Physical mechanism guided city-wide building meter-scale high-resolution dynamic field heterogeneous deduction method

Abstract

This invention pertains to computational fluid dynamics and machine learning technologies, specifically a physical mechanism-guided heterogeneous extrapolation method for high-resolution, meter-level dynamic fields of urban buildings. The method includes the following steps: selecting representative urban blocks to form a sample block set; using a WRF model to provide a regional-scale meteorological background field for the target city; driving a LES model to perform flow field calculations on the sample block set, generating high spatiotemporal resolution flow field data for the sample blocks, and pairing this data with corresponding meteorological background data to generate a training database; constructing and training a deep learning model for latent space injection of meteorological information, establishing an end-to-end nonlinear mapping relationship from the coupling characteristics of urban geometry and meteorological boundaries to the LES fine flow field through a dual-path encoding path and latent space injection mechanism; and dividing the target urban area into multiple local computational units to extrapolate the high-resolution dynamic field of the entire urban area. This invention balances computational efficiency, prediction accuracy, and physical consistency, enabling rapid response to urban dynamic fields.

Description

Technical Field

[0001] This invention relates to the fields of computational fluid dynamics and machine learning, specifically to a physical mechanism-guided heterogeneous extrapolation method for meter-level high-resolution dynamic fields of buildings across the entire urban area, used to achieve efficient prediction and rapid response of high-resolution wind fields under complex terrain conditions in large cities. Background Technology

[0002] Urban wind fields are a crucial factor influencing urban environmental quality, pollutant dispersion, pedestrian comfort, and emergency response decisions. Accurate prediction of high-resolution wind fields at the urban scale is of great significance for urban planning, environmental assessment, and emergency management.

[0003] Currently, the following technical methods are mainly used for urban wind field prediction:

[0004] Method 1: Large Eddy Simulation (LES) Method

[0005] Large eddy simulation (LES) is a turbulence simulation technique that lies between direct numerical simulation (DNS) and Reynolds-averaged flow (RANS). This method can directly analyze large-scale turbulent structures, while employing a subgrid stress model for parameterized closure of subgrid-scale turbulence. LES can accurately capture complex flow phenomena within urban street canyons, such as flow separation and vortex shedding. However, LES is computationally intensive, and high-resolution simulations of large-scale urban areas are difficult to achieve even on supercomputers.

[0006] Method 2: Mesoscale meteorological models (such as WRF)

[0007] WRF (Weather Research and Forecasting Model) is a widely used mesoscale meteorological model. WRF can simulate regional-scale meteorological processes, such as frontal passage and sea-land breeze circulation. However, WRF's resolution is insufficient to resolve microscale (meter- to hundred-meter-scale) flow characteristics in urban areas, and it cannot characterize the disturbance effects of buildings on airflow in detail.

[0008] Method 3: Purely Data-Driven Machine Learning Approach

[0009] In recent years, some studies have attempted to establish a mapping relationship between urban geometry and wind fields using deep learning techniques, such as using convolutional neural network (CNN)-based architectures (e.g., U-Net) to directly predict flow fields from building morphology. However, these methods suffer from the following fundamental technical flaws, severely limiting their practical applications:

[0010] 1. Weak control over meteorological boundary conditions: Existing technologies often treat meteorological conditions (such as wind speed and direction) as scalar parameters or extend them into pseudo-image channels and simply concatenate them with building morphology before inputting them into the network. This approach suffers from severe physical semantic dilution: during the propagation process of meteorological parameters through multiple convolutional operations, their physical semantics as global boundary conditions are gradually diluted and localized, degenerating into local image features, which cannot effectively constrain flow field predictions to conform to atmospheric dynamic boundary conditions. Furthermore, the training process of existing methods often lacks high-fidelity LES data as a supervision target, mostly using low-resolution RANS data or simplified simulation data under idealized conditions. This results in the model establishing only a statistical correlation between geometry and flow field rather than a mapping of physical laws, failing to guarantee the physical consistency of the prediction results.

[0011] 2. High dependence on observational data and lack of cold-start capability: Some super-resolution methods rely on sparse ground observation stations or sensor networks for interpolation reconstruction, and completely fail in sensor-free areas or when sensors fail. These methods are essentially spatiotemporal interpolation techniques and do not have the ability to predict future moments, thus failing to meet the rapid prediction needs in emergency response scenarios.

[0012] 3. Lack of constraints from real physical mechanisms: Most studies use severely simplified boundary conditions (such as uniform inflow, single wind profile, and neutral atmospheric stratification) to generate training data, without considering the complexity of the real atmospheric environment, such as the constraints of mesoscale thermodynamic processes like diurnal variation of atmospheric stability, sea-land breeze circulation, and urban heat island circulation on the street-scale flow field; some methods only perform error correction or super-resolution reconstruction based on existing numerical simulation results, rather than learning the end-to-end physical mapping from boundary conditions to the flow field, resulting in a sharp decline in the prediction accuracy of the model under new meteorological conditions.

[0013] 4. Poor generalization ability and cross-scenario failure: Existing models struggle to adapt to different meteorological scenarios (such as varying wind speeds, directions, and atmospheric stability), resulting in a significant decrease in cross-scenario prediction accuracy. Due to the lack of explicit constraints from real atmospheric boundary conditions, predictions completely fail under atypical weather conditions (such as strong winds, temperature inversions, and complex boundary layer shear). This is primarily because the feature fusion mechanisms of existing technologies fail to effectively inject meteorological information as global boundary conditions, instead treating it as local features. Consequently, the models cannot learn the global constraint mechanism of meteorological boundary conditions on the flow field.

[0014] 5. Insufficient input feature representation leads to weak geometric perception: Some methods only use building height maps or binary masks as input, without introducing geometric features that can characterize the normal and spatial distance relationship of building boundaries, such as the signed distance function (SDF) and its gradient. Building height maps only provide geometric height information, but lack the directional information (normal vector) and distance information (distance to the boundary) of building boundaries. However, key flow phenomena such as the direction of flow separation, the shape of the wake, and the location of vortex shedding are highly sensitive to the geometric gradient and normal information of building boundaries. Due to the lack of these key geometric prior knowledge, the network has difficulty accurately capturing flow phenomena that are sensitive to boundary gradients, such as flow separation, wake direction, and vortex shedding. The prediction results have large errors at building edges and wake regions.

[0015] 6. Lack of physical consistency in the loss function leads to oversmoothing: Most studies only use pure data fitting loss functions such as mean squared error (MSE) without introducing physical constraint terms such as gradient constraints and divergence constraints; MSE loss tends to produce oversmoothed prediction results, resulting in the smoothing of wind speed gradients at building edges, blurring of wake boundaries, and loss of high-frequency turbulence details; it also cannot guarantee the mass conservation law of incompressible fluids (…). The flow field is predicted to have non-physical mass sources and sinks.

[0016] 7. Loss of high-frequency turbulence details: The standard U-Net network architecture is prone to losing high-frequency information during downsampling. Even with skip connections, it is difficult to fully recover extreme wind speeds and fine turbulence structures. Some studies have introduced generative adversarial networks (GANs) to refine the details, but GAN training is unstable, there is a pattern collapse problem, and it is difficult to reproduce and deploy in engineering.

[0017] Method 4: WRF-LES Coupling Method

[0018] To balance large-scale meteorological background with micro-scale flow details, some studies have employed a nested WRF-LES coupling approach. The WRF provides a regional-scale meteorological background field as the boundary condition for the LES, while the LES performs high-resolution calculations in local areas. However, this method still requires LES calculations for each region of interest, which remains computationally burdensome for large-scale urban applications.

[0019] In summary, the existing technology has the following technical problems:

[0020] 1. It is difficult to balance computational efficiency and accuracy: High-precision methods (LES) have excessive computational costs, while fast methods (pure machine learning) lack accuracy and physical consistency.

[0021] 2. Difficulty in large-scale urban applications: Existing methods are insufficient for high-resolution wind field prediction across the entire urban area;

[0022] 3. Training data lacks real physical constraints: The boundary conditions of the training data in pure data-driven methods are too idealized and not coupled with the real meteorological background;

[0023] 4. Lack of systematic technical solutions: Existing methods are all single technologies or combinations of some technologies, lacking a complete technical chain from sample selection, data generation, model training to global extrapolation. Summary of the Invention

[0024] To address the technical problems existing in the prior art, this invention provides a heterogeneous extrapolation method for meter-level high-resolution dynamic fields of urban buildings guided by physical mechanisms. This method balances computational efficiency, prediction accuracy, and physical consistency, and can respond quickly to urban dynamic fields. It meets the needs of practical applications such as emergency response and pollutant diffusion prediction under complex terrain conditions in large cities for real-time high-resolution wind field data.

[0025] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a heterogeneous extrapolation method for meter-level high-resolution dynamic fields of urban buildings guided by physical mechanisms, comprising the following steps:

[0026] S1. Based on urban morphological characteristics, select representative blocks to form a sample block set;

[0027] S2. Use the mesoscale meteorological model WRF to provide a regional-scale meteorological background field for the target city; combine GIS data to extract a fine building geometric model, drive the LES model with atmospheric boundary layer turbulence resolution capability to perform flow field calculations on the sample block set, and generate high spatiotemporal resolution flow field data of the sample blocks; pair the high spatiotemporal resolution flow field data of each sample block with the corresponding meteorological background data to generate a training database.

[0028] S3. Construct and train a deep learning model for latent space injection of meteorological information; in the deep learning model, through a dual-path encoding path and a latent space injection mechanism, establish an end-to-end nonlinear mapping relationship from the coupling features of urban geometry and meteorological boundaries to the LES fine flow field.

[0029] S4. Divide the target city area into multiple local computing units; for each local computing unit, extract its street morphology features and corresponding meteorological background vector, input them into the trained deep learning model, and obtain the high-resolution wind field prediction results of the corresponding local computing unit; through spatial stitching and edge smoothing, eliminate the numerical discontinuity at the boundary of the local computing unit, and reconstruct the prediction results of all local computing units into the full-area high-resolution dynamic field of the target city.

[0030] Compared with the prior art, the technical effects achieved by the present invention include:

[0031] 1. Significantly improved computational efficiency

[0032] By employing a strategy that combines sample selection with machine learning, the enormous computational overhead of full-domain LES simulation is avoided. It can complete wind field predictions under given urban geometry and meteorological conditions within a second-level timescale, achieving a computational speedup of 3-4 orders of magnitude compared to traditional CFD simulations.

[0033] 2. Prediction accuracy maintains physical consistency

[0034] The training data is generated based on WRF-LES multi-scale coupling, possessing realistic boundary conditions and physical consistency. The deep learning model uses the high-fidelity LES flow field as the supervised objective, implicitly learning the motion laws of fluids under complex boundary conditions by learning the end-to-end nonlinear mapping from the coupling characteristics of urban geometry and meteorological boundaries to the fine LES flow field.

[0035] 3. Physical consistency of meteorological boundary condition constraints

[0036] Through the meteorological information latent space injection mechanism, WRF meteorological boundary conditions are injected as global constraints into the deep feature space, avoiding the dilution and localization of meteorological information during multi-layer convolution and maintaining the consistency and physical integrity of meteorological information as a global boundary condition constraint. Functionally, this mechanism is analogous to the boundary condition driving mechanism in the WRF-LES coupling framework, ensuring that the flow field predicted by the model conforms to the large-scale atmospheric dynamic background.

[0037] 4. Enhanced geometric feature representation capability

[0038] By introducing the Signed Distance Function (SDF) and its spatial gradient as input features, the network is able to perceive the normal information and spatial distance relationships of building boundaries, significantly improving the prediction accuracy for gradient-sensitive flow phenomena such as flow separation, wake direction, and vortex shedding. Compared to existing methods that only use building height maps or binary masks, the SDF gradient provides the network with richer geometric prior knowledge, helping the model learn physically correct flow-geometric mapping relationships.

[0039] 5. Strong ability to recover flow field details

[0040] A composite loss function with physical constraints is adopted, specifically including MSE data fitting terms, gradient constraint terms, and divergence constraint terms. The gradient constraint terms sharpen building edges and wake boundaries, and the divergence constraint terms force the mass conservation law to be satisfied. This effectively solves the problem of over-smoothing of the flow field caused by a single MSE loss, ensuring that the prediction results can accurately recover extreme wind speeds and high-frequency turbulence details, while conforming to the mass conservation law of incompressible fluids.

[0041] 6. Strong generalization ability

[0042] The model simultaneously learns the influence of urban geometry and WRF meteorological boundary conditions, enabling it to adapt to various meteorological scenarios such as wind speed, wind direction, and atmospheric stability. Its cross-scenario generalization ability is significantly better than conventional machine learning methods. For training with unseen street geometry combinations or new WRF meteorological boundary condition outputs, the model can achieve reliable physical consistency predictions.

[0043] 7. Systematic technical solutions

[0044] This invention provides a complete technical chain from sample screening, WRF-LES data generation, dual-stream feature decoupling training to full-domain extrapolation. Each stage is interconnected and mutually supportive, forming an efficient dynamic field rapid response system for complex terrains in large cities.

[0045] 8. Cold start capability with zero observation dependency

[0046] Unlike existing super-resolution methods that often rely on dense ground observation stations or sensor networks as input, this invention couples with a WRF mesoscale model, requiring only the meteorological boundary conditions output by the WRF to deduce the micro-flow field. This characteristic makes this method particularly suitable for urban planning in areas without observational data, emergency response in areas where sensors fail, and forecasting of future moments, demonstrating strong engineering practicality and cold-start capability. Attached Figure Description

[0047] Figure 1 This is the overall flowchart of the heterogeneous deduction method for meter-level high-resolution dynamic field of buildings in the entire city area according to the present invention. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0049] Example

[0050] This embodiment takes a certain urban area as the research object. This urban area includes a variety of typical urban forms, from the high-density historical protection area within the second ring road and the super high-rise building cluster in the CBD area to the row-style residential area on the periphery. The surface roughness is complex and varied, making it highly representative.

[0051] See Figure 1 The physical mechanism-guided heterogeneous extrapolation method for meter-level high-resolution dynamic fields of buildings across the entire city in this embodiment includes the following steps:

[0052] S1. Based on the characteristics of urban morphology, select representative blocks to form a sample block set.

[0053] Multidimensional morphological features of blocks across the entire target city are extracted to construct a high-dimensional morphological feature space. The high-dimensional morphological feature space is then orthogonally reduced to extract key principal components. Based on the key principal components, a hybrid sampling strategy is used to select centroid samples and boundary samples from each cluster category to construct a set of physically consistent sample blocks.

[0054] In the high-dimensional morphological feature space, morphological features include, but are not limited to, building density, street valley height, building coverage, floor area ratio, building dispersion, building differentiation, spatial openness, building height standard deviation, building enclosure degree, and sky view factor.

[0055] In this embodiment, principal component analysis is used to orthogonalize and reduce the dimensionality of the high-dimensional morphological feature space. The top principal components whose cumulative explained variance contribution rate exceeds a set threshold (e.g., 85%) are extracted as clustering input variables. The K-means clustering algorithm is used to divide all blocks of the city into K categories based on the main morphological information. Representative sample blocks are selected from each category to form a sample block set.

[0056] By using the above-mentioned hybrid sampling based on multidimensional morphological features, it is ensured that the training data encountered by the subsequent deep learning model contains the most complete morphological change rules in the entire domain. This allows the model to implicitly reconstruct the influence of these complex morphological features on the flow field through convolution operations, even when only basic height field images are input.

[0057] The innovation of this step lies in the fact that, in the face of the reality of a large number and diverse types of urban blocks, a scientific sample selection strategy is used to reduce a large number of complex scenes into a limited number of categories. This significantly reduces the number of computational samples while maintaining the diversity of urban forms, thus providing controllable computational overhead for subsequent physical mechanism-based CFD simulations.

[0058] More specifically, to meet the requirements of computational domain regularity in subsequent Large Eddy Simulation (LES), this embodiment uses a standardized overlapping grid to systematically divide the study area. The standard size of the target training samples is set to 640m × 640m. To eliminate the numerical boundary effect of the flow field caused by the grid boundaries cutting off buildings, an 80m buffer zone is added around the standard grid when actually extracting the physical simulation computational domain, forming an actual computational domain size of 800m × 800m. A sliding window is used to slice the entire urban area, with a sliding step size of 640m, resulting in approximately 1628 standardized candidate grids.

[0059] For each candidate grid, multidimensional morphological features, including building density, street valley height, building coverage, floor area ratio, sky view factor, spatial openness, and building height standard deviation, are extracted to construct a high-dimensional morphological feature matrix. After Z-score standardization of the feature matrix, orthogonal dimensionality reduction is performed using principal component analysis (PCA). The cumulative variance contribution rate of the first four principal components reaches 88.6%, exceeding the set threshold (85%). Therefore, these first four principal components are selected to construct the orthogonal feature space after dimensionality reduction, effectively eliminating multicollinearity among the original features, such as the strong correlation between floor area ratio and building height.

[0060] Based on the dimensionality-reduced feature space, the K-means++ clustering algorithm is used to divide all candidate grids into K=13 morphological categories, where the number of clusters K is determined according to the elbow rule; then, representative grids are selected from each cluster category according to the centroid-boundary hybrid sampling strategy, that is, representative sample blocks are selected.

[0061] Furthermore, centroid and boundary samples are selected from each cluster category to construct a set of physically consistent sample blocks, specifically including:

[0062] Calculate the feature spatial centroid of all block morphological features within each category;

[0063] Centroid samples and boundary samples are selected from each category. The centroid samples are the blocks with the smallest Euclidean distance from the centroid of the feature space, which are used to characterize the general geometric features of this type of urban form. The boundary samples are the top 10% of blocks with the largest Euclidean distance from the centroid of the feature space, which are used to characterize the edge features and morphological extremes of this type of urban form.

[0064] The centroid samples and boundary samples together constitute the sample block set, ensuring that the training data covers a complete distribution from typical to extreme forms.

[0065] In other words, the centroid sample is selected from the grids closest to the centroid of the category feature, representing the typical features of this type of morphology; the boundary sample is selected from the grids furthest from the centroid in the feature space but still within the effective statistical range, to cover the extreme variations of this type of morphology. Ultimately, 50 representative grids were selected to form the sample street set.

[0066] S2. Generate high-fidelity flow field training data with multi-scale coupling.

[0067] The mesoscale meteorological model WRF is used to provide regional-scale initial and time-varying boundary conditions (i.e., regional-scale meteorological background field) for the target city. A detailed building geometry model is extracted from GIS data, and a large eddy simulation (LES) model with atmospheric boundary layer turbulence resolution capability is driven to perform flow field calculations on the sample block set, generating high spatiotemporal resolution flow field data of the sample blocks. The high spatiotemporal resolution flow field data of each sample block is paired with the corresponding meteorological background data to generate a training database.

[0068] The specific steps for generating high spatiotemporal resolution flow field data include:

[0069] The regional-scale meteorological field output by the mesoscale meteorological model WRF is used as the boundary and initial conditions of the LES model.

[0070] Transform detailed architectural geometry models into geometric meshes recognizable by the LES model;

[0071] Driven by the boundary and initial conditions, the LES model performs dynamic solutions for each sample block to obtain a three-dimensional velocity field.

[0072] In this embodiment, a detailed building geometric model is extracted from GIS data, driving a Large Eddy Simulation (LES) model with atmospheric boundary layer turbulence resolution capabilities to perform three-dimensional unsteady LES simulation calculations for atmospheric boundary layer turbulence in each sample block of the sample block set. By numerically solving the spatially filtered incompressible fluid control equations, the dominant turbulence structure at the block scale and above is explicitly resolved, and a subgrid-scale turbulence model is used to characterize the momentum exchange process at the subgrid scale, generating high spatiotemporal resolution flow field data. The high spatiotemporal resolution flow field has a spatial resolution (e.g., 5m level or better) capable of resolving the vortex structure and flow separation characteristics within the street valley.

[0073] The innovation of this step lies in the adoption of the WRF-LES multi-scale coupling framework. WRF provides the regional meteorological background field, enabling LES simulations to be conducted under realistic atmospheric constraints. This invention utilizes flow field data generated by the WRF-LES coupling system, containing complete physical evolution mechanisms and multi-scale turbulence details, for training. This allows the deep learning model to construct a high-dimensional nonlinear mapped manifold constrained by the Navier-Stokes (NS) equations, thereby implicitly learning the motion laws of fluids under complex boundary conditions. Combined with the latent injection of meteorological information, the model can not only recover the mean flow field but also achieve refined reconstruction of high-frequency turbulence details and statistical characteristics such as turbulent kinetic energy. Its physical consistency is significantly better than simple error correction models. The generated training database possesses realistic boundary conditions and physical consistency, providing high-quality training samples for the deep learning model.

[0074] The process of generating high-fidelity flow field training data in this step includes the following key technical features:

[0075] (1) Multi-scale coupling architecture: The WRF model adopts a multi-layer nested grid configuration. The outer coarse resolution grid provides the regional meteorological background, and the inner fine resolution grid covers the target urban area. The horizontal resolution is gradually refined from the outer layer to the inner layer, such as a four-layer nested structure of 9km→3km→1km→333m.

[0076] (2) Boundary condition transfer: The time-varying three-dimensional meteorological field (including wind speed components u, v, w, potential temperature θ, and water vapor mixing ratio) output by the innermost nested WRF model is transferred. By spatial interpolation and temporal resampling, the four side and top boundaries of the LES computational domain are mapped to form time-varying, non-uniform dynamic boundary conditions.

[0077] (3) High-resolution LES calculation: The LES solver is used to numerically solve the filtered incompressible fluid control equations, which include the continuity equation. and momentum equation ,in It is a velocity vector. For pressure, For density, Kinematic viscosity, For the sublattice stress tensor, This is the buoyancy term;

[0078] (4) Subgrid-scale model: The 1.5th order turbulent closure scheme (such as the Deardorff scheme) is adopted, and the subgrid stress is parameterized by introducing and solving the prediction equation of subgrid turbulent kinetic energy (SGS-TKE);

[0079] (5) Grid resolution and time step: The LES computational grid resolution is set to 1-10 meters (e.g., 5m) to resolve the vortex structure and flow separation characteristics at the street valley scale. The time step is adaptively adjusted according to the CFL (Courant-Friedrichs-Lewy) conditions. ;

[0080] (6) Training sample extraction: The simulation duration of each LES case is 2-4 hours, of which the first 0.5-1 hours is the model spin-up period. Only the subsequent period is extracted for training samples. The instantaneous flow field is extracted once every certain time interval (e.g., 30 seconds to 5 minutes) or the flow field is averaged over time (e.g., 10-30 minutes). The statistical average flow field is extracted as the training target.

[0081] More specifically, considering the urban climate characteristics, several statistically representative meteorological scenario combinations were selected from historical meteorological data over the past five years. These combinations include: seasonal prevailing wind conditions, covering typical prevailing wind directions in spring, summer, autumn, and winter; wind speed levels, including low, moderate, and strong winds; and atmospheric stability conditions, including unstable, neutral, and stable stratification states. Through the combination of these multi-dimensional meteorological scenarios, a multi-scenario meteorological driver set covering the city's major large-scale dynamic background throughout the year is constructed for subsequent multi-scale coupled simulations.

[0082] The WRF simulation employs a four-layer nested structure, with horizontal resolution progressively refined from 9 km to 333 m. The initial field and boundary conditions of the WRF model utilize NCEP FNL global reanalysis data. Only the simulation output of the innermost nested region (D04) is extracted as the initial field and dynamic boundary field for the large eddy simulation. The wrf4palm tool is used to interpolate and map the three-dimensional meteorological field (u, v, w wind speed components, potential temperature, and water vapor mixing ratio) of the WRF model onto the four side and top boundaries of the large eddy simulation model, forming time-varying, non-uniform dynamic boundary conditions. Unlike the fixed incoming flow profile or idealized uniform background wind settings commonly used in traditional large eddy simulations, this embodiment achieves a realistic response of the street-scale flow field to the spatiotemporal evolution of large-scale weather processes by introducing the output of a mesoscale meteorological model.

[0083] The PALM large eddy simulation model was used to calculate the flow field of 50 sample blocks under multiple scenarios. The PALM model adopted a 1.5-order turbulent closure scheme, with a grid resolution of 5m. The time step was adaptively adjusted according to the CFL condition. The simulation time for each case was 2 hours, with the first 0.5 hours being the model spin-up period and the last 1.5 hours being used for statistical analysis and training sample extraction.

[0084] The training database generation strategy used in this embodiment has the following characteristics:

[0085] (1) Physical consistency: The flow field of each training sample is obtained by solving the NS equation through LES under the driving force of real WRF meteorological boundary conditions, which includes complete atmospheric dynamic constraints and building geometric boundary perturbations. Compared with training data using idealized uniform inflow or single wind direction profiles, the training data generated in this embodiment truly reflects the spatiotemporal variation characteristics of the atmospheric boundary layer (such as wind speed and direction shear with height, diurnal variation of atmospheric stability, sea-land breeze circulation and other mesoscale processes), enabling the deep learning model to learn the geometric-flow field mapping relationship that conforms to physical laws.

[0086] (2) Multi-scenario coverage: By selecting WRF meteorological background fields with different seasons, different weather systems and different atmospheric stability conditions, a training sample set covering the main meteorological scenarios of the city throughout the year is generated, which enhances the model's ability to generalize to diverse meteorological conditions.

[0087] (3) Multi-scale information coupling: The training data implicitly contains multi-scale coupling relationships from regional meteorological processes to street-scale flow characteristics. The large-scale dynamic constraints provided by WRF and the micro-scale turbulence details analyzed by LES achieve natural coupling at the data level, laying the data foundation for the heterogeneous feature decoupling training of deep learning models.

[0088] (4) High-fidelity supervision signal: The high-resolution flow field provided by LES is used as the supervision target, which includes background average flow, coherent structure at the inertial sub-region scale and small-scale random turbulent fluctuations. It provides deep learning models with a full spectrum of supervision signals from macroscopic flow patterns to microscopic turbulent details, enabling the model to learn average flow field and high-frequency turbulent details at the same time.

[0089] S3. Construct and train a deep learning model for injecting meteorological information into the latent space.

[0090] This step constructs a deep learning model based on heterogeneous feature decoupling, namely a two-stream deep learning network. The two-stream deep learning network has dual encoding paths. Through these dual encoding paths and a latent space injection mechanism, it establishes an end-to-end nonlinear mapping relationship from the coupled features of urban geometry and meteorological boundaries to the LES fine flow field. This step specifically includes:

[0091] S31. Construct a deep learning model for injecting meteorological information into the latent space.

[0092] Specifically, the deep learning model includes a spatial coding path, a meteorological coding path, and a decoding path. The spatial coding path and the meteorological coding path are fused in the deep feature space of the dual-stream deep learning network through a latent space injection mechanism, while the decoding path reconstructs a high-resolution wind field based on the fused features.

[0093] (1) The Spatial Encoder adopts a convolutional neural network architecture with multiple downsampling stages, including N downsampling stages (e.g., N=4 or N=5). Each downsampling stage includes: convolutional layers, using 3×3 or 5×5 convolutional kernels for local feature extraction, with a stride of 1 or 2; Batch Normalization: standardizing the convolutional output to accelerate training convergence and improve numerical stability; Activation functions: using nonlinear activation functions such as ReLU, LeakyReReLU, or GELU to introduce nonlinear expressive power; Residual connections or dense connections: using ResNet or DenseNet modules to enhance feature propagation capabilities and alleviate the gradient vanishing problem; The downsampling operation is: using strided convolution or max pooling to halve the spatial resolution of the spatial feature map.

[0094] The input to the spatial encoding path is a multi-channel geometric feature map consisting of the building height field, building mask, signed distance function (SDF), and its spatial gradient components. ,in The height of the input image. The width of the input image. This refers to the number of geometric feature channels (typically 5 or 6). A multi-channel geometric feature map specifically includes:

[0095] Channel 1: Building height field, representing the building height of each grid point, in meters;

[0096] Channel 2: Building mask, binary mask (0 indicates no building, 1 indicates building).

[0097] Channel 3: Signed Distance Function (SDF), which represents the directed distance from a grid point to the building boundary;

[0098] Channels 4-5: SDF spatial gradient ( , The normal information representing the building boundary provides crucial geometric prior knowledge for the network; among which... Represents the SDF horizontal gradient. The SDF vertical gradient can be calculated using the central difference scheme; the specific formula for calculating the SDF spatial gradient is:

[0099] ;

[0100] ;

[0101] Optional Channel 6: Standard deviation of building height or other morphological characteristics.

[0102] The spatial gradient component includes the partial derivative of the signed distance function (SDF) in the horizontal direction. , This is used to characterize the normal information of building boundaries, enhancing the network's ability to predict flow separation and wake direction; each downsampling stage outputs feature maps with different spatial resolutions, forming a multi-scale spatial feature representation. In this embodiment, for two-dimensional grid points... The signed distance function is defined piecewise:

[0103] ;

[0104] in Represents the set of boundary points of the building. The sign of the SDF (Signal Distance Function) represents the Euclidean distance; the positive and negative signs indicate the exterior and interior of the building, respectively, and the absolute value represents the distance to the boundary. The spatial gradient of the SDF is calculated using the central difference scheme:

[0105] , ;

[0106] in and The grid spacing is defined as follows. The signed distance function (SDF) and its spatial gradient serve as additional input channels, providing the network with unit normal information about the building boundaries:

[0107] ;

[0108] in The value is the numerical stability constant, which is taken as a value in this embodiment. To prevent the denominator from being zero, the unit normal information is closely related to the flow separation direction of the fluid on the building surface. Specifically: fluid separation occurs at the windward edge of the building, and the location of the fluid separation point is determined by the normal of the windward side of the building; while the wake direction, the expansion direction of the wake downstream of the building, is related to the normal height of the building's sides and leeward side; at the sharp edges or curved areas of the building, the direction and frequency of vortex shedding are affected by the boundary curvature and normal changes; in addition, the complex flow structures in the corner areas of the building (such as horseshoe vortices and corner vortices) are highly sensitive to geometric gradients. Therefore, the SDF gradient, as an input feature, provides the network with key geometric prior knowledge, significantly enhancing the network's ability to predict gradient-sensitive flow phenomena such as flow separation points, wake direction, and vortex shedding. In contrast, existing methods that only use building height maps or binary masks lack this key gradient information, resulting in insufficient prediction accuracy at building edges and wake areas.

[0109] The output of the spatial encoding path is a multi-scale spatial feature representation. ,in Indicates the first Feature maps of each stage and For spatial resolution, The number of channels typically increases with depth. These multi-scale spatial feature maps capture urban morphological information from local details (shallow layer) to global structure (deep layer), corresponding to the multi-scale perturbation effect of building geometry as an internal lower boundary condition on the flow field in the LES solution.

[0110] More specifically, the spatial coding path employs four downsampling stages, with each stage configured as follows:

[0111] Stage 1: Input 5-channel geometric feature map (resolution 128×128) → through 2 3×3 convolutional layers (each convolutional layer contains 64 convolutional kernels, stride 1, batch normalization, ReLU activation) → 2×2 max pooling (stride 2) downsampling → output feature map resolution 64×64, number of channels 64;

[0112] Stage 2: Input 64-channel feature map (64×64) → pass through 2 3×3 convolutional layers (128 convolutional kernels) → 2×2 max pooling downsampling → output feature map with a resolution of 32×32 and 128 channels;

[0113] Stage 3: Input 128-channel feature map (32×32) → pass through 2 3×3 convolutional layers (256 convolutional kernels) → 2×2 max pooling downsampling → output feature map with a resolution of 16×16 and 256 channels;

[0114] Stage 4 (bottleneck layer): Input 256-channel feature map (16×16) → through 2 3×3 convolutional layers (512 convolutional kernels) → no downsampling → output feature map with a resolution of 16×16 and 512 channels.

[0115] Each stage of the convolutional layer employs residual connections to enhance feature propagation capabilities. For the output of the i-th stage... The output of the residual block is Where BN is batch normalization. This represents the convolutional transformation at stage i. Residual connections effectively alleviate the gradient vanishing problem in deep networks and enhance the cross-layer propagation capability of features.

[0116] (2) The meteorological encoding path adopts a multi-layer perceptron (MLP) architecture, including... One fully connected layer (typical value) or ).

[0117] The input to the meteorological coding path is the WRF meteorological boundary condition vector. ,in Meteorological parameter dimension; vector The time-varying meteorological field output from the WRF model is extracted, specifically including wind speed components, potential temperature, and turbulence parameters at different altitudes. The meteorological coding path maps the low-dimensional meteorological boundary vector to a high-dimensional meteorological latent space vector through a multi-layer fully connected transformation.

[0118] ;

[0119] in For the first The weight matrix of the layer, For the first The layer's bias vector, For activation functions (such as ReLU, GELU), For meteorological potential space vectors; This is the latent space dimension, which is usually set to the same number of channels as the bottleneck layer of the spatial coding path, such as 512 or 1024.

[0120] The WRF output in the LES solution corresponding to the meteorological coding path acts as a global constraint on the time-varying meteorological fields of the four boundaries and the top boundary, compressing the high-dimensional spatiotemporal non-uniform meteorological field into a compact latent space vector, and capturing the key atmospheric dynamic features that affect the flow field at the street scale.

[0121] (3) The latent space injection mechanism is set in the deep feature space of the network. The latent space injection mechanism expands the high-dimensional meteorological latent space vector to the same size as the spatial features through spatial broadcasting operations, and adjusts the influence strength of meteorological boundary constraints through feature fusion and gating modulation. The processing steps of the latent space injection mechanism specifically include the following:

[0122] 1) Space Broadcasting

[0123] The meteorological latent space vector output by the meteorological coding path Extended to spatial feature maps through spatial copying operations For the same dimensions, the meteorological characteristics obtained after space broadcasting are:

[0124] Where Broadcast represents a broadcast operation, which transmits the vector Replication along spatial dimensions This process generates a three-dimensional tensor containing complete meteorological information for each spatial location.

[0125] 2) Feature fusion

[0126] Meteorological characteristics after space broadcasting Spatial features Feature fusion can be performed. Feature fusion can be achieved using at least one of the following methods:

[0127] Channel concatenation: Spatial and meteorological features are directly spliced ​​together along the channel dimension;

[0128] Adaptive Instance Normalization (AdaIN): ,in and Representing the mean and standard deviation respectively, AdaIN modulates the statistical distribution of the spatial feature map using meteorological latent vectors;

[0129] Feature Affine Transformation: ,in and The scaling factor and bias factor are generated from the meteorological latent vector. This represents element-wise multiplication;

[0130] Cross-Attention mechanism: ,in To query (weather potential vector), and Using keys and values ​​(spatial features), an attention mechanism is used to dynamically modulate the spatial features by meteorological conditions.

[0131] 3) Gated modulation.

[0132] After feature fusion is completed, the relative contributions of spatial features and meteorological features are adaptively adjusted through a gated fusion unit:

[0133] , ;

[0134] in For the gating weight graph, Constrain the value of the sigmoid activation function to . interval, This indicates a channel splicing operation. and All are learnable parameters. The fused feature is obtained by any of the above feature fusion methods. This is the characteristic output after gating and modulation.

[0135] The gating mechanism enables the network to adaptively adjust the influence intensity of meteorological boundary constraints according to the characteristics of different regions: in densely built areas, the gating weights tend to give greater weight to spatial features (emphasizing geometric perturbations); in open areas, the gating weights tend to give greater weight to meteorological features (emphasizing atmospheric dynamic background).

[0136] The latent space injection mechanism is set in the deep feature space of the network, specifically in the bottleneck layer of the spatial encoding path, i.e., the deepest feature map of the encoder. The deeper layers of the decoding path, i.e., one or more upsampling stages near the bottleneck layer of the decoder; or simultaneously set in the bottleneck layer of the spatial coding path and multiple deeper layers of the decoding path, forming a hierarchical meteorological condition injection.

[0137] This latent space injection mechanism is functionally analogous to the boundary condition-driven process in the WRF-LES coupled framework: the meteorological encoding path corresponds to the boundary condition extraction from the WRF output, latent space injection corresponds to the LES solver accepting boundary condition constraints, and the spatial encoding path corresponds to the LES response to building geometric boundary conditions. Unlike existing technologies that simply extend meteorological parameters to image channels and geometric features followed by direct convolution, this dual-stream feature extraction architecture, combined with the latent space injection mechanism, avoids the physical semantic dilution of meteorological boundary conditions during multiple convolution operations, maintaining the consistency and physical integrity of meteorological information as a global boundary condition constraint.

[0138] (4) The decoder adopts a deconvolutional neural network architecture with multiple upsampling stages. Each upsampling stage gradually restores the spatial resolution through transposed convolution or interpolation operations, and receives the corresponding resolution features from the spatial encoding path through dense skip connections. Among them, the dense skip connections adopt a nested dense connection structure to pass the features of different levels of the encoder to the corresponding levels of the decoder, forming a feature pyramid, which is used to recover high-frequency turbulence details.

[0139] More preferably, the deep learning model adopts the U-Net++ architecture, including a nested dense skip connection structure; the spatial encoding path is the encoder part of the U-Net++ architecture, containing N downsampling stages, each stage including a feature extraction module composed of convolutional layers, batch normalization layers, and activation functions, using residual connections or dense connections to enhance feature propagation capability; the meteorological encoding path, as an independent fully connected branch, consists of M fully connected layers, connected to the bottleneck layer of the U-Net++ architecture, mapping the meteorological boundary vector to a latent space vector matching the number of channels in the bottleneck feature map; the dense skip connection passes all feature maps of encoder stage i to decoder stage j (j≥i), achieving multi-scale feature fusion through upsampling and channel concatenation operations. The mathematical expression for the dense skip connection is:

[0140] ;

[0141] Where x represents the feature map, UpSample represents the upsampling operation, and Concat represents the channel concatenation operation.

[0142] This nested, densely connected structure is a core feature of the U-Net++ architecture. Compared to the simple skip connections of the standard U-Net, dense skip connections pass features from different levels of the encoder to the corresponding levels of the decoder, forming a feature pyramid, which helps to recover the high-frequency turbulence details lost during downsampling.

[0143] The output of the decoding path is a high-resolution wind field prediction. It includes two channels: horizontal wind speed component. and The spatial resolution of the predicted wind field is the same as that of the input geometric feature map, for example... The actual dimensions are 640m × 640m, with a grid spacing of 5m.

[0144] S32. Train the constructed deep learning model.

[0145] In this embodiment, the deep learning model is trained under supervision using the training database generated in step S2. High spatiotemporal resolution flow field data calculated by the Large Eddy Simulation (LES) model is used as the supervision target, and a composite loss function including physical constraints is employed to minimize the difference between the predicted and target wind fields. The composite loss function includes a data fitting term, a gradient constraint term, and a divergence constraint term. The data fitting term ensures the accuracy of the flow field prediction, the gradient constraint term sharpens building edges and wake boundaries, and the divergence constraint term enforces the law of conservation of incompressible fluid mass, thereby establishing an end-to-end nonlinear mapping relationship from the coupling characteristics of urban geometry and meteorological boundaries to the fine LES flow field.

[0146] In this embodiment, the composite loss function ensures that the predicted flow field conforms to the basic laws of fluid mechanics through physical constraint terms, and its specific form is as follows:

[0147] ;

[0148] in, For the data fitting term, the mean square error (MSE) or mean absolute error (MAE) between the predicted wind field and the LES target wind field is used, and its mathematical expression is as follows: ,in For wind fields predicted by the network, Target wind field calculated for LES; Indicates the horizontal velocity of the wind field and The various components, This represents the total number of grid points. The data fitting term ensures that the predicted flow field is consistent with the LES target flow field overall and serves as a fundamental monitoring signal.

[0149] This is the gradient constraint term, used to calculate the error between the predicted flow field gradient magnitude and the LES flow field gradient. Its mathematical expression is: This is used to sharpen predictions of building edges and flow separation interfaces, where... Represents the spatial gradient operator, Indicates the wind field number The spatial gradient consists of several components. The gradient constraint term serves several purposes: sharpening building edges, where wind speed gradients change drastically (from zero velocity at the building surface to high shear near the wall), and forcing the predicted flow field to accurately recover these high-frequency gradients; sharpening flow separation interfaces, where shear layers near flow separation points exhibit strong gradient characteristics, and enhancing the network's predictive ability for these interfaces; and sharpening wake boundaries, where velocity gradients exist at the boundary between the wake region and the free flow region, helping to accurately characterize the wake extension range. Compared to using only MSE loss (which tends to produce overly smoothed predictions), the gradient constraint term significantly improves the predicted flow field's ability to recover high-frequency flow structures.

[0150] This is a divergence constraint term used to calculate the horizontal divergence of the predicted flow field. The law of conservation of mass for incompressible fluids is enforced, and its mathematical expression is: ,in Represents the divergence of the horizontal flow field. , These are the two velocity components of the wind field in the horizontal direction. For incompressible fluids, the continuity equation requires that the velocity vector be horizontal; The divergence constraint ensures that the predicted flow field satisfies the law of mass conservation, avoiding the appearance of non-physical mass sources or sinks.

[0151] , These are all weighting coefficients used to balance the contributions of different loss terms, with typical values ​​ranging from [value range missing]. , The choice of weighting coefficient needs to be adjusted according to the specific application scenario: for urban areas with dense buildings and significant flow separation, the weighting coefficient can be appropriately increased. To enhance gradient constraints and sharpen building edges and wake boundaries; for applications with high mass conservation requirements (such as pollutant diffusion simulation), the gradient can be appropriately increased. To strengthen divergence constraints and ensure mass conservation.

[0152] In step S3, the training process of the deep learning model includes the following sub-steps:

[0153] Step S321: Training data preprocessing. The geometric feature maps and LES wind fields in the training database are normalized using zero-mean, unit-variance standardization. The normalized data distribution is as follows:

[0154] ;

[0155] in The mean of the training set. denoted as the standard deviation of the training set. Normalization eliminates the influence of differences in the dimensions and scales of different features, accelerating training convergence and improving numerical stability.

[0156] Step S322, Batch Sampling and Augmentation: Randomly sample small batches of training samples from the training database and apply data augmentation operations such as random flipping and random rotation to the input geometric feature map to improve the model's generalization ability.

[0157] The data augmentation operations specifically include: random horizontal flipping (mirroring the input horizontally with a 50% probability); random vertical flipping (mirroring the input vertically with a 50% probability); random rotation (rotating the input randomly within the range of -90°, +90°); and random scaling (scaling the input size randomly within the range of 0.9, 1.1). It is important to note that the geometric transformations applied to the geometric feature map need to be simultaneously applied to the LES target wind field to maintain input-output consistency.

[0158] Step S323: Forward propagation, inputting randomly sampled small batches of training samples into the deep learning model, and outputting the predicted wind field through the step-by-step transformation of each layer of the model.

[0159] For spatially encoded paths, input → Output multi-scale features{ , ..., }; Meteorological coding path, input m → output meteorological latent space vector m'; latent space injection, inject m' into (or other deeper layers) to obtain fusion features Decode path, input → Output predicted wind field .

[0160] Step S324: Calculate the loss by calculating the loss value of the current batch based on the composite loss function;

[0161] Step S325, Backpropagation and Parameter Update: Stochastic Gradient Descent (SGD) or its variants (such as Adam or AdamW optimizers) are used to calculate the gradient of the composite loss function with respect to the network parameters, and the network parameters are updated.

[0162] ;

[0163] in For network parameters (including weights and biases), This is the learning rate.

[0164] The Adam optimizer adjusts the learning rate for each parameter by adaptively estimating the first and second moments:

[0165] ;

[0166] ;

[0167] ;

[0168] in For the current gradient, and For the estimation of the first and second moments of the gradient, , Exponential decay rate (typical value) , ), Numerical stability constant (e.g.) ).

[0169] Step S326: Iterative training. Repeat steps S322 to S325 until the composite loss function converges or the preset number of iterations is reached. A learning rate decay strategy is used during iterative training to improve convergence accuracy in the later stages of training.

[0170] ,in The initial learning rate, The total number of training rounds. The attenuation coefficient;

[0171] Step S327, Model Validation and Saving: Periodically evaluate the performance of the deep learning model on the validation set, and save the model parameters with the minimum loss on the validation set as the optimal model.

[0172] Model performance was evaluated every 10 epochs on the validation set (e.g., every 5 or 10 epochs), using mean absolute error (MAE) and correlation coefficient. The MAE on the validation set is used as the evaluation metric. The model parameters with the smallest MAE on the validation set are saved as the optimal model. An early stopping strategy (patience=30) is used to prevent overfitting: if the MAE on the validation set does not improve within 30 consecutive epochs, training is stopped and the optimal model parameters are restored.

[0173] Compared to the standard U-Net baseline model that uses only building height maps as input and only employs the MSE loss function, this embodiment uses a deep learning model with decoupled heterogeneous features of two streams, which significantly outperforms the baseline model in the following aspects: thanks to the SDF gradient input and gradient constraint terms, the wind speed gradient recovery capability at the building edge is improved by 25-35%; thanks to the dense skip connections of U-Net++ and the injection of meteorological information into the latent space, the prediction accuracy of the flow field structure in the wake region is improved by 20-30%.

[0174] More specifically, the innovations of this step include:

[0175] (1) Dual-stream architecture for heterogeneous feature decoupling: By independently encoding urban geometric features and meteorological boundary condition features and fusing them in the deep latent space, the physical semantic dilution of meteorological information in the multi-layer convolution process is avoided, and a true physical boundary condition loading mechanism is established.

[0176] (2) Latent space injection mechanism: Cross-dimensional physical boundary condition loading is achieved through space broadcasting and gating fusion. Functionally, it is completely analogous to the boundary condition driving process in the WRF-LES coupling framework, enabling the deep learning model to learn the global constraint effect of meteorological boundary conditions on the flow field.

[0177] (3) Geometric enhancement of SDF and its gradient: Provides key geometric prior knowledge (building boundary normal and distance information) to the network, significantly improving the prediction accuracy of gradient-sensitive flow phenomena such as flow separation, wake direction, and vortex shedding;

[0178] (4) Composite loss function with physical constraints: The gradient constraint term sharpens the building edge and wake boundary, and the divergence constraint term forces the mass conservation law to be satisfied, effectively solving the problem of excessive smoothing of the flow field caused by single MSE loss, and ensuring that the prediction results conform to the basic laws of fluid mechanics.

[0179] (5) U-Net++ nested dense skip connections: It can better recover high-frequency turbulence details through multi-scale feature fusion mechanism, and has stronger feature reuse capability compared with the simple skip connections of standard U-Net.

[0180] S4. Simulate the micro-scale dynamic field of the entire city.

[0181] The target city area is divided into multiple local computing units using a sliding window technique. For each local computing unit, its street morphology features and corresponding meteorological background vectors are extracted and input into a trained deep learning model to obtain the high-resolution wind field prediction results for the corresponding local computing unit. Finally, through spatial stitching and edge smoothing, the numerical discontinuities at the boundaries of the local computing units are eliminated, and the prediction results of all local computing units are reconstructed into the high-resolution dynamic field of the entire target city.

[0182] In this step, the weighted smoothing stitching specifically includes: using a sliding window technique for extrapolation, setting overlapping regions between adjacent windows; and using a distance-based monotonically decreasing weight function to perform a weighted average of the prediction results for the overlapping regions, with the weight function employing either linear or cosine weights.

[0183] , ;

[0184] in The distance from the point to the window boundary. Where is the width of the overlapping region. The weighting function satisfies... That is, the weight is zero at the window boundary; and That is, the weight is greatest at the boundary of the overlapping area.

[0185] This invention couples a WRF mesoscale meteorological model, enabling the derivation of microscopic flow fields using only the meteorological boundary conditions output by the WRF, without relying on ground observation stations. It possesses true zero-observation cold-start capability and future prediction ability. The WRF-LES multi-scale coupling framework generates training data; each training sample is obtained through high-fidelity LES calculations driven by real WRF meteorological boundary conditions, implicitly containing complete atmospheric dynamic constraints and building geometric boundary perturbations, allowing the model to learn physical laws conforming to the NS equation constraints. A composite loss function incorporating physical constraints is employed, specifically including MSE data fitting terms, gradient constraint terms, and divergence constraint terms. Gradient constraints sharpen building edges and wake boundaries, while divergence constraints enforce mass conservation, ensuring predictions conform to fundamental fluid dynamics principles. A U-Net++ nested dense jump connection structure is employed, and a multi-scale feature fusion mechanism better recovers high-frequency turbulence details. Through a dual-flow architecture with heterogeneous feature decoupling, the physical mechanism of the WRF-LES coupling framework is fully functionally replicated: the meteorological encoding path corresponds to the boundary condition extraction from the WRF output, latent space injection corresponds to the LES solver accepting boundary condition constraints, and the spatial encoding path corresponds to the LES response to building geometric boundary conditions, achieving truly physical-guided deep learning modeling.

[0186] As can be seen from the above embodiments, the deduction method of the present invention first constructs a high-dimensional urban morphological feature space containing multi-dimensional indicators such as building density and sky view factor. Principal component analysis is used to orthogonally reduce the dimensionality of the feature data, and a "centroid-boundary" hybrid sampling strategy is employed to select physically representative sample blocks. Secondly, a high-fidelity flow field simulation is performed using the WRF-LES multi-scale coupling framework. WRF provides realistic atmospheric boundary conditions to drive LES calculations, generating a physically consistent training database, with the high-fidelity LES flow field serving as the supervision target. Then, a deep learning model based on heterogeneous feature decoupling is constructed, and the input... This invention utilizes a multi-channel geometric feature map containing a signed distance function and its spatial gradient. It independently encodes urban geometric features and WRF meteorological boundary condition features, injecting meteorological boundary conditions into a deep latent space. Training is performed using a composite loss function incorporating physical constraints to establish an end-to-end nonlinear mapping from the coupled features of urban geometry and meteorological boundaries to the LES fine flow field. This allows the model to learn the physical laws governing how building geometry perturbs the flow field under given real atmospheric boundary conditions. Finally, a sliding window slicing technique is employed for global extrapolation, reconstructing the full-domain dynamic field of the target city with meter-level resolution through spatial stitching and edge smoothing. This invention balances computational efficiency, prediction accuracy, and physical consistency, enabling rapid response to meter-level resolution wind fields under complex terrain conditions in large cities. It can be widely applied in urban planning, environmental assessment, emergency response, and pollutant dispersion prediction.

[0187] A downstream application model coupling example is shown below:

[0188] 1. Pollutant diffusion simulation: Using full-area wind field data as meteorological input for the atmospheric diffusion model, the transmission and diffusion process of pollutants in the urban environment is simulated;

[0189] 2. Urban ventilation assessment: Utilize wind field data to assess the ventilation efficiency of urban ventilation corridors and identify areas with poor ventilation;

[0190] Emergency Response: In emergency events such as chemical spills or fires, rapidly generated wind field data is used to predict the extent of pollutant spread in real time, providing support for emergency decision-making;

[0191] 3. Wind Energy Resource Assessment: Utilize wind farm data to assess the distribution of wind energy resources in urban areas, providing a reference for wind energy development.

[0192] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and do not constitute a limitation on the scope of protection of the present invention. Equivalent substitutions made to the meteorological model type, network structure, or parameter configuration without departing from the core idea of ​​the present invention should all fall within the scope of protection of the present invention.

Claims

1. A physical mechanism-guided method for heterogeneous extrapolation of meter-level high-resolution dynamic fields of buildings across an entire urban area, characterized in that: Includes the following steps: S1. Based on urban morphological characteristics, select representative blocks to form a sample block set; S2. Use the mesoscale meteorological model WRF to provide a regional-scale meteorological background field for the target city; combine GIS data to extract a fine building geometric model, drive the LES model with atmospheric boundary layer turbulence resolution capability to perform flow field calculations on the sample block set, and generate high spatiotemporal resolution flow field data of the sample blocks. The high spatiotemporal resolution flow field data of each sample block are paired with the corresponding meteorological background data to generate a training database. S3. Construct and train a deep learning model for injecting meteorological information into the latent space; In the deep learning model, an end-to-end nonlinear mapping relationship from the coupling features of urban geometry and meteorological boundaries to the LES fine flow field is established through a dual-path encoding path and a latent space injection mechanism. S4. Divide the target city area into multiple local computing units; for each local computing unit, extract its street morphology features and corresponding meteorological background vector, input them into the trained deep learning model, and obtain the high-resolution wind field prediction results of the corresponding local computing unit; through spatial stitching and edge smoothing, eliminate the numerical discontinuity at the boundary of the local computing unit, and reconstruct the prediction results of all local computing units into the full-area high-resolution dynamic field of the target city.

2. The deduction method according to claim 1, characterized in that, Step S1: Extract multidimensional morphological features of blocks across the entire target city and construct a high-dimensional morphological feature space; perform orthogonalization dimensionality reduction on the high-dimensional morphological feature space and extract key principal components; based on the key principal components, use a hybrid sampling strategy to select centroid samples and boundary samples from each cluster category to construct a set of physically consistent sample blocks.

3. The deduction method according to claim 2, characterized in that, In the high-dimensional morphological feature space, morphological features include building density, street valley height, building coverage, floor area ratio, building dispersion, building differentiation, spatial openness, building height standard deviation, building enclosure, and sky view factor.

4. The deduction method according to claim 2, characterized in that, Centroid and boundary samples are selected from each cluster category to construct a set of physically consistent sample blocks, specifically including: Calculate the feature spatial centroid of all block morphological features within each category; Centroid samples and boundary samples are selected from each category. The centroid samples are the blocks with the smallest Euclidean distance from the centroid of the feature space, which are used to characterize the general geometric features of the urban morphology. The boundary samples are the top 10% of blocks with the largest Euclidean distance from the centroid of the feature space, which are used to characterize the edge features and morphological extremes of the urban morphology. The centroid samples and boundary samples together constitute the sample block set.

5. The deduction method according to claim 1, characterized in that, The specific steps in step S2 to generate high spatiotemporal resolution flow field data include: The regional-scale meteorological field output by the mesoscale meteorological model WRF is used as the boundary and initial conditions of the LES model. Transform detailed architectural geometry models into geometric meshes that can be recognized by LES models; Driven by the boundary and initial conditions, the LES model performs dynamic solutions for each sample block to obtain a three-dimensional velocity field. By combining GIS data to extract detailed building geometric models, the LES model with atmospheric boundary layer turbulence analysis capability is driven to perform three-dimensional unsteady large eddy simulation calculations for atmospheric boundary layer turbulence in each sample block in the sample block set. By numerically solving the spatially filtered incompressible fluid control equations, the dominant turbulence structure at the block scale and above is explicitly analyzed, and the momentum exchange process at the subgrid scale is characterized by a subgrid scale turbulence model, generating high spatiotemporal resolution flow field data. The high spatiotemporal resolution flow field has the spatial resolution to analyze the vortex structure and flow separation characteristics within the street valley.

6. The deduction method according to claim 1, characterized in that, Step S3 includes: S31. Construct a deep learning model for latent space injection of meteorological information. The deep learning model includes a spatial coding path, a meteorological coding path, and a decoding path. The spatial coding path and the meteorological coding path are fused in the deep feature space of the dual-stream deep learning network through a latent space injection mechanism. The decoding path reconstructs a high-resolution wind field based on the fused features. S32. The deep learning model is trained under supervision using the training database generated in step S2. The high spatiotemporal resolution flow field data calculated by the LES model is used as the supervision target. A composite loss function containing physical constraints is used to minimize the difference between the predicted wind field and the target wind field. The composite loss function includes a data fitting term, a gradient constraint term, and a divergence constraint term. The data fitting term is used to ensure the accuracy of flow field prediction, the gradient constraint term is used to sharpen building edges and wake boundaries, and the divergence constraint term is used to force the satisfaction of the incompressible fluid mass conservation law, thereby establishing an end-to-end nonlinear mapping relationship from the coupling characteristics of urban geometry and meteorological boundaries to the LES fine flow field.

7. The deduction method according to claim 6, characterized in that, In step S31, the spatial encoding path adopts a convolutional neural network architecture with multiple downsampling stages, including N downsampling stages. Each downsampling stage includes a convolutional layer, a batch normalization layer, an activation function, and dense connections. The downsampling operation is to halve the spatial resolution of the spatial feature map using stride convolution or max pooling. The input of the spatial encoding path is a multi-channel geometric feature map composed of the building height field, building mask, signed distance function, and its spatial gradient components. The spatial gradient components include the partial derivative of the signed distance function in the horizontal direction, which is used to characterize the normal information of the building boundary. Each downsampling stage outputs feature maps with different spatial resolutions, forming a multi-scale spatial feature representation. The input to the meteorological coding path is the WRF meteorological boundary condition vector. ;vector The time-varying meteorological field output by the WRF model is extracted, including wind speed components, potential temperature and turbulence parameters at different altitudes; the meteorological coding path maps the low-dimensional meteorological boundary vector to a high-dimensional meteorological latent space vector through a multi-layer fully connected transformation. The latent space injection mechanism expands the high-dimensional meteorological latent space vector to the same size as the spatial features through spatial broadcasting operations, and adjusts the influence intensity of meteorological boundary constraints through feature fusion and gating modulation. The latent space injection mechanism is set in the bottleneck layer of the spatial coding path, or in the deep layer of the decoding path, or in both the bottleneck layer of the spatial coding path and multiple deep layers of the decoding path, forming a hierarchical meteorological condition injection. The decoding path adopts a deconvolutional neural network architecture with multiple upsampling stages. Each upsampling stage gradually restores the spatial resolution through transposed convolution or interpolation operations, and receives the corresponding resolution features from the spatial encoding path through dense skip connections. The dense skip connections adopt a nested dense connection structure to pass the features of different levels of the encoder to the corresponding levels of the decoder, forming a feature pyramid for recovering high-frequency turbulence details.

8. The deduction method according to claim 7, characterized in that, The processing steps of the latent space injection mechanism include: Spatial broadcasting expands the meteorological latent space vector output by the meteorological coding path to the same size as the spatial feature map through a spatial copying operation, forming a three-dimensional tensor containing complete meteorological information at each spatial location, thus obtaining the meteorological features after spatial broadcasting. Feature fusion combines meteorological features from space broadcasting with spatial features; Gated modulation, after feature fusion is completed, adaptively adjusts the relative contributions of spatial features and meteorological features through a gated fusion unit: ， ； in For the gating weight graph, It is the sigmoid activation function. This indicates a channel splicing operation. and All are learnable parameters. For spatial features, Meteorological characteristics following space broadcasting As a feature of fusion, This is the characteristic output after gating and modulation.

9. The deduction method according to claim 6, characterized in that, The deep learning model adopts the U-Net++ architecture, which includes a nested dense skip connection structure. The spatial encoding path is the encoder part of the U-Net++ architecture, containing N downsampling stages. Each stage includes a feature extraction module consisting of convolutional layers, batch normalization layers, and activation functions. Residual connections or dense connections are used to enhance feature propagation capabilities. The meteorological encoding path, as an independent fully connected branch, consists of M fully connected layers and is connected to the bottleneck layer of the U-Net++ architecture, mapping the meteorological boundary vector into a latent space vector matching the number of channels in the bottleneck feature map. The dense skip connections pass all feature maps from the encoder stage to the decoder stage, achieving multi-scale feature fusion through upsampling and channel concatenation operations.

10. The deduction method according to claim 6, characterized in that, The specific form of the composite loss function is: ； in, For the data fitting term, the mean square error or mean absolute error between the predicted wind field and the LES target wind field is used, and its mathematical expression is: ,in For wind fields predicted by the network, Target wind field calculated for LES; Indicates the horizontal velocity of the wind field and The various components, This represents the total number of grid points. This is the gradient constraint term, used to calculate the error between the predicted flow field gradient magnitude and the LES flow field gradient. Its mathematical expression is: ,in Represents the spatial gradient operator, Indicates the wind field number Spatial gradient of each component; This is a divergence constraint term used to calculate the horizontal divergence of the predicted flow field. The law of conservation of mass for incompressible fluids is enforced, and its mathematical expression is: ,in Represents the divergence of the horizontal flow field. , These are the two velocity components of the wind field in the horizontal direction. For incompressible fluids, the continuity equation requires that the velocity vector be horizontal; ; , These are all weighting coefficients used to balance the contributions of different loss terms.

Images