainly due to complex drainage systems and high spatial heterogeneity of urban surfaces, which pose considerable challenges for accurate and rapid flood simulation. In this study, a new urban drainage supervised flood model (UDFM) for urban flood simulation is proposed. Urban flood processes are decoupled into drainage routing and surface flood inundation. Based on a physical and deep learning drainage model, a hybrid module combining deep learning with dimensionality reduction algorithms is adopted to convert the 1D drainage overflow process into a high-resolution spatiotemporal 2D flood inundation process. Compared with existing advanced surrogate models for rapid flood simulation, UDFM more comprehensively and accurately represents the role of drainage systems in urban flood dynamics, providing high-resolution predictions of flood depth and velocity. When applied to a highly urbanized area in Shenzhen, UDFM-deep learning demonstrates real-time prediction capability and high accuracy, especially in simulating flow velocity, with average Nash-Sutcliffe efficiency coefficients improved by 0.112 and 0.251 compared with response surface models (RSM) and low-fidelity models (LFM), respectively. These findings highlight the critical importance of drainage system overflow in simulating urban surface floods. UDFM improves accuracy, flexibility, interpretability, and scalability without requiring additional physical model construction. This study introduces a new hierarchical surrogate model structure for urban flood simulation, providing valuable insights for rapid flood warning and risk management in urban environments. The drawback of this method is that it does not consider the physical processes of urban hydrology and hydrodynamics and the law of conservation of mass, which cannot regulate the convergence direction of deep learning techniques to align with disciplinary prior knowledge, limiting the generalization ability of the model in complex urban hydrological situations.

[0008] A city flood disaster prediction method is disclosed in Chinese patent application CN109146140A, which comprises the following steps: step 1, dividing the city into catchment areas based on geographic information data; step 2, calculating the water capacity of the catchment area; step 3, based on weather forecast, calculating the water accumulation of the catchment area in the future; step 4, when the water accumulation of the catchment area is greater than the warning threshold, issuing a flood disaster warning. The disadvantage of this method is that it only relies on geographic information data and weather forecast data for prediction, the data source is too single, lacks simulation of such complex runoff generation and runoff concentration mechanism, and is prone to underestimate the flood risk. SUMMARY

[0009] To solve the above technical problems, the present application provides a kind of city hydrology multi-physical field decoupling supervision AI flood real-time simulation method, by hierarchical decoupling city hydrology water power system, combine multi-scale physical constraint supervision, embed physical law such as Saint-Venant equation, shallow water equation in deep learning model training, separate one-dimensional drainage pipe network sub-module and two-dimensional surface flood module, design multi-scale physical constraint loss function for different links, realize micro, meso, macro multi-scale physical constraint fusion, at micro level, embed the residual constraint of Saint-Venant equation in the neural network agent of drainage system;At meso level, embed the residual constraint of shallow water equation in surface flood evolution system;At macro level, realize the water balance constraint of system as a whole;Realize the deep fusion of data-driven rapid calculation and physical mechanism strict constraint.

[0010] The purpose of the present application is to provide a kind of city hydrology multi-physical field decoupling supervision AI flood real-time simulation method, including obtaining historical rainfall data and real-time rainfall data, training stage and application stage, the training stage includes the following steps:

[0011] Step 01: input the historical rainfall data into one-dimensional drainage pipe network sub-module, calculate micro constraint data, the micro constraint data includes node overflow and water accumulation;

[0012] Step 02: input the historical rainfall data into two-dimensional surface flood sub-module, generate meso reference data, the meso reference data includes surface flood spatio-temporal distribution;

[0013] Step 03: empirical orthogonal function analysis is carried out on two-dimensional water power data, and the first k order principal component is extracted, and the data dimension is reduced, wherein, M is the number of spatial points;

[0014] Step 04: use multi-physical field constraint supervision module to constrain training, through hierarchical supervision structure architecture, learn rainfall-overflow-flood mapping relationship by using recurrent neural network, while applying micro, meso and macro constraint loss;

[0015] The application stage comprises the following sub-steps:

[0016] Step 11: input the real-time rainfall data into the one-dimensional drainage pipe network sub-module to calculate the node overflow data Q ovf ;

[0017] Step 12: the flood decoder maps the overflow data into the EOF time mode, combines the spatial mode extracted in the training stage, and reconstructs the high-resolution flood depth and flow velocity distribution,

[0018]

[0019] wherein, X is the original data set matrix, i is the index value of the mode, e i is the spatial mode, a i is the predicted time mode, T is the transpose symbol, is the real space, M is the size of the time dimension, N is the size of the spatial dimension.

[0020] Preferably, the one-dimensional drainage pipe network sub-module is composed of continuity equation and momentum equation, respectively expressed as:

[0021]

[0022]

[0023] wherein, A is the cross-sectional area of the water, t is the time, Q is the cross-sectional flow, Q = u c A , u c is the average flow velocity of the cross section, x is the coordinate along the axis of the pipe, q l is the inflow rate of the side of the unit length pipe, h is the water depth of the cross section, g is the acceleration of gravity, s 0 is the pipe bottom slope, s f is the friction slope, u l is the flow velocity of the side inflow.

[0024] In any of the above schemes, preferably, the two-dimensional surface flood sub-module couples the runoff process by using a full-distributed hydrological calculation method,

[0025]

[0026]

[0027]

[0028] wherein, x and y are spatial coordinates, h is water depth, z is water surface height, u and v are flow velocities in the directions of x and y respectively; V |is flow velocity, , r is rainfall intensity, f is infiltration intensity, c is drainage intensity, n is roughness coefficient, g is gravitational acceleration.

[0029] In any of the above schemes, preferably, the multi-physical field constraint supervision module realizes deep integration of mechanism and data through hierarchical physical constraints based on a three-level architecture of decoupling-constraint-prediction.

[0030] In any of the above schemes, preferably, the multi-physical field constraint supervision module reconstructs the loss function for microscopic physical constraints into a microscopic mass conservation loss, which forces the one-dimensional pipe network module to satisfy the continuity equation and momentum equation residual constraints of Saint-Venant equation, and the loss function is

[0031]

[0032] wherein, γ 1 and γ 2 are weight coefficients, E is the mathematical expectation of the loss random variable.

[0033] In any of the above schemes, preferably, the multi-physical field constraint supervision module reconstructs the loss function for mesoscopic physical constraints into a mesoscopic momentum conservation loss, which forces the two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of shallow water equation, and the loss function is

[0034]

[0035] wherein, γ 3 is a weight coefficient.

[0036] In any of the above schemes, preferably, the multi-physical field constraint supervision module reconstructs the loss function of macroscopic physical constraints into a macroscopic water balance loss, forcing the system as a whole to satisfy the global water balance, including the water exchange between the pipe network and the ground, and the loss function is

[0037]

[0038] wherein, γ 4 is a weight coefficient, Q in is the inflow of the target hydrological system, Q out is the outflow of the target hydrological system, Ω is the calculation domain of the flood evolution simulation, ΔV total is the total water storage change of the system.

[0039] In any of the above schemes, preferably, the total loss function is integrated as

[0040] L total L data L micro L meso L macro

[0041] In any of the above schemes, preferably, the hierarchical supervision structure is a rainfall-overflow encoder-drainage dynamic information supervision-overflow-flood decoder.

[0042] In any of the above schemes, preferably, the rainfall-overflow encoder uses a recurrent neural network to extract rainfall time series features, capturing the dynamic effects of rainfall intensity, duration and spatial distribution.

[0043] In any of the above schemes, preferably, the overflow-flood decoder combines data dimensionality reduction techniques to compress high-dimensional hydrodynamic data into low-dimensional time modes, maps overflow data to flood spatio-temporal distribution through a recurrent neural network, and realizes sub-second fast prediction.

[0044] The present application proposes an AI flood real-time simulation method for urban hydrology multi-physical field decoupling supervision, which breaks through the traditional model "either-or" driving mode, and provides a new path for solving urban flood real-time early warning, drainage system optimization and other key problems, and has important practical significance for improving urban flood risk management capability. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 ​​​​​This is a flowchart of a preferred embodiment of the AI-based real-time flood simulation method for urban hydrological multiphysics field decoupling supervision according to the present invention.

[0046] Figure 2 This is a schematic diagram of the structure of an embodiment of the AI ​​real-time flood simulation method for urban flood intelligent physical prediction model according to the present invention, which is based on the decoupled supervision of urban hydrological multi-physics fields. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] like Figure 1 As shown, an AI-based real-time flood simulation method with decoupled supervision of urban hydrology multiphysics fields is implemented in step 100, which involves acquiring historical rainfall data and real-time rainfall data.

[0050] Perform step 110 to train the model, including the following steps:

[0051] Execute step 111, input the historical rainfall data into the one-dimensional drainage network submodule, and calculate the micro-constraint data. The micro-constraint data includes nodal overflow and water accumulation. The one-dimensional drainage network submodule consists of a continuity equation and a momentum equation, which are expressed as follows:

[0052]

[0053]

[0054] in, A The cross-sectional area of ​​the water passage. t For time, Q For cross-sectional flow, Q = u c A , u c The cross-sectional average velocity is... x The coordinates are along the pipeline axis. q l The side inflow rate per unit length of pipe. h The depth of the cross-section of the water passage. g It is the acceleration due to gravity. s 0 represents the bottom slope of the pipeline. s f For frictional gradient, u l The velocity of the side inflow.

[0055] The step 112 is performed to input the historical rainfall data into a two-dimensional surface flood sub-module to generate mesoscopic reference data, the mesoscopic reference data including a surface flood spatio-temporal distribution, the two-dimensional surface flood sub-module coupling a runoff process by using a full-distributed hydrological calculation method,

[0056]

[0057]

[0058]

[0059] wherein, x and y are spatial coordinates, h is a water depth, z is a water surface height, u and v are flow velocities in directions of x and y respectively; V is a flow velocity, , r is a rainfall intensity, f is an infiltration intensity, c is a drainage intensity, n is a roughness coefficient, g is a gravitational acceleration.

[0060] The step 113 is performed to perform an empirical orthogonal function analysis on the two-dimensional hydrodynamic data to extract the first k principal components to reduce the data dimension, wherein, M is a spatial point number.

[0061] The step 114 is performed to use a multi-physical field constraint supervision module to constrain training, to learn a rainfall-overflow-flood mapping relationship by using a recurrent neural network through a hierarchical supervision structure architecture, and to simultaneously apply microscopic, mesoscopic and macroscopic constraint losses to ensure that an output of the neural network model conforms to a physical law, the multi-physical field constraint supervision module being based on a three-level architecture of decoupling-constraint-prediction to realize deep fusion of mechanism and data through hierarchical physical constraints.

[0062] The loss function of the multi-physical field constraint supervision module for the microscopic physical constraint is reconstructed as a microscopic mass conservation loss, the microscopic mass conservation loss forcing the one-dimensional pipe network module to satisfy a continuity equation and a momentum equation residual constraint of the Saint-Venant equation, the loss function being

[0063]

[0064] wherein, γ 1 and γ 2 are weight coefficients, E is a mathematical expectation of a loss random variable.

[0065] The multi-physical field constraint supervision module reconstructs the loss function for mesoscopic physical constraints into a mesoscopic momentum conservation loss, forcing the two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of the shallow water equation, and the loss function is

[0066]

[0067] Wherein, γ 3 is a weight coefficient.

[0068] The multi-physical field constraint supervision module reconstructs the loss function for macroscopic physical constraints into a macroscopic water balance loss, forcing the system as a whole to satisfy the global water balance, including the water exchange between the pipe network and the surface, and the loss function is

[0069]

[0070] Wherein, γ 4 is a weight coefficient, Q in is the inflow of the target hydrological system, Q out is the outflow of the target hydrological system, Ω is the calculation domain of flood evolution simulation, ΔV total is the total water storage change of the system.

[0071] The total loss function set is

[0072] L total = L data + L micro + L meso + L macro .

[0073] The hierarchical supervision structure is a rainfall-inflow encoder-drainage dynamic information supervision-inflow-flood decoder.

[0074] The rainfall-inflow encoder uses a recurrent neural network to extract rainfall time series features, capturing the dynamic effects of rainfall intensity, duration and spatial distribution.

[0075] The inflow-flood decoder combines data dimensionality reduction techniques to compress high-dimensional hydrodynamic data into low-dimensional time modes, maps inflow data to flood spatio-temporal distribution through a recurrent neural network, and realizes sub-second fast prediction.

[0076] Step 120 is performed to use the model to implement prediction of urban waterlogging, including the following sub-steps:

[0077] Step 121 is performed to input the real-time rainfall data into a one-dimensional drainage network submodule to calculate node overflow data Q ovf ;

[0078] Step 122 is performed, in which the flood decoder maps the overflow data into an EOF time mode, combines the spatial mode extracted in the training stage, and reconstructs the high-resolution flood depth and flow velocity distribution,

[0079]

[0080] wherein, X is the original data set matrix, and is the flood depth and flow velocity description data at different times and spatial positions, i is the index value of the mode, from 1 to k , indicating different empirical orthogonal function modes, e i is the spatial mode, a i is the predicted time mode, T is the transpose symbol, is the real space, M is the size of the time dimension, N is the size of the spatial dimension.

[0081] Example Two

[0082] The prior art proposes a new hierarchical surrogate model structure for urban flood simulation, providing valuable insights for rapid flood warning and risk management in urban environments. The disadvantage of this method is that it does not consider the physical processes of urban hydrology and water dynamics and the law of conservation of mass, which cannot standardize the convergence direction of the deep learning technique to align with the subject prior knowledge, resulting in limited generalization ability of the model under complex urban hydrological situations. Specifically, urban flood evolution essentially follows the principles of mass and momentum conservation described by the Saint-Venant equations. However, when UDFM converts the drainage overflow process into a two-dimensional flooding process, it mainly relies on data-driven deep learning modules and fails to explicitly embed the physical mechanisms of surface-pipe network water flow interaction (such as key hydrodynamic processes like orifice outflow and unsteady flow in open channels). The absence of such physical mechanisms may lead to flow balance deviations or abnormal flow velocity distributions under extreme rainfall conditions, drainage system blockage, or pipe network pressure flow in non-design conditions. For example, in the Shenzhen case, although high-resolution prediction is achieved, when facing sudden changes in drainage capacity, the model may accumulate water volume calculation errors due to the lack of mass conservation constraints.

[0083] The "urban hydrology multi-physical field decoupling supervision AI flood real-time simulation method" provided by the application is just for the core problem of the existing model in the calculation efficiency and the lack of physical mechanism fusion: through hierarchical decoupling of the urban hydrology and water power system (separation of one-dimensional drainage pipe network sub-module and two-dimensional surface flood module), combined with multi-scale physical constraint supervision (microscopic pipe network mass conservation, mesoscopic surface flow momentum conservation, macroscopic system water balance), the physical laws such as Saint-Venant equation and shallow water equation are embedded in the deep learning model training, realizing the deep integration of "data-driven rapid calculation" and "strict constraint of physical mechanism". The method breaks through the traditional model "non-this or that" driving mode, provides a new path for solving the key problems of urban flood real-time early warning, drainage system optimization and the like, and has important practical significance for improving the urban flood risk management ability.

[0084] The application realizes efficient, accurate and physical law-compliant urban flood simulation through hierarchical decoupling, multi-scale physical constraint embedding and deep learning technology. The specific technical scheme is as follows:

[0085] I. Core technical framework: hierarchical decoupling and multi-physical field constraint fusion

[0086] The urban hydrology and water power system is decoupled into "efficient one-dimensional drainage pipe network sub-module" (Saint-Venant equation driven) and "lightweight two-dimensional surface flood module" (data driven combined with shallow water equation constraint), and a "micro-mesoscopic-macroscopic" multi-scale physical constraint supervision system is constructed:

[0087] 1. Microscopic constraint (drainage pipe network layer): based on Saint-Venant equation, mass and momentum conservation constraints are applied to pipe network nodes and pipe sections to ensure node water balance (inflow = outflow + water accumulation change) and avoid non-physical overflow flow prediction;

[0088] 2. Mesoscopic constraint (surface flood layer): based on the two-dimensional shallow water equation, momentum conservation and mass conservation joint constraints are applied to the surface flow field to ensure that the velocity and water depth distribution conforms to the physical law of water flow motion;

[0089] 3. Macroscopic constraint (system layer): through the global water balance equation, the total inflow, total outflow and total storage capacity relationship of the entire urban water system is constrained to ensure macroscopic physical consistency.

[0090] II. Model architecture: urban flood intelligent physical prediction model (IPPM4UF), as shown in Figure 2 The "decoupling-constraint-prediction" three-level architecture is constructed, including three core modules:

[0091] 1. Decoupled physical module (microscopic + mesoscopic)

[0092] One-dimensional drainage pipe network sub-module: in the simulation of urban drainage pipe network, one-dimensional Saint-Venant equation set is the basic equation for describing the unsteady flow in the pipe, which consists of continuity equation and momentum equation, as follows:

[0093] The continuity equation and the momentum equation are respectively expressed as:

[0094]

[0095]

[0096] wherein, A : cross-sectional area of water (m 2 ), which is related to pipe fullness, pipe diameter, etc. t : time (s) Q : cross-sectional flow (m 3 / s), Q uA , u is the cross-sectional average flow velocity; x : coordinate along the pipe axis (m) q l : lateral inflow rate per unit length of pipe (m 2 / s), such as rainwater inflow or leakage. h : water depth of cross section (m), reflecting the water level height in the pipe; g : gravitational acceleration (m / s 2 ); s 0: pipe bottom slope (dimensionless); s f : friction slope (dimensionless), reflecting the energy loss caused by water flow resistance; u l : flow velocity of lateral inflow (m / s).

[0097] The continuity equation ensures the conservation of water flow mass in the pipe, i.e. the change of cross-sectional water volume per unit time is equal to the sum of the difference between inflow and outflow and lateral inflow; the momentum equation considers the effects of gravity, friction and lateral inflow on water flow movement, and describes the change of flow with time and space.

[0098] Two-dimensional surface flood sub-module: in the simulation of urban surface flood, two-dimensional shallow water equation set is used to describe the water flow movement state and characteristics of urban surface during flood period, showing the diffusion of flood in horizontal direction, the change of flow velocity and the water depth distribution in different areas. The present application adopts a fully distributed hydrological calculation method, coupling the runoff process to the right side of the continuity equation, such as the additional source term.

[0099]

[0100] ​

[0101]

[0102] where, t denotes time; x and y denotes spatial coordinates; h is the water depth; z is the water surface height; u and v are the flow velocities in the x and y directions, respectively; V is the magnitude of the flow velocity, i.e. ; r is the rainfall intensity; f is the infiltration intensity; c is the drainage intensity; n is the roughness coefficient; g is the gravitational acceleration;

[0103] 2. Multi-physical field constraint supervision module

[0104] Based on the "decoupling-constraint-prediction" three-level architecture, the model realizes the deep integration of mechanism and data through hierarchical physical constraints. The following is the loss function reconstruction for micro, meso, and macro physical constraints:

[0105] (1) Microscopic mass conservation loss (Saint-Venant equation constraint)

[0106] Core goal: Force one-dimensional pipe network module to satisfy the continuity equation and momentum equation residual constraints of the Saint-Venant equation.

[0107] Loss function design:

[0108]

[0109] where the first term: based on the continuity equation residual, ensure the balance of cross-section water volume change and side inflow per unit time; the second term: based on the momentum equation residual, constrain the physical consistency of flow change, gravity, friction, and side inflow momentum transmission; the weight coefficient γ 1, γ 2 dynamically adjusts the balance between data fitting and physical constraints.

[0110] (2) Meso-momentum conservation loss (shallow water equation constraint)

[0111] Core goal: Force two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of the shallow water equation.

[0112] Loss function design:

[0113]

[0114] where, x-direction and y-direction momentum equation residuals: constrain the momentum variation of surface water flow in the horizontal direction, pressure gradient and friction resistance, respectively; roughness coefficient n The physical mechanism of surface resistance is introduced by the Manning formula; the residual term directly reflects the degree of momentum conservation that the shallow water equation does not satisfy.

[0115] Pure data-driven models (such as UDFM) do not explicitly consider the law of momentum conservation, resulting in distortion of key hydrodynamic characteristics such as water flow direction and flow velocity distribution. For example, in areas such as overpass topography, the predicted results may appear to violate the actual water flow direction. The present invention can minimize the momentum equation residual, and the flow velocity field predicted by the model u , v strictly satisfies the law of momentum conservation, especially in complex topography such as steep slopes and river junctions, which can more accurately simulate the water flow turning and energy dissipation process.

[0116] (3) Macroscopic water balance loss (global conservation constraint)

[0117] Core goal: Force the system as a whole to satisfy the global water balance, including the water exchange between the pipe network and the surface.

[0118] Loss function design:

[0119]

[0120] where, the first term: global integral based on the continuity equation of the two-dimensional shallow water equation, constraining the overall water balance of the surface runoff-infiltration-drainage process; the second term: based on the input and output flow difference and the total water change at the interface between the pipe network and the surface, to ensure the water conservation across modules; ΔV total is the total water storage change of the system, which is calculated in real time through differentiable programming technology.

[0121] In urban flood simulation, global water balance is the core index to test the physical reasonableness of the model. The traditional numerical driving artificial intelligence model has the problem of inaccurate water exchange across systems, which makes the local conservation and global balance disconnected, resulting in the accumulation of errors in long-time simulation, making the flood simulation deviate from the actual situation. This design can minimize the global integral term to ensure that the surface water change at any time is strictly balanced with the rainfall, infiltration and drainage process.

[0122] (4) Total loss function integration

[0123] L total = L data + L micro + L meso+ L macro

[0124] Optimization strategy:

[0125] Hierarchical weight distribution: γ 1~ γ 4 Adopt adaptive learning mechanism, dynamically adjust the constraint strength according to the training stage;

[0126] Physically guided gradient propagation: through the automatic differentiation of the residual term, the direct correction of the physical equation to the neural network parameters is realized.

[0127] 3. Artificial intelligence prediction module

[0128] Rainfall-inflow encoder: adopt recurrent neural network (such as LSTM) to extract rainfall time series features, capture the dynamic influence of rainfall intensity, duration and spatial distribution;

[0129] Inflow-flood decoder: combined with data dimensionality reduction techniques such as empirical orthogonal function (Empirical Orthogonal Function, EOF), high-dimensional hydrodynamic data (water depth, flow velocity) are compressed into low-dimensional time modes, through recurrent neural network mapping inflow data to flood spatio-temporal distribution, realizing sub-second level fast prediction;

[0130] Multi-task learning: jointly optimize the rainfall-inflow-flood mapping relationship, require the decoupling variables of each system to meet the multi-scale physical constraints at the same time.

[0131] Example three

[0132] A kind of city hydrology multi-physical field decoupling supervised AI flood real-time simulation method, the steps are as follows:

[0133] 1. Training phase (data-physical collaborative learning)

[0134] S1: Physical module data generation

[0135] Input historical rainfall data to one-dimensional drainage pipe network sub-module, calculate node inflow and water accumulation (microscopic constraint data); input to two-dimensional hydrodynamic model (such as MIKEURBAN) to generate surface flood spatio-temporal distribution (mesoscopic reference data).

[0136] S2: EOF dimensionality reduction processing

[0137] Perform empirical orthogonal function analysis on two-dimensional hydrodynamic data (water depth, flow velocity) to extract the first k principal components (, M is the number of spatial points), reduce the data dimension.

[0138] S3: Multi-physical field constraint training

[0139] ​Through a hierarchical supervision structure architecture (rainfall-overflow encoder-drainage dynamic information supervision-overflow-flood decoder), a rainfall-overflow-flood mapping relationship is learned by using a recurrent neural network, while microscopic, mesoscopic and macroscopic constraint losses are applied to ensure that the model output conforms to the physical law.

[0140] 2. Application stage (real-time flood prediction)

[0141] S4: Real-time rainfall input and decoupling calculation

[0142] Real-time rainfall data is input into a one-dimensional drainage pipe network submodule, and node overflow data is quickly calculated to retain the engineering scheduling compatibility of the physical model (such as pump station and gate control).

[0143] S5: Spatiotemporal reconstruction of flood

[0144] The flood decoder maps the overflow data to the EOF time mode, and combines the spatial mode extracted in the training stage to reconstruct the high-resolution flood depth and flow velocity distribution:

[0145]

[0146] wherein, e i for the spatial mode, a i for the predicted time mode.

[0147] The application adopts decoupled hierarchical modeling to separate the high-consumption two-dimensional hydrodynamic module and the efficient one-dimensional pipe network module, and replaces complex physical calculation with deep learning, taking into account accuracy and efficiency;

[0148] The application adopts multi-scale physical constraints, which first embeds the core constraints of Saint-Venant equation (microscopic), shallow water equation (mesoscopic) and water balance equation (macroscopic) into deep learning, realizes "physical mechanism can be verified, and prediction results can be explained";

[0149] The application adopts data-physical bidirectional fusion, and the physical module provides prior constraints for the data-driven model, and the deep learning feeds back to the physical model optimization, forming a closed loop of "mechanism guiding learning and learning correcting mechanism";

[0150] The application improves real-time performance and generalization ability, and EOF dimensionality reduction and LSTM prediction realize sub-second level calculation, supporting real-time early warning under complex conditions such as extreme rainfall and pipe network scheduling.

[0151] For better understanding of the present application, the above detailed description is made in combination with the specific embodiments of the present application, but is not a limitation to the present application. Any simple modification made to the above embodiments according to the technical essence of the present application still belongs to the scope of the technical scheme of the present application. In the specification, each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be understood by mutual reference. For the system embodiments, since they basically correspond to the method embodiments, the description is relatively simple, and the relevant parts can be understood by referring to the part of the method embodiments.

Claims

1. A real-time AI-based flood simulation method with decoupled supervision of multi-physics fields in urban hydrology, comprising acquiring historical and real-time rainfall data, a training phase, and an application phase, characterized in that, The training phase includes the following steps: Step 01: Input the historical rainfall data into the one-dimensional drainage network submodule to calculate the micro-constraint data, which includes node overflow and water accumulation. Step 02: Input the historical rainfall data into the two-dimensional surface flood submodule to generate meso-level reference data, which includes the spatiotemporal distribution of surface floods; Step 03: Perform empirical orthogonal function analysis on the two-dimensional hydrodynamic data to extract the first k principal components and reduce the data dimensionality, where k≪M, and M is the number of spatial points; Step 04: Use the multiphysics constraint supervision module to constrain training. Through a hierarchical supervision structure, the recurrent neural network is used to learn the mapping relationship between rainfall, overflow and flood. At the same time, micro, meso and macro constraint losses are applied. The multiphysics constraint supervision module is based on a three-level architecture of decoupling-constraint-prediction. It achieves deep integration of mechanism and data through hierarchical physical constraints. The multiphysics constraint monitoring module reconstructs the loss function for microphysical constraints into a micro-mass conservation loss. This micro-mass conservation loss forces the one-dimensional pipeline module to satisfy the residual constraints of the continuity equation and momentum equation of the Saint-Venant equation. The loss function is... ， in, γ 1 and γ 2 is the weighting coefficient. E For the mathematical expectation of the loss random variable, A The cross-sectional area of ​​the water passage. t For time, Q For cross-sectional flow, Q = u c A , u c The cross-sectional average velocity is... x The coordinates are along the pipeline axis. q l The side inflow rate per unit length of pipe. h The depth of the cross-section of the water passage. g It is the acceleration due to gravity. s 0 represents the bottom slope of the pipeline. s f For frictional gradient, u l The velocity of the inflow from the side; The application phase includes the following sub-steps: Step 11: Input the real-time rainfall data into the one-dimensional drainage network submodule and calculate the node overflow data. Q ovf ; Step 12: The flood decoder maps the overflow data to EOF time modes and, combined with the spatial modes extracted during the training phase, reconstructs high-resolution flood depth and velocity distributions. ， in, X The original dataset matrix, i The index value of the modality. e i For spatial modes, a i For the predicted time modes, T It is the transpose symbol. For the real number space, M For the size of the time dimension, N This represents the size of the spatial dimension.

2. The AI-based real-time flood simulation method with decoupled supervision of urban hydrological multiphysics fields as described in claim 1, characterized in that, The one-dimensional drainage network submodule consists of a continuity equation and a momentum equation, which are expressed as follows: ， 。 3. The AI-based real-time flood simulation method with decoupled supervision of urban hydrological multiphysics fields as described in claim 2, characterized in that, The two-dimensional surface flood submodule employs a fully distributed hydrological computation method to couple the runoff generation process. ， ， ， in, x and y For spatial coordinates, h Because of the water depth, z The height of the water surface. u and v They are along x and y Flow velocity in direction; | V | represents the flow rate. , r Rainfall intensity, f Infiltration strength, c For drainage strength, n Roughness coefficient g This is the acceleration due to gravity.

4. The AI-based real-time flood simulation method with decoupled supervision of urban hydrological multiphysics fields as described in claim 3, characterized in that, The multiphysics constraint monitoring module reconstructs the loss function for mesophysical constraints into a mesophysical momentum conservation loss, forcing the two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of the shallow water equation. The loss function is... ， in, γ 3 represents the weighting coefficient.

5. The AI-based real-time flood simulation method with decoupled supervision of urban hydrological multiphysics fields as described in claim 4, characterized in that, The multiphysics constraint monitoring module reconstructs the loss function for macroscopic physical constraints into a macroscopic water balance loss, forcing the entire system to satisfy global water balance, including water exchange between the pipe network and the surface. The loss function is... ， in, γ 4 represents the weighting coefficient. Q in For the inflow of the target hydrological system, Q out Let Ω be the outflow of the target hydrological system, Ω be the computational domain for flood evolution simulation, and ΔV be the value of V. total This represents the change in the total water storage of the system.

6. The AI-based real-time flood simulation method with decoupled supervision of urban hydrological multiphysics fields as described in claim 5, characterized in that, The total loss function is integrated as L total = L data + L micro + L meso + L macro 。 7. The AI-based real-time flood simulation method for decoupled supervision of urban hydrological multiphysics fields as described in claim 6, characterized in that, The hierarchical supervision architecture consists of a rainfall-overflow encoder, a drainage dynamic information supervisor, and an overflow-flood decoder.

8. The AI-based real-time flood simulation method with decoupled supervision of urban hydrological multiphysics fields as described in claim 7, characterized in that, The rainfall-overflow encoder uses a recurrent neural network to extract rainfall time series features and capture the dynamic effects of rainfall intensity, duration, and spatial distribution.

Images