A neural network-based urban street canyon flow field prediction method and device

Abstract

The application discloses a kind of urban street canyon flow field prediction method and device based on neural network.The method comprises: according to street canyon inclination parameter, adjusting the spatial position of grid node in calculation region, obtain different inclination urban street canyon parameterized deformation grid;Based on the high-fidelity numerical solution of fluid mechanics control equation to complete multiple street canyon working conditions, construct offline flow field database containing multiple physical fields;Through intrinsic orthogonal decomposition, low-dimensional space mode and mode coefficient are extracted, and reduced-order dynamic system is obtained through Galileo projection, and neural network turbulence closure prediction model is trained;In online stage, the turbulence closure term is predicted in real time through state feedback mechanism, and it is coupled with reduced-order dynamic system and low-dimensional space mode base function and is solved, and the optimal mode coefficient of physical self-consistent constraint is obtained, and high-dimensional flow field is reconstructed.The application improves the efficiency and stability of urban street canyon flow field prediction, and is used for urban wind environment rapid evaluation.

Description

Technical Field

[0001] This application relates to the field of numerical simulation and modeling technology, and in particular to a method and apparatus for predicting urban street valley flow fields based on neural networks. Background Technology

[0002] With the continuous development of urban construction, urban street valley wind environment assessment, air quality prediction, and building cluster ventilation optimization have become important components of smart city and digital twin city construction. In these application scenarios, it is often necessary to repeatedly conduct flow field and pollutant diffusion simulations under numerous operating conditions with different building inclination angles, street valley structures, and incoming flow conditions to support rapid analysis and optimization design of the urban microenvironment. However, traditional high-fidelity computational fluid dynamics methods require solving large-scale fluid dynamics control equations, resulting in long computation times, high computing costs, and low efficiency in scanning multiple operating parameters, making it difficult to meet the needs of rapid prediction and real-time analysis.

[0003] Current methods for rapid flow field prediction mainly fall into two categories: order reduction modeling and data-driven methods. One category is the traditional order reduction model based on intrinsic orthogonal decomposition, which achieves rapid solutions by extracting low-dimensional modes of the flow field. However, in complex conditions such as street valleys, shear layers, and recirculation zones, low-dimensional truncation can lead to the loss of high-order turbulent dissipation information, resulting in modal energy drift, solution oscillations, and non-physical divergences, leading to insufficient solution stability. The other category is end-to-end flow field prediction methods based on deep learning. While these methods can improve prediction speed, they are not constrained by the physical equations of fluid dynamics, making it difficult to strictly satisfy mass and momentum conservation. They also exhibit weak generalization ability under parameter variations and are prone to local flow field distortion. Furthermore, when building geometric parameters change, traditional global dynamic meshing methods are prone to causing local mesh stretching, intersections, or even negative volumes, leading to numerical solution divergence and failing to support automated parameter scanning and online prediction.

[0004] Therefore, how to improve the prediction stability under complex turbulent street valley conditions while preserving physical conservation constraints, and at the same time solve the grid distortion problem in the building parameterization process, so as to achieve efficient and stable prediction of urban street valley flow field and pollutant diffusion, is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present application provides a method and apparatus for predicting urban street valley flow fields based on neural networks. This method constructs an adaptive distance-weighted node displacement function based on street valley inclination parameters and the spatial location of grid nodes to continuously reconstruct the geometric structure of urban street valleys with different inclination angles and the numerical simulation grid. Combined with intrinsic orthogonal decomposition modeling, Galilean projection, and machine learning turbulence closure prediction mechanisms, it obtains optimal modal coefficients with physical self-consistent constraints while ensuring the conservation of fluid dynamics laws. This improves the accuracy and stability of street valley flow and pollutant diffusion prediction, achieving rapid and high-precision prediction of urban street valley flow fields. The method and apparatus for predicting urban street valley flow fields based on neural networks provided in this application are implemented as follows: This application provides a method for predicting urban street valley flow fields based on neural networks, including: Obtain the street valley inclination angle parameters, and adjust the spatial position of the grid nodes in the calculation area according to the street valley inclination angle parameters to obtain the parameterized deformable grid of urban street valleys with different inclination angles; Numerical solution processing of the parameterized deformable mesh and multiple sets of street valley working parameters is performed based on the fluid dynamics control equations to obtain an offline flow field database including multi-physics field data. The multiphysics field data in the offline flow field database are subjected to intrinsic orthogonal decomposition to reduce dimensionality, resulting in low-dimensional spatial modal basis functions and modal coefficients; Based on the fluid dynamics control equations, Galerkin projection and closed-loop model training are performed on the low-dimensional spatial modal basis functions, modal coefficients, and turbulence data to obtain a reduced-order dynamic system and a neural network turbulence closed-loop prediction model. Obtain the target geometric parameters corresponding to the target working condition to be predicted, and input the target working condition geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term. The target geometric parameters include the target street valley inclination parameter. The reduced-order dynamic system, the real-time low-dimensional turbulence closure term, and the low-dimensional spatial modal basis function are coupled and solved, and the high-dimensional flow field is reconstructed to obtain the high-dimensional flow field of the target working condition.

[0006] In some embodiments, the numerical solution process based on the fluid dynamics control equations to solve the parameterized deformable mesh and multiple sets of street valley working parameters yields an offline flow field database including multiphysics data, comprising: The building tilt angle is sampled and processed to obtain a set of street valley working condition parameters; For the parameterized deformable mesh corresponding to each set of street valley working parameters in the street valley working parameter set, steady-state numerical solution processing is performed based on the Reynolds-averaged Navier-Stokes equation and the pollutant transport control equation to obtain full-order flow field data, which includes velocity field, pressure field, turbulent viscosity field and pollutant concentration field. The full-order flow field data and corresponding grid data are classified, stored, and shared with the grid topology to obtain an offline flow field database.

[0007] In some embodiments, the step of performing intrinsic orthogonal decomposition and dimensionality reduction processing on the multiphysics field data in the offline flow field database to obtain low-dimensional spatial modal basis functions and modal coefficients includes: The multi-physics field data in the offline flow field database are processed to construct a snapshot matrix, thereby obtaining a high-dimensional flow field snapshot matrix corresponding to each physical field; The high-dimensional flow field snapshot matrix is ​​subjected to eigenorthogonal decomposition and truncated according to a preset modal energy ratio threshold to obtain the low-dimensional spatial modal basis functions corresponding to each physical field. Modal projection processing is performed on the full-order flow field data and the low-dimensional spatial modal basis functions to obtain the modal coefficients corresponding to each physical field.

[0008] In some embodiments, the step of performing Galilean projection and closure model training on the low-dimensional spatial modal basis functions, modal coefficients, and turbulence data based on the fluid dynamics control equations to obtain a reduced-order dynamic system and a neural network turbulence closure prediction model includes: Based on the fluid dynamics control equations, the low-dimensional space modal basis functions and modal coefficients are subjected to Galerkin projection to obtain a reduced-order dynamic system; Training samples were constructed and network training was performed on the street valley working parameters, the modal coefficients corresponding to the velocity field, and the turbulent viscosity field in the full-order flow field data to obtain a neural network turbulence closure prediction model. Establish the correspondence between the low-dimensional turbulence closure term of the reduced-order dynamic system and the low-dimensional turbulence viscosity modal coefficients output by the neural network turbulence closure prediction model, complete the mapping matching process, and obtain the reduced-order dynamic system and the neural network turbulence closure prediction model that can be coupled.

[0009] In some embodiments, the step of obtaining the target geometric parameters corresponding to the target working condition to be predicted, and inputting the target working condition geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing, yields a real-time low-dimensional turbulence closure term. The target geometric parameters include target street valley inclination parameters, including: The target working condition geometric parameters and the initial low-dimensional flow field modal coefficients are subjected to feature integration processing to obtain neural network input feature data; The neural network input feature data is input into the neural network turbulence closure prediction model for prediction calculation processing to obtain low-dimensional turbulence viscosity modal coefficients; The low-dimensional turbulent viscosity modal coefficients are subjected to closed-term transformation to obtain real-time low-dimensional turbulent closed terms.

[0010] In some embodiments, the coupled solution of the reduced-order dynamic system, the real-time low-dimensional turbulence closure term, and the low-dimensional spatial modal basis function, along with high-dimensional flow field reconstruction processing, to obtain the high-dimensional flow field of the target condition includes: The reduced-order dynamic system and the real-time low-dimensional turbulence closed term are coupled and iteratively solved to obtain the convergent low-dimensional modal coefficients of each physical field. The low-dimensional modal coefficients of each converged physical field and the low-dimensional spatial modal basis functions are subjected to high-dimensional flow field reconstruction processing to obtain high-dimensional distribution data of each physical field; The high-dimensional distribution data of each physical field are classified and integrated to obtain the high-dimensional flow field of the target working condition.

[0011] This application provides an embodiment of an urban street valley flow field prediction device based on a neural network, comprising: The processing module is used to obtain the street valley inclination angle parameters, and adjust the spatial position of the grid nodes in the calculation area according to the street valley inclination angle parameters to obtain the parameterized deformable grid of urban street valleys with different inclination angles; The processing module is also used to perform numerical solution processing on the parameterized deformable mesh and multiple sets of street valley working parameters based on the fluid dynamics control equations to obtain an offline flow field database including multi-physics field data. The processing module is also used to perform intrinsic orthogonal decomposition and dimensionality reduction processing on the multiphysics field data in the offline flow field database to obtain low-dimensional spatial modal basis functions and modal coefficients; The training module is used to perform Galilean projection and closed-loop model training on the low-dimensional spatial modal basis functions, modal coefficients and turbulence data based on the fluid dynamics control equations, so as to obtain a reduced-order dynamic system and a neural network turbulence closed-loop prediction model. The input module is used to obtain the target geometric parameters corresponding to the target working condition to be predicted, and input the target working condition geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term. The target geometric parameters include the target street valley inclination parameter. The processing module is also used to couple and solve the reduced-order dynamic system, the real-time low-dimensional turbulence closure term and the low-dimensional spatial modal basis function, and to reconstruct the high-dimensional flow field to obtain the high-dimensional flow field of the target working condition.

[0012] The computer device provided in this application includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the method described in this application.

[0013] The computer-readable storage medium provided in this application embodiment stores a computer program thereon, which, when executed by a processor, implements the method provided in this application embodiment.

[0014] This application provides a method and apparatus for predicting urban street valley flow fields based on neural networks. By constructing an adaptive distance-weighted node displacement function based on street valley inclination parameters and grid node spatial positions, it achieves continuous reconstruction of the geometric structure and computational grid of urban street valleys with different inclination angles, avoiding the grid distortion problem that occurs in traditional parameterization processes. An offline flow field database is built based on high-fidelity numerical simulation. Low-dimensional spatial modes are extracted through intrinsic orthogonal decomposition, and a reduced-order model is established. Furthermore, turbulent viscosity modal coefficients are predicted using building geometric parameters and velocity field modal coefficients. In the online stage, turbulent closure terms are predicted in real time through state feedback and coupled iteratively with the reduced-order model to finally reconstruct the high-dimensional flow field under the target working condition. This achieves the organic integration of parametric modeling, reduced-order solution, machine learning, and physical constraint optimization, significantly improving prediction efficiency and accuracy while ensuring the conservation of fluid dynamics laws, realizing rapid and high-precision prediction of urban street valley flow fields, and solving the technical problems mentioned in the background art. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram illustrating the implementation process of a neural network-based urban street valley flow field prediction method provided in this application embodiment; Figure 2 A schematic diagram illustrating the implementation process of acquiring an offline flow field database, provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a neural network-based urban street valley flow field prediction device provided in this application embodiment; Figure 4 A local algebraic mesh deformation diagram provided in an embodiment of this application; Figure 5 A result comparison diagram provided for an embodiment of this application; Figure 6 This is a comparison chart of online prediction time provided for an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0019] Figure 1 This is a schematic diagram illustrating the implementation process of a neural network-based urban street valley flow field prediction method provided in this application embodiment, including steps 101 to 106. Wherein, Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order for a neural network-based urban street valley flow field prediction method. The execution order can be adjusted to achieve the desired final result. Figure 1 The steps shown can be performed in parallel or in reverse order.

[0020] Step 101: Obtain the street valley inclination angle parameters, and adjust the spatial position of the grid nodes in the calculation area according to the street valley inclination angle parameters to obtain the parameterized deformable grid of urban street valleys with different inclination angles.

[0021] In this embodiment, an initial structured computational domain for the urban street valley is first established, and a corresponding structured orthogonal computational mesh is generated. A local bounding box is delineated around the building area to be deformed. Displacement updates are performed only on the mesh nodes inside the bounding box, while the mesh outside the bounding box retains its original shape. Let the bounding box region be:

[0022] For any grid node within the computational domain:

[0023] Define the building tilt angle parameter as follows:

[0024] Let the reference coordinates of the original boundary of the building in the x direction be... The building height is H; the slope angle of the street valley is... At that time, the offset of the building's top boundary along the x-direction is H·tan The target x-coordinate corresponding to the building boundary is The displacement of the building boundary nodes is thus derived: =

[0025] The above-mentioned relationship between the adaptive distance weighting function and the nodal displacement calculation of the building boundary node displacement is called the adaptive distance weighting nodal displacement function.

[0026] The distance from the node to the building boundary is:

[0027] Construct an adaptive distance weighting function:

[0028] in: Represents the planar coordinates of the i-th grid node within the computational domain. These are the spatial coordinates of the node in the x and y directions, respectively. This is the attenuation coefficient, used to control the rate attenuation of node displacement with respect to the distance from the node to the building boundary; it is a positive constant. This represents the maximum impact within the local influence domain.

[0029] The nodal displacement is then expressed as:

[0030] The updated node coordinates are:

[0031]

[0032] By using a distance-weighted mechanism, nodes closer to the building boundary are given larger displacements, while the displacements of nodes farther away from the building boundary gradually decrease, thereby achieving continuous and smooth deformation of the mesh nodes.

[0033] Furthermore, to avoid local mesh distortion when the tilt angle is large, mesh quality constraints can be applied to the nodal displacements:

[0034] in: This is the maximum displacement threshold allowed by the mesh quality.

[0035] The displacement of nodes within the bounding box is allocated using the aforementioned bilinear spatial decay rule. The displacement of a node gradually decreases as its spatial distance from the center of the bounding box increases, thus enabling a smooth and continuous transition between the deformed and undeformed mesh regions.

[0036] The entire mesh deformation process is completed through explicit algebraic calculations, eliminating the need to solve the Laplace equation, spring smoothing equation, or global radial basis function dynamic mesh control equations. This effectively avoids local mesh stretching, mesh line intersections, and negative volume problems caused by traditional global dynamic meshing. After completing the local mesh update, the local mesh is updated using `mesh.movePoints()`, and a new `polyMesh` file is output for subsequent high-fidelity CFD (Computational Fluid Dynamics) solutions. Furthermore, the topological connectivity of the mesh elements remains unchanged under different building parameter conditions, with only the node coordinates differing. Therefore, a shared access method is used for the common mesh topology connection file, storing only the node coordinates and flow field result files differentiated for each condition separately, reducing storage resource consumption during large-scale parameter scans.

[0037] The deformation effect of the local bounding box algebraic mesh corresponding to this application is as follows: Figure 4 As shown, the diagram includes the original mesh, the local deformation results, and the selected area. Smooth and continuous deformation of the local mesh is achieved through bilinear attenuation displacement.

[0038] Step 102: Based on the fluid dynamics control equations, the parameterized deformable mesh and multiple sets of street valley working parameters are numerically solved to obtain an offline flow field database including multi-physics data.

[0039] In this embodiment of the application, the building tilt angle is discretely sampled within a preset parameter space to generate multiple sets of different street valley working condition parameters; for each working condition, a parameterized mesh with local deformation is used to perform steady-state high-fidelity numerical solution based on the Reynolds-averaged Navier-Stokes fluid dynamics control equations and a turbulence model.

[0040] For each set of street valley working parameters, the adaptive distance-weighted node displacement function in step 101 is called according to the corresponding building tilt angle parameters to perform continuous deformation processing on the initial structured mesh and generate the corresponding parametric deformable mesh.

[0041] Because a unified mesh topology is maintained under different operating conditions, with only the spatial coordinates of the nodes changing, all operating conditions share the same topological connection relationship, thus ensuring the consistency of data in the subsequent intrinsic orthogonal decomposition and order reduction modeling processes.

[0042] For incompressible flow, the continuity equation satisfies:

[0043] The corresponding RANS (Reynolds-Averaged Navier-Stokes) momentum equation is:

[0044] in: For velocity components; For pressure; For fluid density; Dynamic viscosity; For the Reynolds stress tensor components.

[0045] To solve for the pollutant concentration field, the steady-state pollutant transport control equation is adopted:

[0046] in, Let be the time-averaged velocity component in the j-th spatial direction, and be the average flow velocity under the Reynolds average (RANS) framework; Let be the concentration gradient of the pollutant in the j-th spatial direction, i.e., the rate of change of concentration along the spatial direction; The j-th spatial coordinate direction; The concentration of pollutants; The molecular diffusion coefficient; Turbulent viscosity; For turbulent Schmidt number; This refers to the source of pollutants.

[0047] The full-order flow field data of velocity field, pressure field, turbulent viscosity field and pollutant concentration field under each working condition are calculated; the full-order flow field data of all working conditions are organized and stored together with the corresponding grid data to form a structured offline flow field database.

[0048] To reduce the pressure of storing a large amount of repetitive mesh data during the construction of a large-scale offline database, this embodiment adopts a mesh topology sharing mechanism based on soft links: under different building parameter conditions, although the coordinates of nodes inside the local bounding box will undergo local deformation with the change of building inclination angle, the overall mesh topology connection relationship remains unchanged. That is, the topology files such as owner, neighborhood, faces, and boundary between mesh cells do not need to be generated repeatedly; only the locally deformed points node coordinate files and the corresponding flow field result files are saved for different conditions. For the polyMesh topology structure files that remain unchanged, soft links are used for sharing and calling.

[0049] Step 103: Perform intrinsic orthogonal decomposition to reduce the dimensionality of the multiphysics data in the offline flow field database to obtain low-dimensional spatial modal basis functions and modal coefficients.

[0050] In this embodiment, for various physical field data such as velocity field, pressure field, turbulent viscosity field, and pollutant concentration field in the offline flow field database, corresponding high-dimensional flow field snapshot matrices are constructed respectively; intrinsic orthogonal decomposition is performed on each snapshot matrix, and the first few main modes are selected and retained according to the modal energy ratio to obtain the low-dimensional space modal basis functions corresponding to each physical field; the full-order flow field data of each physical field are projected onto the corresponding low-dimensional space modal basis functions, and the low-dimensional modal coefficients corresponding to each physical field are calculated, thus completing the mapping transformation from high-dimensional flow field to low-dimensional modal space.

[0051] For the velocity field, its low-dimensional POD (Proper Orthogonal Decomposition) expansion is as follows:

[0052] in: The flow field is a mean flow field; These are the modal basis functions in the POD space; These are the corresponding modal coefficients; To preserve the modal order.

[0053] Similarly, the pressure field, pollutant concentration field, and turbulent concentration field are respectively expressed as:

[0054]

[0055]

[0056] in: The pressure modal coefficient; The pollutant concentration modal coefficient; These are the turbulent viscosity modal coefficients; These are the basis functions for the corresponding spatial modes.

[0057] Through the above processing, the low-dimensional spatial modal basis functions and modal coefficients corresponding to the velocity field, pressure field, turbulent viscosity field and pollutant concentration field are obtained, providing basic data for subsequent order reduction modeling and machine learning closure prediction.

[0058] Step 104: Based on the fluid dynamics control equations, Galilean projection and closed-loop model training are performed on the low-dimensional space modal basis functions, modal coefficients and turbulence data to obtain the reduced-order dynamic system and the neural network turbulence closed-loop prediction model.

[0059] In this embodiment, the low-dimensional spatial modal basis functions and modal coefficients of the velocity field are substituted into the fluid dynamics control equations, and the high-dimensional control equations are mapped to the low-dimensional modal space using the Galerkin projection method, resulting in a reduced-order dynamic system containing constant terms, linear terms, nonlinear convection terms, and low-dimensional turbulence closure terms.

[0060] Substituting the POD expansion into the RANS governing equations and employing the Galerkin projection method, a low-dimensional reduced-order dynamic system can be obtained:

[0061] in: For constant terms; It is a linear term; It is a nonlinear convection term; It is a low-dimensional turbulent closed term generated after truncation of higher-order modes.

[0062] The loss of high-order turbulence dissipation information during the mode truncation process corresponding to the low-dimensional turbulence closure term is the core factor causing the traditional reduced-order model to exhibit solution oscillations and non-physical divergences under complex turbulent conditions.

[0063] Furthermore, training samples were constructed using offline high-fidelity flow field data, with building geometry parameters and velocity field modal coefficients as inputs and turbulent viscosity modal coefficients as outputs, to train a neural network turbulence closure prediction model.

[0064] Specifically, the neural network closure term prediction process can be expressed as follows:

[0065] in: These are the building's geometric parameters; These are the current low-dimensional velocity mode coefficients; This represents a nonlinear mapping relationship within a neural network.

[0066] Network input state is defined as:

[0067] The network output is:

[0068] in: Indicates building geometric parameters; Represents the low-dimensional modal coefficients of the velocity field; This represents the turbulent viscosity modal coefficient.

[0069] Step 105: Obtain the target geometric parameters corresponding to the target working condition to be predicted. Input the target working condition geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term.

[0070] In this embodiment, for the target street valley condition to be predicted, local bounding box algebraic mesh deformation processing is first performed to obtain the parameterized mesh corresponding to the target condition, and the modal coefficients of the low-dimensional flow field are initialized. The building geometry parameters of the target condition and the current low-dimensional flow field modal coefficients are used as inputs and fed into the trained neural network turbulence closure prediction model. The low-dimensional turbulence viscosity modal coefficients are calculated in real time through state feedback. The low-dimensional turbulence viscosity modal coefficients are substituted into the turbulence viscosity field POD expansion for high-dimensional reconstruction to obtain the real-time turbulence viscosity field. The turbulence diffusion term in the reduced-order model is calculated based on the real-time turbulence viscosity field to form the real-time low-dimensional turbulence closure term.

[0071] In the online prediction phase, for new building conditions, the system first performs local bounding box algebraic mesh deformation, then initializes the low-dimensional velocity and pressure field modal coefficients; then, it inputs the current low-dimensional velocity state and geometric parameters into the neural network to predict the low-dimensional turbulent viscosity modal coefficients in real time, and further obtains the effective diffusion coefficient through turbulent Schmidt number transformation.

[0072] Step 106 involves coupling the reduced-order dynamic system, the real-time low-dimensional turbulence closure term, and the low-dimensional spatial modal basis function to obtain the high-dimensional flow field of the target working condition.

[0073] In this embodiment of the application, the real-time low-dimensional turbulence closure term obtained in step 105 is substituted into the reduced-order dynamics system and coupled with the reduced-order flow field control equation for iterative solution. The low-dimensional modal coefficients of the velocity field and pressure field are updated in real time until the convergence condition is met, and the converged low-dimensional modal coefficients of the velocity field and pressure field are obtained.

[0074] Subsequently, high-dimensional reconstruction is performed based on the converged low-dimensional modal coefficients of the velocity field and the pressure field, combined with the corresponding POD space modal basis functions, to obtain the high-dimensional velocity field and high-dimensional pressure field corresponding to the target working condition.

[0075] After obtaining the high-dimensional flow field and the corresponding low-dimensional velocity mode coefficients, the low-dimensional velocity mode coefficients and the real-time low-dimensional turbulence closure term obtained in step 105 are substituted into the pollutant transport control equation for solution to obtain the low-dimensional pollutant mode coefficients; further, high-dimensional reconstruction is performed by combining the pollutant POD spatial mode basis functions to obtain the high-dimensional pollutant concentration field corresponding to the target operating condition.

[0076] Finally, the high-dimensional velocity field, high-dimensional pressure field, and high-dimensional pollutant concentration field are classified and integrated to form the high-dimensional flow field of the target working condition.

[0077] The prediction performance of this method is compared with that of high-fidelity CFD results, for example... Figure 5 As shown, (a) is the CFD calculation result, (b) is the prediction result of this application, and (c) is the absolute error between the two. It can be seen that the prediction result of this application is highly consistent with the high-fidelity CFD result, and the prediction accuracy meets the requirements of engineering applications.

[0078] Comparison of online prediction time between traditional high-fidelity CFD methods and this application: Figure 6 As shown, the traditional high-fidelity CFD method requires approximately 210 seconds to solve a single case, while the online prediction method of this invention only requires approximately 13.3 seconds, which can significantly improve prediction efficiency.

[0079] This application's embodiments construct an adaptive distance-weighted node displacement function based on street valley inclination parameters and grid node spatial positions to achieve continuous reconstruction of urban street valley geometry and computational grids at different inclination angles, avoiding the grid distortion problem generated in traditional parametric modeling. An offline flow field database is built based on high-fidelity numerical simulation, and a low-dimensional reduced-order model is established through intrinsic orthogonal decomposition and Galilean projection. Turbulent viscosity modal coefficients are predicted using building geometric parameters and velocity field modal coefficients. In the online stage, the turbulent viscosity field is predicted in real time through state feedback, updating the diffusion term in the reduced-order model and performing coupled iterative solutions to obtain the optimal modal coefficients with physical self-consistent constraints. Finally, a high-dimensional flow field under the target working condition is reconstructed. Compared with traditional high-fidelity CFD methods, this application significantly improves computational efficiency while ensuring prediction accuracy, achieving rapid and high-precision prediction of urban street valley flow fields.

[0080] In the above Figure 1 Based on the above, this application embodiment also provides a schematic diagram of the implementation process for obtaining an offline flow field database. For example... Figure 2 As shown, steps 201 to 203 are included: Step 201: Perform parameter sampling processing on the building tilt angle to obtain the set of street valley working condition parameters.

[0081] In this embodiment, a parameter space for predicting the street valley flow field is predefined, including building inclination angle variables. Within a reasonable range of parameter values, parameters are combined using uniform discrete sampling to generate multiple sets of non-overlapping street valley working condition parameters. For each set of building inclination angle parameters, based on the initial structured mesh, an adaptive distance-weighted node displacement function is constructed according to the building inclination angle parameters and the spatial position of mesh nodes. This function continuously deforms the mesh nodes within the local bounding box region, generating a corresponding parametric deformed mesh. Grid nodes closer to the building boundary receive larger displacement updates, while the displacements of nodes farther from the building boundary gradually decrease, thus ensuring the continuity and smoothness of the mesh deformation process and avoiding local mesh distortion. All working condition parameters together constitute the street valley working condition parameter set.

[0082] Step 202: For each set of street valley working condition parameters in the street valley working condition parameter set, steady-state numerical solution processing is performed based on the Reynolds-averaged Navier-Stokes equation and the pollutant transport control equation to obtain full-order flow field data, which includes velocity field, pressure field, turbulent viscosity field and pollutant concentration field.

[0083] In this embodiment, for each set of street valley working condition parameters in the street valley working condition parameter set, the adaptive distance-weighted node displacement function defined in step 101 is first called according to the corresponding building inclination angle parameter to perform continuous deformation processing on the initial structured grid and generate the corresponding parameterized deformed grid; then, the Reynolds-averaged Navier-Stokes equations are used as the fluid dynamics control equations, and the steady-state numerical solution of the flow and pollutant transport process in the computational domain is performed in combination with the turbulence model. After the iterative residuals converge to the set threshold, the full-order flow field data corresponding to the working condition is output. The full-order flow field data specifically includes the complete spatial distribution data of four types of physical fields: velocity field, pressure field, turbulent viscosity field, and pollutant concentration field.

[0084] Step 203: Classify and store the full-order flow field data and the corresponding grid data, and perform grid topology sharing processing to obtain an offline flow field database.

[0085] In this embodiment, all flow field data for all operating conditions are categorized and archived according to operating condition identifiers, and stored independently. For grid data, since only the spatial coordinates of nodes change under different building parameter operating conditions, while the topological connections between grid cells remain unchanged, all operating conditions share a unified grid topology structure. Based on this characteristic, a shared access method is used for the common grid topology structure file, storing only the differentiated node coordinates and flow field result files for each operating condition separately. After all flow field data and corresponding grid data are uniformly organized, a structured offline flow field database is formed.

[0086] This application embodiment, by sampling the parameters of building tilt angle, can systematically cover typical working conditions under the target scenario, providing a sufficient and reasonably distributed sample basis for subsequent order reduction modeling, and ensuring the parameter generalization ability of the order reduction model; based on the parameterized deformable mesh, steady-state numerical solution is carried out to obtain high-precision full-order flow field data, providing a reliable data benchmark for order reduction model training and accuracy verification; adopting a classified storage and mesh topology sharing mechanism, only storing differentiated node coordinates and flow field data, and reusing general mesh topology files, effectively reducing the storage resource occupation of multi-working-condition data and improving the management and calling efficiency of large-scale working-condition databases.

[0087] In some embodiments, the multi-physics data in the offline flow field database is subjected to intrinsic orthogonal decomposition to reduce dimensionality and obtain low-dimensional spatial modal basis functions and modal coefficients, including: performing snapshot matrix construction processing on the multi-physics data in the offline flow field database to obtain high-dimensional flow field snapshot matrices corresponding to each physical field.

[0088] Specifically, for the four types of physical fields in the offline flow field database—velocity field, pressure field, turbulent viscosity field, and pollutant concentration field—the full-order flow field data corresponding to each working condition are vectorized and organized, with the flow field data of each working condition serving as an independent snapshot vector. The snapshot vectors of all working conditions under the same physical field are arranged sequentially according to the working condition order to construct the high-dimensional flow field snapshot matrix corresponding to each type of physical field.

[0089] Furthermore, the high-dimensional flow field snapshot matrix is ​​subjected to eigenorthogonal decomposition, and the order is truncated according to the preset modal energy ratio threshold to obtain the low-dimensional spatial modal basis functions corresponding to each physical field.

[0090] Specifically, an intrinsic orthogonal decomposition operation is performed on the high-dimensional flow field snapshot matrix of each type of physical field to obtain a spatial mode sequence and corresponding eigenvalues ​​sorted in descending order of modal energy proportion. According to the preset modal energy retention threshold, the first few major modes that meet the energy proportion requirements are retained, and the higher-order minor modes with too low energy proportion are removed, finally obtaining the low-dimensional spatial mode basis function corresponding to the physical field.

[0091] Furthermore, modal projection processing is performed on the full-order flow field data and the low-dimensional spatial modal basis functions to obtain the modal coefficients corresponding to each physical field.

[0092] Specifically, the full-order flow field data of each working condition in the offline flow field database are projected onto the low-dimensional spatial modal basis functions of the corresponding physical field. The modal coefficients corresponding to each order mode under each working condition are obtained through projection operation. Thus, the high-dimensional flow field can be approximated by a linear combination of the low-dimensional modal coefficients and the corresponding spatial modal basis functions, realizing the dimensionality reduction mapping of high-dimensional flow field data to low-dimensional modal space.

[0093] This application's embodiments construct high-dimensional flow field snapshot matrices separately for each physical field, ensuring the independence and data consistency of the dimensionality reduction process for each physical field, thus laying a standardized data foundation for subsequent field-specific order reduction solutions. Through eigenorthogonal decomposition and order truncation based on energy proportion, high-dimensional flow field data can be compressed into a low-dimensional modal space while preserving mainstream field information, significantly reducing the degrees of freedom in the solution, which is the core support for achieving rapid prediction. The low-dimensional modal coefficients obtained through modal projection can linearly reconstruct the original flow field with the spatial modal basis functions, ensuring the reversibility and accuracy controllability of the dimensionality reduction process, and achieving accurate mapping and transformation between high-dimensional flow fields and low-dimensional modal spaces.

[0094] In some embodiments, a reduced-order dynamic system and a neural network turbulence closure prediction model are obtained by performing Galerkin projection on low-dimensional spatial modal basis functions, modal coefficients and turbulence data based on fluid dynamics control equations and training a closed model. This includes: performing Galerkin projection on low-dimensional spatial modal basis functions and modal coefficients based on fluid dynamics control equations to obtain a reduced-order dynamic system.

[0095] Specifically, the low-dimensional modal basis functions and modal coefficients of the velocity field are substituted into the Reynolds-averaged Navier-Stokes hydrodynamic control equations. Using the low-dimensional modal basis functions as weights, Galerkin projection operations are performed on the control equations, mapping the high-dimensional partial differential control equations in the full-grid space to the low-dimensional modal space. This yields a reduced-order dynamic system containing constant terms, linear terms, nonlinear convection terms, and low-dimensional turbulence closure terms. The low-dimensional turbulence closure terms correspond to the high-order turbulence dissipation information lost during the low-dimensional modal truncation process, and are the core factor causing solution oscillations and non-physical divergences in traditional reduced-order models under complex turbulent conditions.

[0096] Furthermore, training samples were constructed and network training was performed on the modal coefficients corresponding to the street valley working parameters and velocity fields of each group, as well as the turbulent viscosity field in the full-order flow field data, to obtain a neural network turbulence closure prediction model.

[0097] Specifically, the geometric parameters of each group of street valley working conditions and the low-dimensional modal coefficients of the velocity field under the corresponding working conditions are used as input features. The low-dimensional modal coefficients obtained by dimensionality reduction of the turbulent viscosity field in the full-order flow field data are used as training labels, and a training sample set is constructed by matching each group. A neural network prediction model is built, with the mean squared error as the loss function. The network weight parameters are iteratively updated through the error backpropagation algorithm until the loss function converges to a preset threshold, thus completing the offline training of the neural network turbulent closure prediction model. The trained neural network only outputs the low-dimensional modal coefficients corresponding to the turbulent viscosity and does not directly predict the complete full-dimensional flow field data such as the velocity field and pressure field.

[0098] Furthermore, a corresponding transformation relationship is established between the low-dimensional turbulence closure term of the reduced-order dynamic system and the low-dimensional turbulence viscosity modal coefficients output by the neural network turbulence closure prediction model. After completing the mapping matching process, a reduced-order dynamic system and a neural network turbulence closure prediction model that can be coupled are obtained.

[0099] Specifically, a corresponding conversion relationship is established between the low-dimensional turbulent viscosity modal coefficients output by the neural network and the low-dimensional turbulent closure terms in the reduced-order dynamic system. This allows the prediction results of the neural network to be directly substituted into the reduced-order dynamic system for iterative solution, ultimately resulting in a reduced-order dynamic system and neural network turbulent closure prediction model that can be coupled online.

[0100] This application's embodiments utilize Galerkin projection based on the fluid dynamics control equations to transform high-dimensional partial differential equations into a set of low-dimensional ordinary differential equations. This ensures that the reduced-order system strictly adheres to physical laws such as the conservation of mass and momentum, avoiding the physical distortion problems that easily occur in purely data-driven models. By training a neural network with operating condition geometric parameters and velocity field modal coefficients as inputs and turbulent viscosity modal coefficients as labels, it can accurately fit the high-order turbulent dissipation effects lost due to modal truncation, thus compensating for the accuracy shortcomings of traditional POD reduced-order models from a physical mechanism perspective. Through mapping and matching between closed terms and network outputs, seamless integration between the neural network turbulent closed model and the reduced-order dynamic system is achieved, allowing them to collaboratively participate in the solution, forming a reduced-order solution framework that combines physical constraints and data correction.

[0101] In some embodiments, the target geometric parameters corresponding to the target operating condition to be predicted are obtained, and the target operating condition geometric parameters and the initial low-dimensional flow field modal coefficients are input into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain real-time low-dimensional turbulence closure terms. This includes: performing feature integration processing on the target operating condition geometric parameters and the initial low-dimensional flow field modal coefficients to obtain neural network input feature data.

[0102] Specifically, for the target street valley condition to be predicted, its corresponding building geometric parameters are defined, and the low-dimensional flow field modal coefficients of the target condition are initialized as initial flow field state data. The geometric parameters of the target condition and the initial low-dimensional flow field modal coefficients are unified in format and spliced ​​together to obtain standardized neural network input feature data, forming a state feedback input structure that combines geometric parameters and flow field state.

[0103] Furthermore, the neural network input feature data is input into the neural network turbulence closure prediction model for prediction calculation processing to obtain low-dimensional turbulence viscosity modal coefficients.

[0104] Specifically, the neural network input feature data is input into a pre-trained neural network turbulence closure prediction model. The feature mapping is calculated through the forward operation of the model to obtain the corresponding low-dimensional turbulence viscosity modal coefficients. This prediction process only outputs low-dimensional modal coefficients related to turbulence, with low computational overhead, and can achieve rapid response in online scenarios.

[0105] Furthermore, the low-dimensional turbulent viscosity modal coefficients are subjected to closed-term transformation to obtain real-time low-dimensional turbulent closed terms.

[0106] Specifically, based on the predetermined mapping transformation relationship, the predicted low-dimensional turbulent viscosity mode coefficients are converted into real-time low-dimensional turbulent closure terms that are adapted to the reduced-order dynamic system. This is used to supplement the high-order turbulent dissipation information lost during the low-dimensional mode truncation process and to provide closure conditions for the subsequent iterative solution of the reduced-order dynamic system.

[0107] This application embodiment integrates the target working condition geometric parameters and the current flow field modal coefficients as inputs to form a state feedback prediction structure, enabling the closure term to be dynamically updated with the iterative state of the flow field, adapting to the changes in turbulent characteristics during the flow field evolution process; the neural network only outputs low-dimensional turbulent viscosity modal coefficients, with low computational dimensionality and fast forward inference speed, which can meet the real-time calling requirements in the online solution process without significantly increasing the overall computation time; the result after the closure term transformation can be directly substituted into the reduced-order dynamic system, ensuring the continuity of the online prediction and reduced-order solution process, and realizing the rapid and accurate injection of turbulent closure effect.

[0108] In some embodiments, the reduced-order dynamic system, the real-time low-dimensional turbulence closure term and the low-dimensional spatial modal basis function are coupled and solved, and the high-dimensional flow field is reconstructed to obtain the high-dimensional flow field of the target working condition. This includes: coupling the reduced-order dynamic system and the real-time low-dimensional turbulence closure term to solve iteratively and obtain the converged low-dimensional modal coefficients of each physical field.

[0109] Furthermore, the low-dimensional modal coefficients and low-dimensional spatial modal basis functions of each converged physical field are subjected to high-dimensional flow field reconstruction processing to obtain high-dimensional distribution data of each physical field.

[0110] Specifically, the low-dimensional modal coefficients of each physical field obtained by convergence are linearly combined with the low-dimensional spatial modal basis functions of the corresponding physical field, and mapped node by node back to the full grid space of the original computational domain to obtain high-dimensional flow field data, thus completing the inverse mapping transformation from low-dimensional modal space to high-dimensional physical space.

[0111] Furthermore, the high-dimensional distribution data of each physical field are classified and integrated to obtain the high-dimensional flow field of the target working condition.

[0112] Specifically, high-dimensional distributed data such as velocity field and pressure field are integrated into complete high-dimensional flow field data, and structured and archived according to spatial coordinates and physical quantity categories to finally obtain the complete high-dimensional flow field under the target working condition.

[0113] This application's embodiments improve the prediction accuracy of turbulent closure terms under street valley conditions by embedding real-time low-dimensional turbulent closure terms predicted by neural networks into a reduced-order dynamic system for coupled iterative solution. This enhances the accuracy and convergence stability of the reduced-order flow field solution. High-dimensional reconstruction is achieved through a linear combination of low-dimensional modal coefficients and spatial modal basis functions, restoring a spatial flow field distribution consistent with the full-order solution resolution, balancing solution efficiency and spatial integrity. Standardized high-dimensional flow field results are output after classification and integration, enabling direct integration with subsequent engineering applications such as urban wind environment assessment and air quality analysis, thus improving the method's engineering applicability.

[0114] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in this embodiment is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the methods shown in this embodiment or the accompanying drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0115] like Figure 3 As shown in the illustration, this application also provides an urban street valley flow field prediction device 300 based on a neural network. The device includes: The processing module 301 is used to obtain the street valley inclination angle parameters, and adjust the spatial position of the grid nodes in the calculation area according to the street valley inclination angle parameters to obtain the parameterized deformable grid of urban street valleys with different inclination angles.

[0116] The processing module 301 is also used to perform numerical solution processing on the parameterized deformable mesh and multiple sets of street valley working parameters based on the fluid dynamics control equations, so as to obtain an offline flow field database including multi-physics field data.

[0117] The processing module 301 is also used to perform intrinsic orthogonal decomposition and dimensionality reduction processing on the multiphysics field data in the offline flow field database to obtain low-dimensional spatial modal basis functions and modal coefficients.

[0118] Training module 302 is used to perform Galerkin projection and closed-loop model training on low-dimensional spatial modal basis functions, modal coefficients and turbulence data based on fluid dynamics control equations, so as to obtain a reduced-order dynamic system and a neural network turbulence closed-loop prediction model.

[0119] The input module 303 is used to obtain the target geometric parameters corresponding to the target working condition to be predicted, and input the target working condition geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term. The target geometric parameters include the target street valley inclination parameters.

[0120] The processing module 301 is also used to couple the solution of the reduced-order dynamic system, the real-time low-dimensional turbulence closure term and the low-dimensional spatial modal basis function, and to reconstruct the high-dimensional flow field to obtain the high-dimensional flow field of the target working condition.

[0121] Some modules in the apparatus described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0122] The apparatus or module described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. For ease of description, the above apparatus is described by dividing it into various modules according to their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0123] The methods, apparatus, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, such as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of a memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included within it for implementing various functions can also be considered as structures within the hardware component. Alternatively, the device used to implement various functions can be viewed as either a software module that implements the method or a structure within a hardware component.

[0124] This application also provides an apparatus, the apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method described in this application.

[0125] This application also provides a non-volatile computer-readable storage medium storing a computer program or instructions thereon, which, when executed, enables the method described in this application embodiment to be implemented.

[0126] Furthermore, in the various embodiments of the present invention, each functional module can be integrated into a processing module, or each module can exist independently, or two or more modules can be integrated into a single module.

[0127] The aforementioned storage media include, but are not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions.

[0128] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0129] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0130] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A neural network-based urban street canyon flow field prediction method, characterized in that, include: Obtain the street valley inclination angle parameters, and adjust the spatial position of the grid nodes in the calculation area according to the street valley inclination angle parameters to obtain the parameterized deformable grid of urban street valleys with different inclination angles; Numerical solution processing of the parameterized deformable mesh and multiple sets of street valley working parameters is performed based on the fluid dynamics control equations to obtain an offline flow field database including multi-physics field data. The multiphysics field data in the offline flow field database are subjected to intrinsic orthogonal decomposition to reduce dimensionality, resulting in low-dimensional spatial modal basis functions and modal coefficients; Based on the fluid dynamics control equations, Galerkin projection and closed-loop model training are performed on the low-dimensional spatial modal basis functions, modal coefficients, and turbulence data to obtain a reduced-order dynamic system and a neural network turbulence closed-loop prediction model. Obtain the target geometric parameters corresponding to the target working condition to be predicted, and input the target geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term. The target geometric parameters include the target street valley inclination parameter. The reduced-order dynamic system, the real-time low-dimensional turbulence closure term and the low-dimensional spatial modal basis function are coupled and solved, and the high-dimensional flow field is reconstructed to obtain the high-dimensional flow field of the target working condition. The parameterized deformable mesh and multiple sets of street valley working parameters are numerically solved based on the fluid dynamics control equations to obtain an offline flow field database including multi-physics data, including: The building tilt angle is sampled and processed to obtain a set of street valley working condition parameters; For the parameterized deformable mesh corresponding to each set of street valley working parameters in the street valley working parameter set, steady-state numerical solution processing is performed based on the Reynolds-averaged Navier-Stokes equation and the pollutant transport control equation to obtain full-order flow field data, which includes velocity field, pressure field, turbulent viscosity field and pollutant concentration field. The full-order flow field data and corresponding grid data are classified, stored, and shared with the grid topology to obtain an offline flow field database.

2. The method according to claim 1, characterized in that, The process of performing intrinsic orthogonal decomposition and dimensionality reduction on the multiphysics field data in the offline flow field database to obtain low-dimensional spatial modal basis functions and modal coefficients includes: The multi-physics field data in the offline flow field database are processed to construct a snapshot matrix, thereby obtaining a high-dimensional flow field snapshot matrix corresponding to each physics field; The high-dimensional flow field snapshot matrix is ​​subjected to eigenorthogonal decomposition and truncated according to a preset modal energy ratio threshold to obtain the low-dimensional spatial modal basis functions corresponding to each physical field. Modal projection processing is performed on the full-order flow field data and the low-dimensional spatial modal basis functions to obtain the modal coefficients corresponding to each physical field.

3. The method according to claim 1, characterized in that, The process involves Galilean projection and closed-loop model training on the low-dimensional modal basis functions, modal coefficients, and turbulence data based on the fluid dynamics control equations to obtain a reduced-order dynamic system and a neural network turbulence closure prediction model, including: Based on the fluid dynamics control equations, the low-dimensional space modal basis functions and modal coefficients are subjected to Galerkin projection to obtain a reduced-order dynamic system; Training samples were constructed and network training was performed on the street valley working parameters, the modal coefficients corresponding to the velocity field, and the turbulent viscosity field in the full-order flow field data to obtain a neural network turbulence closure prediction model. Establish the correspondence between the low-dimensional turbulence closure term of the reduced-order dynamic system and the low-dimensional turbulence viscosity modal coefficients output by the neural network turbulence closure prediction model, complete the mapping matching process, and obtain the reduced-order dynamic system and the neural network turbulence closure prediction model that can be coupled.

4. The method according to claim 1, characterized in that, Obtaining the target geometric parameters corresponding to the target operating condition to be predicted, and then inputting the target geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term, includes: The target geometric parameters and the initial low-dimensional flow field modal coefficients are subjected to feature integration processing to obtain neural network input feature data; The neural network input feature data is input into the neural network turbulence closure prediction model for prediction calculation processing to obtain low-dimensional turbulence viscosity modal coefficients; The low-dimensional turbulent viscosity modal coefficients are subjected to closed-term transformation to obtain real-time low-dimensional turbulent closed terms.

5. The method according to claim 1, characterized in that, The process of coupling the solution of the reduced-order dynamic system, the real-time low-dimensional turbulence closure term, and the low-dimensional spatial modal basis functions, and reconstructing the high-dimensional flow field to obtain the high-dimensional flow field of the target working condition includes: The reduced-order dynamic system and the real-time low-dimensional turbulence closed term are coupled and iteratively solved to obtain the convergent low-dimensional modal coefficients of each physical field. The low-dimensional modal coefficients of each converged physical field and the low-dimensional spatial modal basis functions are subjected to high-dimensional flow field reconstruction processing to obtain high-dimensional distribution data of each physical field; The high-dimensional distribution data of each physical field are classified and integrated to obtain the high-dimensional flow field of the target working condition.

6. A device for predicting urban street valley flow field based on neural networks, characterized in that, include: The processing module is used to obtain the street valley inclination angle parameters, and adjust the spatial position of the grid nodes in the calculation area according to the street valley inclination angle parameters to obtain the parameterized deformable grid of urban street valleys with different inclination angles; The processing module is also used to perform numerical solution processing on the parameterized deformable mesh and multiple sets of street valley working parameters based on the fluid dynamics control equations to obtain an offline flow field database including multi-physics field data. The processing module is also used to perform intrinsic orthogonal decomposition and dimensionality reduction processing on the multiphysics field data in the offline flow field database to obtain low-dimensional spatial modal basis functions and modal coefficients; The training module is used to perform Galilean projection and closed-loop model training on the low-dimensional spatial modal basis functions, modal coefficients and turbulence data based on the fluid dynamics control equations, so as to obtain a reduced-order dynamic system and a neural network turbulence closed-loop prediction model. The input module is used to obtain the target geometric parameters corresponding to the target working condition to be predicted, and input the target geometric parameters and the initial low-dimensional flow field modal coefficients into the neural network turbulence closure prediction model for state feedback closure prediction processing to obtain the real-time low-dimensional turbulence closure term. The target geometric parameters include the target street valley inclination parameter. The processing module is also used to couple and solve the reduced-order dynamic system, the real-time low-dimensional turbulence closure term and the low-dimensional spatial modal basis function, and to reconstruct the high-dimensional flow field to obtain the high-dimensional flow field of the target working condition. The processing module is also used to numerically solve the parameterized deformable mesh and multiple sets of street valley working parameters based on the fluid dynamics control equations, to obtain an offline flow field database including multiphysics data, wherein: The building tilt angle is sampled and processed to obtain a set of street valley working condition parameters; For the parameterized deformable mesh corresponding to each set of street valley working parameters in the street valley working parameter set, steady-state numerical solution processing is performed based on the Reynolds-averaged Navier-Stokes equation and the pollutant transport control equation to obtain full-order flow field data, which includes velocity field, pressure field, turbulent viscosity field and pollutant concentration field. The full-order flow field data and corresponding grid data are classified, stored, and shared with the grid topology to obtain an offline flow field database.

7. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

Images