AI flood real-time simulation method for urban hydrological multi-physics field decoupling supervision

Abstract

The invention provides an AI flood real-time simulation method for urban hydrological multi-physics field decoupling supervision, and the method comprises the steps: constructing an intelligent physical prediction model of urban flood, carrying out the layered decoupling of an urban hydrological hydrodynamic system, separating a one-dimensional drainage pipe network submodule from a two-dimensional surface flood module, designing a multi-scale physical constraint loss function for different links, and carrying out the real-time simulation of the AI flood. Microcosmic, mesoscopic and macroscopic multi-scale physical constraint fusion is realized, and residual constraints of the Saint-Venant equation are embedded in a neural network agent of a drainage system in a microcosmic level; in the mid-level, embedding the residual constraint of the shallow water equation in the surface flood routing system; in the macroscopic level, the water balance constraint of the whole system is realized; urban hydrology and hydrodynamic force are decoupled, and multi-scale physical field supervision constraint is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy projects, and in particular to an AI real-time flood simulation method for decoupling and supervision of multi-physical fields in urban hydrology. Background Art

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

[0003] In the field of urban flood risk management, flood simulation models are important tools for ensuring urban safety. However, existing single-driven models (physically driven numerical models, data-driven proxy 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. For example, multi-module structured models, by coupling hydrological runoff modules, one-dimensional Navier-Stokes equation modules, two-dimensional shallow water equation modules, etc., can simulate urban hydrodynamic processes more comprehensively, but the calculation cost is extremely high. Especially two-dimensional high-resolution hydrodynamic models, although they have high simulation accuracy for surface flood dynamics, their running time is long and they cannot meet the real-time requirements. In actual flood emergency management, accurate flood prediction information is often required to guide decision-making in a short time, and the high calculation cost of such models makes it difficult to be widely applied in practice.

[0005] To address the issue of timeliness, data-driven surrogate models, such as Response Surface Models (RSM) and Low-Fidelity Models (LFM), have become the main strategies for improving computational efficiency. As an end-to-end model, the response surface model approximates the behavior of the high-fidelity model through statistical or empirical formulas, greatly reducing the computational cost. However, it is difficult for RSM to ensure 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 the low-fidelity model retains physical information to a certain extent, it is prone to losing 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 go wrong, resulting in numerical instability or even simulation failure. Moreover, the common defect of these numerically driven surrogate models is the lack of interpretability and physical consistency. In practical application scenarios, urban flood problems are not only related to complex water flow movements but also closely related to various factors such as the topography and geomorphology of the city, the layout of drainage pipe networks, and land use types. The lack of interpretability makes it impossible for the model to provide a clear physical mechanism explanation for decision-makers when facing practical problems, and it is difficult to effectively analyze and adjust the model results according to the actual situation. The real urban hydro-hydraulic system follows strict physical laws, such as mass conservation and momentum conservation. However, it is often difficult for the above-mentioned surrogate models to incorporate these physical laws into the model construction process. Taking mass conservation as an example, in the drainage pipe network, the inflow of water into a node should be equal to the sum of the outflow of water from the node and the change in the stored water volume at the node. However, the surrogate model may not accurately reflect this relationship, resulting in unreasonable simulation results at the physical level. This lack of physical consistency makes the model output results with a large deviation from the actual situation when simulating extreme rainfall or special drainage conditions, and it cannot provide a reliable basis for urban flood control and disaster reduction.

[0006] Existing studies have shown that for the complex behavior of urban hydro-hydraulic systems such as urban floods, through system decoupling, especially by emphasizing the attention to the dynamic information of the drainage system, the accuracy and interpretability of the model can be improved to a certain extent. However, the existing methods only stay at the level of "physical supervision information" and do not deeply integrate physical mechanisms (such as the core constraints of the Saint-Venant equation and the shallow water equation), so they cannot essentially explain the causal relationship of the flood process and have insufficient adaptability to complex urban environments (such as differences in pipe network layout and terrain changes).

[0007] In November 2024, the article "Incorporating dynamic drainage supervision into deep learning for accurate real-time flood simulation in urban areas" by Hancheng Ren, Bo Pang, Gang Zhao, Haijun Yu, Peinan Tian, and Chenran Xie was published in *Water Research*. The article proposed that urban flooding has become a common problem faced by cities globally. The dynamics of urban floods differ significantly from those of natural watersheds, mainly due to the complex drainage system and the high spatial heterogeneity of urban surfaces, which pose considerable challenges to accurate and rapid flood simulation. In this study, an urban drainage supervised flood model (UDFM) for urban flood simulation was proposed. The urban flood process was decoupled into drainage paths and surface flood inundation. Based on physical and deep learning drainage models, a hybrid module combining deep learning and dimensionality reduction algorithms was used to convert the 1D drainage overflow process into a high-resolution spatio-temporal 2D flood process. Compared with existing advanced surrogate models for rapid flood simulation, UDFM more comprehensively and accurately represents the role of the drainage system in urban flood dynamics, providing high-resolution flood depth and velocity predictions. When applied to a highly urbanized area in Shenzhen, UDFM-deep learning demonstrated real-time prediction capabilities and high accuracy, especially in simulating flow velocities, with the average Nash efficiency coefficient increasing by 0.112 and 0.251 compared to the response surface model (RSM) and the low-fidelity model (LFM), respectively. These findings emphasize the crucial importance of drainage system overflows in simulating urban surface floods. UDFM improves accuracy, flexibility, interpretability, and scalability without requiring additional physical model construction. This study introduced 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 hydrodynamics and the law of mass conservation, and it cannot regulate the convergence direction of deep learning techniques to align with prior disciplinary knowledge, resulting in limited generalization ability of the model in complex urban hydrographic scenarios.

[0008] The Chinese invention patent application with the publication number CN109146140A discloses a method for predicting urban flood disasters, including: Step 1, dividing the urban catchment area based on geographic information data; Step 2, calculating the water capacity of the catchment area; Step 3, calculating the accumulated water volume in the catchment area for a future period based on weather forecasts; Step 4, when the accumulated water volume in 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 the simulation of such complex runoff generation and concentration mechanisms, and is prone to underestimating flood risks. Summary of the Invention

[0009] To solve the above technical problems, an AI real-time flood simulation method with decoupled supervision of multi-physical fields in urban hydrology proposed by the present invention, through hierarchical decoupling of the urban hydrological hydrodynamic system, combined with multi-scale physical constraint supervision, embeds physical laws such as the Saint-Venant equation and the shallow water equation into the deep learning model training, separates the one-dimensional drainage pipe network sub-module from the two-dimensional surface flood module, designs multi-scale physical constraint loss functions for different links, realizes the integration of multi-scale physical constraints at the micro, meso, and macro levels. At the micro level, the residual constraint of the Saint-Venant equation is embedded in the neural network agent of the drainage system; at the meso level, the residual constraint of the shallow water equation is embedded in the surface flood evolution system; at the macro level, the overall water balance constraint of the system is realized; realizes the deep integration of data-driven fast calculation and strict physical mechanism constraints.

[0010] The purpose of the present invention is to provide an AI real-time flood simulation method with decoupled supervision of multi-physical fields in urban hydrology, including obtaining historical rainfall data and real-time rainfall data, a training stage and an application stage. The training stage includes the following steps: Step 01: Input the historical rainfall data into the one-dimensional drainage pipe network sub-module to calculate micro-constraint data, where the micro-constraint data includes node overflow and accumulated water volume; Step 02: Input the historical rainfall data into the two-dimensional surface flood sub-module to generate meso-reference data, where the meso-reference data includes the spatio-temporal distribution of surface floods; Step 03: Perform empirical orthogonal function analysis on the two-dimensional hydrodynamic data to extract the first k order principal components to reduce the data dimension, where , M is the number of spatial points; Step 04: Use the multi-physical field constraint supervision module for constrained training. Through a hierarchical supervision structure, use a recurrent neural network to learn the rainfall-overflow-flood mapping relationship, and at the same time impose micro, meso, and macro constraint losses; The application stage includes the following sub-steps: Step 11: Input the real-time rainfall data into the one-dimensional drainage pipe network sub-module to calculate node overflow dataQ ovf ; 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 velocity distribution.

[0011]

[0012] where, X is the original dataset 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 number space, M is the size of the time dimension, N is the size of the spatial dimension.

[0013] Preferably, the one-dimensional drainage pipe network sub-module is composed of the continuity equation and the momentum equation, which are respectively expressed as:

[0014]

[0015]

[0016] where, A is the cross-sectional area of flow, t is the time, Q is the cross-sectional flow rate, Q = u c A , u c is the average cross-sectional velocity, x is the coordinate along the pipeline axis direction, q l is the lateral inflow rate per unit length of the pipeline, h is the water depth of the cross-sectional area of flow, g is the acceleration of gravity, s 0 is the pipeline bottom slope, s f is the friction slope, u l is the velocity of the lateral inflow.

[0017] In any of the above solutions, preferably, the two-dimensional surface flood sub-module couples the runoff generation process using the fully distributed hydrological calculation method.

[0018]

[0019]

[0020]

[0021] Among them, x and y are spatial coordinates, h is the water depth, z is the water surface height, u and v are the flow velocities along x and y directions respectively; | V | is the flow velocity, , r is the rainfall intensity, f is the infiltration intensity, c is the drainage intensity, n is the roughness coefficient, g is the acceleration of gravity.

[0022] Preferably in any of the above solutions, the multi-physical field constraint supervision module is based on a three-level architecture of decoupling-constraint-prediction, and realizes the deep fusion of mechanism and data through hierarchical physical constraints.

[0023] Preferably in any of the above solutions, the loss function of the multi-physical field constraint supervision module for microscopic physical constraints is reconstructed into a microscopic mass conservation loss, and the microscopic mass conservation loss forces the one-dimensional pipe network module to satisfy the residual constraints of the continuity equation and the momentum equation of the Saint-Venant equation. The loss function is

[0024]

[0025] Among them, γ 1 and γ 2 are weight coefficients, E is the mathematical expectation of the loss random variable.

[0026] Preferably in any of the above solutions, the loss function of the multi-physical field constraint supervision module for mesoscopic physical constraints is reconstructed 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. The loss function is

[0027]

[0028] Among them, γ 3 is the weight coefficient.

[0029] Preferably in any of the above solutions, the loss function of the multi-physical field constraint supervision module for macroscopic physical constraints is reconstructed into a macroscopic water balance loss, forcing the overall system to satisfy the global water balance, including the water exchange between the pipe network and the surface. The loss function is

[0030]

[0031] Among them, γ 4 is the 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 computational domain of flood routing simulation, and ΔV total is the change in the total water storage of the system.

[0032] Preferably, in any of the above solutions, the total loss function is integrated into L total = L data + L micro + L meso + L macro .

[0033] Preferably, in any of the above solutions, the hierarchical supervision structure is a rainfall-overflow encoder—drainage dynamic information supervision—overflow-flood decoder.

[0034] Preferably, in any of the above solutions, the rainfall-overflow encoder uses a recurrent neural network to extract the characteristics of the rainfall time series and capture the dynamic effects of rainfall intensity, duration, and spatial distribution.

[0035] Preferably, in any of the above solutions, the overflow-flood decoder combines data dimensionality reduction technology to compress high-dimensional hydrodynamic data into low-dimensional time modes, maps the overflow data to the flood spatio-temporal distribution through a recurrent neural network, and realizes sub-second fast prediction.

[0036] The present invention proposes an AI flood real-time simulation method for decoupling and supervising multi-physical fields in urban hydrology, breaking through the "either-or" driving mode of traditional models, providing a new path for solving key problems such as urban flood real-time warning and drainage system optimization, and having important practical significance for improving the urban flood risk management ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flowchart of a preferred embodiment of the AI flood real-time simulation method for decoupling and supervising multi-physical fields in urban hydrology according to the present invention.

[0038] Figure 2 is a structural schematic diagram of an embodiment of the urban flood intelligent physical prediction model of the AI flood real-time simulation method for decoupling and supervising multi-physical fields in urban hydrology according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0040] Embodiment 1 As Figure 1 shown, an AI real-time flood simulation method for decoupling and supervision of multi-physical fields in urban hydrology performs step 100 to obtain historical rainfall data and real-time rainfall data.

[0041] Perform step 110 to train the model, including the following steps: Perform step 111, input the historical rainfall data into the one-dimensional drainage pipe network sub-module to calculate microscopic constraint data, where the microscopic constraint data includes node overflow and water accumulation. The one-dimensional drainage pipe network sub-module consists of a continuity equation and a momentum equation, which are respectively expressed as:

[0042]

[0043]

[0044] Among them, A is the cross-sectional area of flow, t is time, Q is the sectional flow rate, Q = u c A , u c is the average sectional flow velocity, x is the coordinate along the pipeline axis direction, q l is the lateral inflow rate per unit length of the pipeline, h is the water depth of the cross-sectional area of flow, g is the acceleration due to gravity, s 0 is the pipeline bottom slope, s f is the friction slope, u l is the flow velocity of the lateral inflow.

[0045] Perform step 112, input the historical rainfall data into the two-dimensional surface flood sub-module to generate mesoscopic reference data, where the mesoscopic reference data includes the spatio-temporal distribution of surface floods. The two-dimensional surface flood sub-module couples the runoff generation process using a fully distributed hydrological calculation method,

[0046]

[0047]

[0048]

[0049] Among them, x and y are spatial coordinates, h is the water depth, z is the water surface height, u and v are respectively the flow velocities along x and y directions; | V | is the flow velocity, , r is the rainfall intensity, f is the infiltration intensity, c is the drainage intensity, n is the roughness coefficient, g is the acceleration due to gravity.

[0050] Execute step 113 to perform empirical orthogonal function analysis on the two-dimensional hydrodynamic data, extract the first k principal components, and reduce the data dimension. Among them, , M is the number of spatial points.

[0051] Execute step 114 to perform constrained training using the multi-physics field constraint supervision module. Through a hierarchical supervision structure architecture, use a recurrent neural network to learn the rainfall-overflow-flood mapping relationship, and at the same time impose microscopic, mesoscopic, and macroscopic constraint losses to ensure that the output of the neural network model conforms to physical laws. The multi-physics field constraint supervision module is based on a three-level architecture of decoupling-constraint-prediction, and realizes the deep integration of mechanism and data through hierarchical physical constraints.

[0052] The loss function of the multi-physics field constraint supervision module for microscopic physical constraints is reconstructed as a microscopic mass conservation loss. The microscopic mass conservation loss forces the one-dimensional pipe network module to satisfy the residual constraints of the continuity equation and momentum equation of the Saint-Venant equation. The loss function is

[0053]

[0054] Among them, γ 1 and γ 2 are weight coefficients, E is the mathematical expectation of the loss random variable.

[0055] The loss function of the multi-physics field constraint supervision module for mesoscopic physical constraints is reconstructed as a mesoscopic momentum conservation loss, which forces the two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of the shallow water equation. The loss function is

[0056]

[0057] Among them, γ 3 is the weight coefficient.

[0058] The multi-physical field constraint supervision module reconstructs the loss function for the macroscopic physical constraint into a macroscopic water balance loss, forcing the overall system to satisfy the global water balance, including the water exchange between the pipe network and the surface. The loss function is

[0059]

[0060] where γ 4 is the 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 computational domain of flood routing simulation, and ΔV total is the change in the total water storage of the system.

[0061] The total loss function is integrated as L total = L data + L micro + L meso + L macro .

[0062] The hierarchical supervision structure is a rainfall-overflow encoder—drainage dynamic information supervision—overflow-flood decoder.

[0063] The rainfall-overflow encoder uses a recurrent neural network to extract the characteristics of rainfall time series and capture the dynamic effects of rainfall intensity, duration, and spatial distribution.

[0064] The overflow-flood decoder combines data dimensionality reduction technology to compress high-dimensional hydrodynamic data into low-dimensional time modes, and maps the overflow data to the flood spatio-temporal distribution through a recurrent neural network to achieve sub-second rapid prediction.

[0065] Execute step 120 to perform real-time prediction of urban flood using the model, including the following sub-steps: Execute step 121, input the real-time rainfall data into the one-dimensional drainage pipe network sub-module to calculate the node overflow data Q ovf ; Execute step 122, the flood decoder maps the overflow data to the EOF time mode, and reconstructs the high-resolution flood depth and velocity distribution in combination with the spatial mode extracted in the training stage,

[0066]

[0067] where Xis the original data set matrix, describing the flood depth and flow velocity at different time and spatial locations. i The index value of the mode, from 1 to k , represents 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 number space, M is the size of the time dimension, N is the size of the spatial dimension.

[0068] Embodiment 2 The existing technology proposes a new hierarchical surrogate model structure for urban flood simulation, which provides 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 and material conservation laws of urban hydrology and hydrodynamics, and cannot regulate the convergence direction of deep learning technology to be close to disciplinary prior knowledge, resulting in limited generalization ability of the model in complex urban hydrological situations. Specifically, the evolution of urban floods essentially follows the principles of mass conservation and momentum conservation described by the Saint-Venant equations. 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 mechanism of surface-pipeline network water flow interaction (such as orifice outflow, open channel unsteady flow and other key hydrodynamic processes). The lack of this physical mechanism may cause the model to have flow balance deviation or abnormal flow velocity distribution under non-design conditions such as extreme rainfall conditions, drainage system blockage or pipeline pressure flow. For example, in the Shenzhen case, although high-resolution prediction is achieved, when facing the scenario of sudden change in drainage capacity, water volume calculation errors may be accumulated due to the lack of material conservation constraints.

[0069] The "AI real-time flood simulation method with decoupled supervision of urban hydrological multi-physics fields" proposed in this invention is aimed at the core problem of insufficient integration of computational efficiency and physical mechanism in existing models: by layered decoupling of urban hydrological and hydrodynamic systems (separation of one-dimensional drainage pipe network submodule and two-dimensional surface flood module), combined with multi-scale physical constraint supervision (micro-pipe network mass conservation, meso-surface flow momentum conservation, macro-system water balance), physical laws such as Saint-Venant equation and shallow water equation are embedded in deep learning model training to achieve the deep integration of "data-driven fast calculation" and "strict physical mechanism constraints". This method breaks through the "either-or" driving mode of traditional models, provides a new path for solving key problems such as real-time warning of urban floods and optimization of drainage systems, and has important practical significance for improving urban flood risk management capabilities.

[0070] Through hierarchical decoupling, multi-scale physical constraint embedding, and deep learning technology, the present invention realizes efficient, accurate, and physically consistent urban flood simulation. The specific technical solutions are as follows: I. Core technical framework: Integration of hierarchical decoupling and multi-physical field constraints Decouple the urban hydrological and hydrodynamic system into an "efficient one-dimensional drainage pipe network sub-module" (driven by the Saint-Venant equations) and a "lightweight two-dimensional surface flood module" (data-driven combined with the shallow water equation constraints), and construct a "microscopic-mesoscopic-macroscopic" multi-scale physical constraint supervision system: 1. Microscopic constraints (drainage pipe network layer): Based on the Saint-Venant equations, impose mass and momentum conservation constraints on the pipe network nodes and pipe segments to ensure the water balance at the nodes (inflow = outflow + change in water storage), and avoid non-physical overflow prediction; 2. Mesoscopic constraints (surface flood layer): Based on the two-dimensional shallow water equations, impose joint constraints of momentum conservation and mass conservation on the surface flow field to ensure that the velocity and water depth distributions conform to the physical laws of water flow; 3. Macroscopic constraints (system layer): Through the global water balance equation, constrain the relationship between the total inflow, total outflow, and total water storage of the entire urban water system to ensure macroscopic physical consistency.

[0071] II. Model architecture: Urban flood intelligent physical prediction model (IPPM4UF), as Figure 2 shown, construct a three-level architecture of "decoupling-constraint-prediction", including three core modules: 1. Decoupled physical module (microscopic + mesoscopic) One-dimensional drainage pipe network sub-module: In the simulation of urban drainage pipe networks, the one-dimensional Saint-Venant equations are the basic equations describing unsteady flow in pipes, consisting of the continuity equation and the momentum equation, as follows: The continuity equation and the momentum equation are respectively expressed as:

[0072]

[0073]

[0074] Among them, A : Cross-sectional area (m 2 ), related to the pipe fullness, pipe diameter, etc.; t : Time (s); Q : Cross-sectional flow rate (m 3 / s), Q = uA , u is the cross-sectional average velocity; x : Coordinate along the pipe axis direction (m); q l : Lateral inflow rate per unit length of the pipe (m2 / s), such as rainwater inflow or leakage. h : Water depth of the cross-section (m), reflecting the water level height in the pipeline; g : Acceleration due to gravity (m / s 2 ); s 0: Pipeline bottom slope (dimensionless); s f : Friction slope (dimensionless), reflecting the energy loss caused by water flow resistance; u l : Velocity of the lateral inflow (m / s).

[0075] The continuity equation ensures the conservation of water mass in the pipeline, that is, the change in cross-sectional water volume per unit time is equal to the sum of the inflow and outflow flow differences and the lateral inflow; the momentum equation considers the effects of gravity, frictional resistance, and lateral inflow on the water flow movement, and describes the change of flow rate over time and space.

[0076] Two-dimensional surface flood sub-module: In urban surface flood simulation, the two-dimensional shallow water equations describe the water flow movement state and characteristics of the urban surface during floods, presenting the diffusion of floods in the horizontal direction, the change in flow velocity, and the water depth distribution in different regions. The present invention adopts a fully distributed hydrological calculation method, coupling the runoff generation process into the right side of equations such as the continuity equation to express additional source terms.

[0077]

[0078]

[0079]

[0080] Among them, t represents time; x and y represent spatial coordinates; h is the water depth; z is the water surface height; u and v are the flow velocities along x and y directions respectively; | V | is the magnitude of the flow velocity, that is ; r is the rainfall intensity; f is the infiltration intensity; c is the drainage intensity; n is the roughness coefficient; g is the acceleration due to gravity; 2. Multi-physical field constraint supervision module 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 reconstruction of the loss function for micro, meso, and macro physical constraints: (1) Microscopic mass conservation loss (Saint-Venant equation constraint) Core objective: Force the one-dimensional pipe network module to satisfy the residual constraints of the continuity equation and momentum equation of the Saint-Venant equation.

[0081] Loss function design:

[0082]

[0083] Among them, the first term: Based on the residual of the continuity equation, ensure the balance of the water volume change in the cross-section and the lateral inflow per unit time; the second term: Based on the residual of the momentum equation, constrain the physical consistency of the flow rate change with gravity, friction, and lateral inflow momentum transfer; the weight coefficient γ 1. γ 2 Dynamically adjust the balance between data fitting and physical constraints.

[0084] (2) Mesoscopic momentum conservation loss (shallow water equation constraint) Core objective: Force the two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of the shallow water equation.

[0085] Loss function design:

[0086]

[0087] Among them, the x-direction and y-direction momentum equation residuals: respectively constrain the momentum change, pressure gradient, and frictional resistance of the surface water flow in the horizontal direction; the roughness coefficient n Introduce the physical mechanism of surface resistance through the Manning formula; the residual term directly reflects the degree of unfulfilled momentum conservation of the shallow water equation.

[0088] Pure data-driven models (such as UDFM) do not explicitly consider the law of momentum conservation, resulting in distorted simulations of key hydrodynamic characteristics such as water flow direction and velocity distribution. For example, in areas with sudden terrain changes such as overpasses, prediction results that violate the actual water flow direction may occur. The present invention can minimize the momentum equation residual, and the velocity field predicted by the model u , v Strictly satisfies the law of momentum conservation. Especially in complex terrains such as steep slopes and river confluences, it can more accurately simulate the water flow turning and energy dissipation processes.

[0089] (3) Macroscopic water volume balance loss (global conservation constraint) Core objective: Force the overall system to satisfy the global water volume balance, including the water volume exchange between the pipe network and the surface.

[0090] Loss function design:

[0091]

[0092] Among them, the first term: the global integral based on the continuity equation of the two-dimensional shallow water equations, which constrains the overall water balance of the surface runoff-infiltration-drainage process; the second term: based on the input-output flow difference and the total water volume change at the interface between the pipe network and the surface, ensuring the water volume conservation across modules; ΔV total is the change in the total water storage of the system, which is calculated in real time through differentiable programming technology.

[0093] In urban flood simulation, the overall water balance of the entire region is the core indicator to test the physical rationality of the model. Traditional numerical-driven artificial intelligence models have problems with inaccurate water volume exchange across systems, causing the disconnection between local conservation and global balance. As a result, during long-term simulations, the accumulation of errors causes the flood simulation to deviate from the actual situation and diverge. This design can ensure the strict balance between the surface water volume change and the rainfall, infiltration, and drainage processes at any time by minimizing the global integral term.

[0094] (4)Integration of the total loss function L total = L data + L micro + L meso + L macro Optimization strategy: Hierarchical weight allocation: γ 1~ γ 4 adopt an adaptive learning mechanism to dynamically adjust the constraint strength according to the training stage; Physics-guided gradient propagation: directly correct the neural network parameters by the physical equations through the automatic differentiation of the residual terms.

[0095] 3. Artificial intelligence prediction module Rainfall-overflow encoder: Use a recurrent neural network (such as LSTM) to extract the characteristics of the rainfall time series and capture the dynamic effects of rainfall intensity, duration, and spatial distribution; Overflow-flood decoder: Combine data dimensionality reduction techniques, such as Empirical Orthogonal Function (EOF), compress the high-dimensional hydrodynamic data (water depth, flow velocity) into low-dimensional time modes, and map the overflow data to the flood spatio-temporal distribution through a recurrent neural network to achieve sub-second fast prediction; Multi-task learning: Jointly optimize the rainfall-overflow-flood mapping relationship, requiring the decoupled variables of each system to simultaneously satisfy multi-scale physical constraints.

[0096] Example 3 An AI real-time flood simulation method for decoupled supervision of multi-physical fields in urban hydrology, the steps are as follows: 1. Training stage (data-physics collaborative learning) S1: Physical module data generation Input historical rainfall data into the one-dimensional drainage pipe network sub-module to calculate node overflows and water accumulation (microscopic constraint data); input it into the two-dimensional hydrodynamic model (such as MIKE URBAN) to generate the spatio-temporal distribution of surface floods (mesoscopic reference data).

[0097] S2: EOF dimensionality reduction processing Perform empirical orthogonal function analysis on the two-dimensional hydrodynamic data (water depth, flow velocity), and extract the first k principal components ( , M is the number of spatial points), reducing the data dimension.

[0098] S3: Multi-physical field constraint training Through a hierarchical supervision structure architecture (rainfall-overflow encoder-drainage dynamic information supervision-overflow-flood decoder), use a recurrent neural network to learn the rainfall-overflow-flood mapping relationship, and at the same time impose microscopic, mesoscopic, and macroscopic constraint losses to ensure that the model output conforms to physical laws.

[0099] 2. Application stage (real-time flood prediction) S4: Real-time rainfall input and decoupled calculation Input real-time rainfall data into the one-dimensional drainage pipe network sub-module, quickly calculate node overflow data, and retain the engineering scheduling compatibility of the physical model (such as pumping station, gate control).

[0100] S5: Flood spatio-temporal reconstruction The flood decoder maps the overflow data into 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:

[0101]

[0102] Among them, e i is the spatial mode, a i is the predicted time mode.

[0103] The present invention adopts a decoupled hierarchical modeling method, separates the high-consumption two-dimensional hydrodynamic module from the high-efficiency one-dimensional pipe network module, replaces complex physical calculations through deep learning, and takes into account both accuracy and efficiency; The present invention adopts multi-scale physical constraints and for the first time embeds the core constraints of the Saint-Venant equation (microscopic), shallow water equation (mesoscopic), and water balance equation (macroscopic) into deep learning, achieving "verifiable physical mechanisms and interpretable prediction results". The present invention adopts data-physics bidirectional fusion. The physical module provides prior constraints for the data-driven model, and deep learning feeds back to optimize the physical model, forming a closed loop of "mechanism guiding learning and learning correcting the mechanism". The present invention improves the real-time performance and generalization ability. EOF dimensionality reduction and LSTM prediction achieve sub-second calculation and support real-time early warning under complex working conditions such as extreme rainfall and pipe network scheduling.

[0104] For a better understanding of the present invention, the above has been described in detail in conjunction with specific embodiments of the present invention, but it is not a limitation of the present invention. Any simple modification made to the above embodiments based on the technical essence of the present invention still belongs to the scope of the technical solution of the present invention. Each embodiment in this specification focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment.

Claims

1. An AI real-time flood simulation method with decoupled supervision of multiple physical fields in urban hydrology, including obtaining historical rainfall data and real-time rainfall data, a training stage, and an application stage, characterized in that, The training stage includes the following steps: Step 01: Input the historical rainfall data into the one-dimensional drainage network sub-module to calculate the microscopic constraint data, where the microscopic constraint data includes node overflows and water accumulation; Step 02: Input the historical rainfall data into the two-dimensional surface flood sub-module to generate the mesoscopic reference data, where the mesoscopic reference data includes the spatio-temporal distribution of surface floods; Step 03: Perform empirical orthogonal function analysis on the two-dimensional hydrodynamic data, extract the first k principal components, and reduce the data dimension, where , M is the number of spatial points; Step 04: Use the multi-physical field constraint supervision module for constrained training. Through a hierarchical supervision structure architecture, utilize a recurrent neural network to learn the rainfall-overflow-flood mapping relationship, and at the same time impose microscopic, mesoscopic, and macroscopic constraint losses; The application stage includes the following sub-steps: 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 ; 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; ， Among them, X is the original dataset matrix, i is the index value of the modality, e i is the spatial modality, a i is the predicted temporal modality, T is the transpose symbol, is the real number space, M is the size of the time dimension, N is the size of the spatial dimension.

2. The AI real-time flood simulation method for urban hydrological multi-physical field decoupling supervision according to claim 1, characterized in that The one-dimensional drainage network sub-module consists of the continuity equation and the momentum equation, which are respectively expressed as: ， ， Among them, A is the cross-sectional area of flow, t is the time, Q is the cross-sectional flow rate, Q = u c A , u c is the average cross-sectional flow velocity, x is the coordinate along the pipeline axis direction, q l is the lateral inflow rate per unit length of the pipeline, h is the water depth of the cross-sectional area of flow, g is the acceleration due to gravity, s 0 is the pipeline bottom slope, s f is the friction slope, u l is the flow velocity of the lateral inflow.

3. The AI real-time flood simulation method for decoupled supervision of urban hydrological multi-physical fields according to claim 2, characterized in that The two-dimensional surface flood sub-module couples the runoff generation process using a fully distributed hydrological calculation method; ， ， ， Among them, x and y are spatial coordinates, h is the water depth, z is the water surface height, u and v are respectively the flow velocities along x and y directions; | V | is the flow velocity, , r is the rainfall intensity, f is the infiltration intensity, c is the drainage intensity, n is the roughness coefficient, g is the acceleration due to gravity.

4. The AI flood real-time simulation method for urban hydrological multi-physical field decoupling supervision according to claim 3, wherein The multi-physical field constraint supervision module is based on a three-level architecture of decoupling-constraint-prediction, and realizes the deep integration of mechanism and data through hierarchical physical constraints.

5. The AI real-time flood simulation method for urban hydrological multi-physical field decoupling supervision according to claim 4, characterized in that The loss function of the multi-physical field constraint supervision module for microscopic physical constraints is reconstructed as the microscopic mass conservation loss, and the microscopic mass conservation loss forces the one-dimensional pipe network module to satisfy the residual constraints of the continuity equation and the momentum equation of the Saint-Venant equation. The loss function is ， Among them, γ 1 and γ 2 are weight coefficients, E is the mathematical expectation of the loss random variable.

6. The AI real-time flood simulation method with decoupled supervision of urban hydrological multi-physical fields according to claim 5, characterized in that, The loss function of the multi-physical field constraint supervision module for mesoscopic physical constraints is reconstructed as the mesoscopic momentum conservation loss, which forces the two-dimensional surface flood module to satisfy the momentum conservation and mass transport constraints of the shallow water equation. The loss function is ， Among them, γ 3 is the weight coefficient.

7. The AI real-time flood simulation method with decoupled supervision of urban hydrological multi-physical fields according to claim 6, characterized in that, The loss function of the multi-physical field constraint supervision module for macroscopic physical constraints is reconstructed as the macroscopic water balance loss, which forces the overall system to satisfy the global water balance, including the water exchange between the pipe network and the surface. The loss function is ， Among them, γ 4 is the 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 computational domain for flood routing simulation, ΔV total is the change in the total water storage of the system.

8. The AI real-time flood simulation method for decoupled supervision of urban hydrological multi-physical fields according to claim 7, characterized in that, The total loss function is integrated into L total = L data + L micro + L meso + L macro 。 9. The AI real-time flood simulation method for decoupled supervision of urban hydrological multi-physical fields according to claim 8, wherein The hierarchical supervision structure architecture is rainfall-overflow encoder—drainage dynamic information supervision—overflow-flood decoder.

10. The AI real-time flood simulation method for decoupled supervision of urban hydrological multi-physical fields according to claim 9, wherein, The rainfall-overflow encoder uses a recurrent neural network to extract the rainfall time series features and capture the dynamic impacts of rainfall intensity, duration, and spatial distribution.

Images