A fast prediction method for urban multiscale wind field based on geometric-driven neural operator

Abstract

The application discloses a kind of urban multiscale wind field fast prediction method based on geometric driving neural operator.Firstly, the high-fidelity expression of urban building geometry is carried out using signed distance field (SDF).By extracting multi-scale global geometry code, the input features of the query point containing local geometric features, position information and signed distance field are constructed, and the complex global wind field prediction is transformed into efficient local problem solving.The trained model can directly predict the wind speed amplitude and turbulent kinetic energy of any point in space according to the input building geometry data, and then reconstruct the continuous flow field distribution.The method realizes the decoupling of computational complexity and spatial resolution, has high precision and second-level reasoning speed, can capture building detail flow and street scale wake characteristics, and significantly improves the generalization ability and engineering application value in large-scale complex urban environment.

Description

Technical Field

[0001] This invention belongs to the field of urban wind environment numerical simulation and artificial intelligence technology, specifically involving a geometry-driven neural operator learning method that directly maps extracted building geometric information to the flow field function space. Background Technology

[0002] With the acceleration of urbanization, wind field distribution in complex built environments has a significant impact on pedestrian comfort, pollutant dispersion, and building ventilation performance. Traditional computational fluid dynamics (CFD) methods can obtain high-accuracy wind field results by discretizing and numerically solving the fluid control equations. However, these methods typically rely on complex geometric modeling and mesh generation processes, resulting in high computational costs and making it difficult to meet the needs of rapid evaluation and real-time prediction of multiple design options at the urban scale. Furthermore, when building designs are frequently adjusted, repeated mesh generation and numerical calculations are required, leading to low overall design efficiency.

[0003] In recent years, data-driven methods based on deep learning have been used to accelerate flow field prediction. These methods reduce computational costs to some extent by learning the mapping relationship between input and output. For example, NVIDIA's patent US11,954,791 B2 (2024) proposes a fluid simulation method based on a three-dimensional convolutional network. Its abstract indicates that this method does not require solving the Navier-Stokes equations to enforce the fluid incompressibility condition. Instead, it generates density tensors and rigid body mapping tensors for a large number of particles in a sub-region, serving as input channels for a network with a three-dimensional spatial structure. The network performs a series of operations on the input channels to predict the updated position and update velocity of each particle at the end of the frame, thus handling fluid simulation tasks with tens of millions of particles in a nearly infinitely expanding simulation domain. Further analysis of the specification reveals that this scheme essentially uses a three-dimensional convolutional neural network to tensorize the fluid particles in the sub-domain and uses operations such as dilated convolution and depth-to-space transformation to predict correction factors to simulate fluid dynamics processes. However, such methods typically rely on regular or unstructured meshes as input. On the one hand, methods based on regular grids require voxelization of building geometry, which can easily lead to the loss of geometric boundary information, and it is difficult to balance large-scale scope and local details under limited computing resources. On the other hand, methods based on unstructured grids or graph structures have a certain degree of geometric adaptability, but still rely on the grid generation process, and have high computational complexity in large-scale urban scenarios, which limits their application efficiency.

[0004] To further enhance the model's ability to represent continuous spatial mappings, neural operator methods have emerged. These methods aim to learn the mapping relationship from input functions to output functions, thus overcoming the limitations of traditional discretization representations to some extent. Representative works include the Fourier Neural Operator (FNO), which uses coordinate transformation techniques to process the geometry of compressor blades and parameterizes the operator in frequency space to achieve rapid flow field prediction (see Jiang et al.'s "Prediction of compressor blade cascade flow field based on Fourier neural operator" published in Aerospace Science and Technology in 2025). Furthermore, for efficient solutions to large-scale three-dimensional partial differential equations (PDEs), the abstract of Microsoft's patent US 2023 / 0146819 A1 shows that this scheme downsamples the input data after transforming it to the frequency domain to retain a subset of frequency modes, applies nonlinear transformations to the frequency mode subsets at various stages of the neural network, and then upsamples and inversely transforms it back to the original domain to generate simulation results. Through frequency domain downsampling and nonlinear transformation techniques, the scalability of the neural network in solving large-scale PDEs (such as carbon dioxide injection simulations) is optimized. However, in practical urban wind field prediction tasks, existing neural operator methods still have the following shortcomings: i) The adaptability of geometric representation to scene scale is limited. Although the Geometric Information Neural Operator (GINO) attempts to combine Graph Neural Operator (GNO) and Graph Neural Operator (FNO) to handle changing geometries using the symbolic distance field (SDF) (see the paper "Geometry-Informed Neural Operator for Large-Scale 3D PDEs" published by Li et al. on the preprint platform arXiv in 2023), its computational efficiency and spatial geometric continuity in large-scale urban building cluster scenes still face challenges. ii) Insufficient analytical capability of multi-scale flow features. Although the multi-scale neural operator (MNO) enhances the representation of three-dimensional irregular point cloud flow fields by extracting global dimension compressed attention, local graph attention and micro-point-by-point information in parallel (see the paper "MNO: Multiscale Neural Operator for 3D Computational Fluid Dynamics" published by Wang et al. on the preprint platform arXiv in 2026), existing sampling mechanisms and feature extraction methods often fail to achieve completely "meshless" prediction while ensuring accuracy when facing highly irregular building boundaries and highly dynamic flow field gradient changes in cities.

[0005] The aforementioned limitations affect the prediction accuracy and generalization ability of the model in complex urban environments. Therefore, there is an urgent need for an efficient wind field prediction method that can directly utilize building geometry information, does not rely on gridding, and has the ability to express both global and local flow characteristics, in order to meet the application needs of rapid multi-scale evaluation and optimization design in complex urban environments. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a fast prediction method for multi-scale wind fields in cities based on geometry-driven neural operators. This method can achieve efficient prediction of three-dimensional wind field distribution in complex built environments without the need for complex grid partitioning, meeting the dual requirements of urban planning and engineering design for computational efficiency and prediction accuracy.

[0007] The present invention provides a method for rapid prediction of urban multi-scale wind fields based on geometry-driven neural operators, comprising the following steps: Step S1: A wind field prediction model is obtained through model training. The wind field prediction model takes the building geometry data of the urban building complex as input and outputs the corresponding flow field physical quantities. The flow field physical quantities specifically include wind speed amplitude and turbulent kinetic energy. Step S2: For the building geometry data of the new urban building complex, directly input the trained wind field prediction model and output the corresponding flow field physical quantities. The specific steps are as follows: Step S201: Constructing architectural geometric representation and symbolic distance field: Obtain architectural geometric data of the target area, wherein the architectural geometric data is represented in the form of a three-dimensional mesh or point cloud; Step S202: Extract multi-scale global geometric codes: Extract multi-scale geometric features based on the building geometric data obtained in step S201, and generate global geometric codes to characterize the overall spatial structure; Step S203: Construct the input features for any query point within the computational domain, specifically including the following sub-steps: Step S2031: Construct a local calculation template: For each query point, randomly select multiple reference points within its neighborhood. The query point and the reference points together constitute a local calculation template. Step S2032: Extract local geometric features: Extract the local subdomain features corresponding to the local calculation template from the global geometric code generated in step S202, and use them as the local geometric features of the query point; Step S2033: Construct the location information feature of the query point: Based on the local calculation template in step S2031, obtain the spatial coordinate information of the query point, and perform high-dimensional mapping on the spatial coordinate information through location encoding to construct the location information feature of the query point; Step S2034: Feature Fusion: The local geometric features extracted in step S2032, the location information features constructed in step S2033, and the symbolic distance field at the query point are fused to form the final input feature vector of the query point; Step S204: Wind field prediction: Input the input features of the query point constructed in step S203 into the trained wind field prediction model to realize the mapping from geometric information to wind field and predict the wind speed amplitude and turbulent kinetic energy. Step S205: Spatial wind field reconstruction: Based on the flow field physical quantities predicted by multiple query points, the spatial wind field is continuously reconstructed to generate continuously distributed flow field physical quantities within the computational domain.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: i) By using SDF-based geometric representation, high-fidelity representation of complex building boundaries is achieved, fundamentally eliminating the dependence on mesh generation process and significantly improving the adaptability to any complex geometric shape. ii) By constructing a local computation template, the global wind field prediction problem is transformed into a local subproblem, decoupling computational complexity from spatial resolution, and supporting second-level inference in large-scale urban scenarios while ensuring prediction accuracy; iii) By modeling multi-scale geometric features and fusing global information, it is possible to simultaneously capture the fine flow separation structure at the building edge and the wake organization characteristics at the block scale, thereby improving the accuracy and stability of predictions of key physical quantities such as physical field distribution. iv) Establishes a mapping relationship from geometry to physical fields in a continuous function space, with good cross-scale and cross-layout generalization ability, and supports unified prediction of multiple physical quantities such as wind speed and turbulence parameters. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the geometric structure of an ideal building complex in Embodiment 1 of the present invention.

[0010] Figure 2 This is a schematic diagram of the geometric structure of a real urban building complex in Embodiment 2 of the present invention.

[0011] Figure 3 This is a flowchart illustrating the rapid urban wind field prediction method provided in an embodiment of the present invention.

[0012] Figure 4 This is a schematic diagram of the overall architecture of the wind field prediction model provided in an embodiment of the present invention.

[0013] Figure 5-1 This is a schematic diagram illustrating the multi-scale global geometric coding construction process provided in an embodiment of the present invention.

[0014] Figure 5-2 This is a schematic diagram illustrating the local computation template construction process provided in an embodiment of the present invention.

[0015] Figure 5-3 This is a schematic diagram of the feature aggregation and wind field reconstruction process provided in an embodiment of the present invention.

[0016] Figure 6-1 The typical cross-sectional wind speed amplitude (V) in Embodiment 1 of the present invention mag A cloud map comparing the CFD reference value and the predicted value.

[0017] Figure 6-2 The typical cross-sectional wind speed amplitude (V) in Embodiment 1 of the present invention mag A quantitative comparison chart of CFD reference values ​​and predicted values.

[0018] Figure 7-1 This is a cloud map comparing the CFD reference value and the predicted value of the turbulent kinetic energy (k) of a typical cross section in Embodiment 1 of the present invention.

[0019] Figure 7-2 This is a quantitative comparison chart of the CFD reference value and the predicted value of the turbulent kinetic energy (k) of a typical cross section in Embodiment 1 of the present invention.

[0020] Figure 8-1 The typical cross-sectional wind speed amplitude (V) in Embodiment 2 of the present invention mag A cloud map comparing the CFD reference value and the predicted value.

[0021] Figure 8-2 The typical cross-sectional wind speed amplitude (V) in Embodiment 2 of the present invention mag A quantitative comparison chart of CFD reference values ​​and predicted values.

[0022] Figure 9-1 This is a cloud map comparing the CFD reference value and the predicted value of the turbulent kinetic energy (k) of a typical cross section in Embodiment 2 of the present invention.

[0023] Figure 9-2This is a quantitative comparison chart of the CFD reference value and the predicted value of the turbulent kinetic energy (k) of a typical cross section in Embodiment 2 of the present invention. Detailed Implementation

[0024] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0025] like Figure 3 As shown, the fast prediction method for urban multi-scale wind fields based on geometry-driven neural operators provided by this invention includes the following steps: Step S1: Obtain the wind field prediction model through model training. The wind field prediction model takes the building geometry data of the urban building complex as input and outputs the corresponding flow field physical quantities. The specific process is as follows: Step S101: Obtain architectural geometry data of typical urban building clusters to construct a benchmark sample of building clusters.

[0026] like Figure 1 As shown, in the first embodiment of the present invention, the method for constructing a benchmark sample of a building complex is as follows: constructing an ideal building complex benchmark sample. Specifically, a regular building array is constructed as the basic geometric model, and diverse ideal building complex samples are generated by deleting, adjusting the size, or changing the orientation of some buildings to enhance geometric diversity.

[0027] Specifically, the method for deleting some buildings is to randomly delete about 30% of the buildings to change the floor plan density; The methods for adjusting the size or changing the direction are as follows: stretching approximately 30% of the building vertically to two or three times its original height to enhance height variability, and rotating approximately 30% of the building by 45° to break orthogonal symmetry.

[0028] like Figure 2 As shown, in the second embodiment of the present invention, the method for constructing a benchmark sample of building clusters is as follows: constructing a benchmark sample of real urban building clusters. Specifically, local building cluster samples are extracted from real urban data using a sliding window method, preferably retaining geometrically complete building structures; at the same time, the extracted building clusters can be oriented to further enrich the data distribution.

[0029] Step S102: Based on the building cluster benchmark samples constructed in Step S101, computational fluid dynamics (CFD) numerical simulation is used to perform fluid simulation calculations (specifically referring to the paper "AIJ guidelines for practical applications of CFD to pedestrian wind environment around buildings" published by Tominaga et al. in the Journal of Wind Engineering and Industrial Aerodynamics in 2008). The flow field physical quantities corresponding to each building cluster benchmark sample are obtained, and these flow field physical quantities are defined as CFD reference values. Specifically, these flow field physical quantities include wind speed amplitude (V). mag ) and turbulent kinetic energy (k).

[0030] Step S103, Construct training dataset: Associate the building geometry data obtained in step S101 with the corresponding flow field physical quantities obtained in step S102. Each set of associated building geometry data and flow field physical quantities constitutes a training sample, and multiple training samples constitute a benchmark dataset.

[0031] Step S104: Input the benchmark dataset into the wind field prediction model to be trained. Use building geometry data as input and flow field physical quantities as supervision labels for training, and finally obtain the trained wind field prediction model.

[0032] Among them, such as Figure 4 As shown, the specific process of training the wind field prediction model is as follows: Step S1041, Forward prediction and supervision of the model: Input the benchmark dataset into the wind field prediction model to be trained to obtain the predicted values ​​of the corresponding flow field physical quantities. Calculate the difference between the predicted values ​​and the corresponding CFD reference values ​​through the error metric function to obtain the model error.

[0033] In a preferred embodiment, the error metric function is a volume-weighted mean square error loss function, which simultaneously affects the wind speed amplitude (V). mag Constraints are imposed on spatial volume weights and turbulent kinetic energy (k). This loss function improves the model training stability and the physical consistency of the prediction results by introducing spatial volume weights and co-optimizing multiple physical quantities, thereby enhancing the prediction accuracy of wind fields in complex cities.

[0034] In a preferred embodiment, the step of inputting the benchmark dataset into the wind field prediction model to be trained specifically includes the following steps: Step 1: Extract the global geometric coding field. For each building cluster sample in the benchmark dataset, a symbolic distance field is constructed based on its building geometry data, and multi-scale geometric features are extracted. A feature fusion network with an encoder-decoder structure is used to propagate and aggregate information across different spatial scales, generating a global geometric coding field covering the entire computational domain. This global geometric coding field stores global geometric information about the impact of building layout on wind flow and serves as a repository for subsequent local feature extraction of query points, rather than being a direct input to the prediction model. In a preferred embodiment, the symbolic distance field is an SDF (Specialized Distance Field).

[0035] Step 2: Constructing query point input features and supervision signals. Query points are selected within the computational domain, and a local computational template is constructed for each query point to extract the geometric features of the query point and its neighboring region from the global geometric coding field, serving as local geometric features. Simultaneously, the spatial coordinates of the query point are position-encoded to generate high-dimensional positional information features. The local geometric features, high-dimensional positional information features, and symbolic distance field are combined to form the input features of the query point. Preferably, the input features of the query point can be further constructed to include global operating condition parameters, such as incoming wind speed, wind direction and atmospheric boundary layer feature parameters, and the CFD reference value corresponding to the query point can be used as a supervision signal for model training.

[0036] Step S1042, iterative optimization of model parameters. Based on the model error obtained in step S1041, a loss function is constructed, and an optimization algorithm is used to iteratively update the network parameters of the wind field prediction model; the network parameters include weight parameters and bias parameters. The optimization algorithm can be a second-order optimization algorithm such as Adaptive Momentum Estimation (Adam), AdamW with weight decay, or L-BFGS.

[0037] Step S1043, Model convergence determination. Repeat steps S1041 to S1042 until the model error converges or the preset number of training rounds is reached, and the trained wind field prediction model is obtained.

[0038] In some implementations, sampling strategies or physical constraints can be introduced during training to improve the model's predictive ability and physical consistency in complex regions. Through the above training process, a trained wind field prediction model is obtained.

[0039] Step S2: For the building geometry data of the new urban building complex, directly input it into the trained wind field prediction model. Without the need for mesh generation, the corresponding flow field physical quantities can be output, enabling rapid assessment of the urban outdoor wind field. Combined with... Figure 3 , Figure 4 , Figure 5-1 , Figure 5-2 and Figure 5-3The specific steps are as follows: Step S201: Constructing the architectural geometry representation and symbolic distance field. Acquire architectural geometry data for the target area, represented as a 3D mesh or point cloud. Based on the architectural geometry data, construct a symbolic distance field within a 3D computational domain to describe the geometric relationship between any point in space and the building surface. Preferably, the symbolic distance field is an SDF (Signal Distance Field), used to characterize the signed distance from a spatial point to the building surface and its internal / external positional relationship, thereby achieving a continuous expression of complex geometric boundaries.

[0040] Step S202: Extract multi-scale global geometric codes: Extract multi-scale geometric features based on the building geometric data obtained in step S201, and generate global geometric codes to characterize the overall spatial structure.

[0041] Specifically, such as Figure 5-1 As shown, spherical neighborhoods of different scales are constructed on the building surface or within its neighborhood, where x i Let be any point in space, and let d be the distance from all other points in space to this point. ij This approach captures local details such as building edges through a small-scale neighborhood (radius r1), represents global information such as street block structure through a large-scale neighborhood (radius r2), and propagates information across different spatial scales through a multi-scale feature fusion mechanism (e.g., using a hierarchical feature learning network with an encoder-decoder architecture). This results in a global geometric code representing the impact of the overall urban layout on traffic flow. The extraction and fusion of these multi-scale geometric features can be implemented based on a point cloud multi-scale feature extraction architecture, such as the PointNet++ model proposed by Qi et al. at the Advances in Neural Information Processing Systems conference in 2017.

[0042] Step S203: Construct the input features for any query point within the computational domain, such as... Figure 5-2 As shown. Specifically, it includes the following sub-steps: Step S2031: Construct a local computation template. For each query point, randomly select multiple reference points within its neighborhood. The query point and the reference points together constitute a local computation template. The local computation template serves as a basic computational unit, used to transform the global prediction problem into local sub-problems, thereby decoupling computational complexity from spatial scale.

[0043] Step S2032: Extract local geometric features. Extract local subdomain features corresponding to the local computation template from the global geometric code generated in step S202, as the local geometric features of the query point. Specifically, the global geometric code generated in step S202 is a continuous function covering the entire computation domain, providing local geometric context features for any point in space. Using the spatial coordinates of each point in the local computation template constructed in step S2031 as an index, perform point-by-point sampling in the global geometric code to obtain a set of high-dimensional feature vectors corresponding to the query point and its neighboring reference points, as the local geometric features of the query point (e.g., ...). Figure 5-3 The 192-channel local geometric feature shown here, where 192 represents the number of information channels of the feature.

[0044] Step S2033: Construct the location information feature of the query point. To supplement the spatial information in the local geometric features of the query point, based on the local calculation template of step S2031, the spatial coordinate information of the query point is obtained, and the spatial coordinate information is mapped in a high dimension through position encoding to construct the location information feature of the query point.

[0045] Preferably, the original 3D spatial coordinates (x, y, z) can be transformed into 512-dimensional high-dimensional feature (i.e., 512 channels) spatial coordinate information, such as... Figure 5-3 As shown.

[0046] Step S2034: Feature Fusion. The local geometric features extracted in step S2032, the location information features constructed in step S2033, and the symbolic distance field at the query point are fused to form the final input feature vector of the query point. For example... Figure 5-3 As shown.

[0047] Preferably, the input features can be further integrated with global operating condition parameters to characterize external incoming flow conditions, such as incoming wind speed, wind direction, and atmospheric boundary layer characteristic parameters, so that the model can be applied to different operating conditions.

[0048] Step S204: Wind Field Prediction. The input features of the query point constructed in step S203 are input into the trained wind field prediction model to achieve the mapping from geometric information to the wind field, and to predict the wind speed amplitude (V). mag The prediction is made using the kinetic energy (k) and turbulent kinetic energy (k).

[0049] like Figure 5-3 As shown, in some implementations, the wind field prediction model may include a multilayer perceptron (MLP) feature aggregation network to fuse information from multiple reference points within a local computation template and output the flow field physical quantities at the query point.

[0050] Step S205: Spatial wind field reconstruction. Based on the flow field physical quantities predicted from multiple query points, the spatial wind field is continuously reconstructed to generate continuously distributed flow field physical quantities within the computational domain, thereby achieving a continuous representation of the wind field across the entire computational domain.

[0051] Preferably, the predicted values ​​of discrete query points can be interpolated and fused using the inverse distance weighting (IDW) method to achieve continuous reconstruction of the spatial wind field, thereby generating continuously distributed flow field physical quantities and realizing continuous characterization of the wind environment in the entire computational domain.

[0052] Using the above method, during the model inference stage, only new urban building geometry data and optional operating parameters need to be input to obtain the spatially continuous three-dimensional wind field distribution within seconds, enabling rapid gridless assessment of complex urban environments.

[0053] Example To verify the effectiveness and applicability of the method of the present invention, two typical scenarios are selected as examples for illustration, namely Example 1 (ideal building complex) and Example 2 (real urban building complex).

[0054] In this embodiment, all buildings are placed within a computational domain with a flow length of 300 m, a lateral width of 200 m, and a vertical height of 100 m. The geometric center of the building complex is located at the center of the computational domain. The incoming flow conditions are set according to the atmospheric boundary layer, using a logarithmic wind profile, where the reference wind speed at a height of 10 m is 10 m / s, the ground roughness length is 1 m, and the von Kármán constant is taken as 0.40-0.41.

[0055] Example 1: Ideal Building Complex like Figure 1 As shown, Example 1 is an ideal building complex generated based on a regular building array and through parametric perturbation. Specifically, a 9×9 regularly arranged, equally spaced cubic building array is used as a reference, with a spacing of 5m between adjacent buildings. Based on this, the following three types of geometric perturbations are introduced to enhance sample diversity: approximately 30% of the buildings are randomly deleted to change the planar density; of the remaining buildings, approximately 30% are vertically stretched to two or three times their original height to enhance height variability; and approximately 30% of the buildings are randomly selected and rotated 45° to break orthogonal symmetry. Through these perturbations, the initial regular array evolves into a complex building complex configuration with multi-scale geometric features.

[0056] like Figure 6-1 and Figure 7-1 As shown, the model in Example 1 is used to represent the wind speed amplitude (V) magA contour plot comparing the CFD reference values ​​and predicted values ​​of tidal energy (k) and turbulent kinetic energy (k). To comprehensively evaluate the model performance, two typical sections were selected within the computational domain: a longitudinal section at y=0 m and a horizontal section at z=2 m (pedestrian height). Sampling lines were set along the vertical line at the building center and the horizontal line at z=2 m to compare and analyze the model's predicted values ​​with the CFD reference values. From the perspective of spatial distribution characteristics, the model can accurately reconstruct the flow field structure. In the vertical direction, it successfully captured the development process of the atmospheric boundary layer and the characteristics of the shear layer at the top of the building; in the horizontal direction, it can meticulously depict the narrow-channel acceleration effect and the leeward wake structure inside the building array.

[0057] Figure 6-2 and Figure 7-2 The model in Example 1 is shown for wind speed amplitude (V) mag A quantitative comparison of CFD reference and predicted values ​​for V and turbulent kinetic energy (k). The quantitative results show that at V mag For prediction, the mean absolute error (MAE) for the y-section is 0.264 m / s, and the root mean square error (RMSE) is 0.405 m / s; for the z-section, the MAE is 0.253 m / s, and the RMSE is 0.382 m / s. For k-prediction, the MAE for the y-section is 0.239 m² / s², and the RMSE is 0.364 m² / s², with similar error levels for the z-section.

[0058] Example 2: Real Urban Building Complex like Figure 2 As shown, Example 2 illustrates a complex building layout extracted from a real urban area. The geometric prototype of this scenario originates from real urban morphology, exhibiting significant spatial irregularities and authentic urban characteristics. To construct training and testing samples, a sliding window sampling strategy is employed to extract local building clusters from a large-scale urban dataset: a cubic sampling window with a side length of 100 m is used to scan the urban geometry at 50 m steps, and geometric integrity filtering is applied to buildings within the window, retaining only those completely within the window. To further enrich the data distribution, each extracted sub-region is rotated in a plane at 30° intervals, generating multiple geometric instances with different orientations.

[0059] like Figure 8-1 and Figure 9-1 As shown, the model in Example 2 demonstrates the effect of wind speed amplitude (V) mag A contour plot comparing the CFD reference values ​​and predicted values ​​of tidal energy (k). Compared to regular building clusters, this scenario has a more complex geometry, with more severe flow separation and shear layer structures at the building edges. Overall, the model can predict the dominant flow patterns around the buildings relatively well, including key features such as inflow around the buildings, wake separation, and local acceleration.

[0060] Figure 8-2 and Figure 9-2 The model in Example 2 is shown for wind speed amplitude (V) mag A quantitative comparison of CFD reference and predicted values ​​for V and turbulent kinetic energy (k). The quantitative analysis results show that at V mag In terms of prediction, the MAE for the y-section is 0.297 m / s, and the RMSE is 0.648 m / s; the MAE for the z-section is 0.258 m / s, and the RMSE is 0.490 m / s. For the k-section prediction, the MAE for the y-section is 0.194 m² / s², and the RMSE is 0.332 m² / s²; the MAE for the z-section is 0.185 m² / s², and the RMSE is 0.363 m² / s². Comparison of the sampling lines shows that the model predictions generally agree well with the CFD reference values. In areas obstructed by buildings, the model effectively reflects the drastic changes in velocity; in the wake recovery region, the prediction results exhibit high smoothness and numerical stability, with the overall error controlled at a low level.

[0061] Based on the test results of Examples 1 and 2, the method of the present invention achieves rapid flow field prediction within an acceptable engineering accuracy range for both idealized building arrays and real urban layouts without the need for mesh generation. In terms of inference speed, this method improves upon traditional steady-state RANS numerical simulation by approximately two orders of magnitude, enabling the prediction of flow field physical quantities for a single scenario within seconds. These results verify the effectiveness, efficiency, and scalability of the method of the present invention in the assessment of complex urban wind fields.

[0062] The above description is only a preferred embodiment of the present invention. Those skilled in the art can make various modifications or substitutions without departing from the technical principles of the present invention, and all such equivalent transformations should fall within the protection scope of the present invention.

Claims

1. A fast prediction method for urban multi-scale wind fields based on geometrically driven neural operators, characterized in that... Includes the following steps: Step S1: A wind field prediction model is obtained through model training. The wind field prediction model takes the building geometry data of the urban building complex as input and outputs the corresponding flow field physical quantities. The flow field physical quantities specifically include wind speed amplitude and turbulent kinetic energy. Step S2: For the building geometry data of the new urban building complex, directly input the trained wind field prediction model and output the corresponding flow field physical quantities. The specific steps are as follows: Step S201: Constructing architectural geometric representation and symbolic distance field: Obtain architectural geometric data of the target area, wherein the architectural geometric data is represented in the form of a three-dimensional mesh or point cloud; Step S202: Extract multi-scale global geometric codes: Extract multi-scale geometric features based on the building geometric data obtained in step S201, and generate global geometric codes to characterize the overall spatial structure; Step S203: Construct the input features for any query point within the computational domain, specifically including the following sub-steps: Step S2031: Construct a local calculation template: For each query point, randomly select multiple reference points within its neighborhood. The query point and the reference points together constitute a local calculation template. Step S2032: Extract local geometric features: Extract the local subdomain features corresponding to the local calculation template from the global geometric code generated in step S202, and use them as the local geometric features of the query point; Step S2033: Construct the location information feature of the query point: Based on the local calculation template in step S2031, obtain the spatial coordinate information of the query point, and perform high-dimensional mapping on the spatial coordinate information through location encoding to construct the location information feature of the query point; Step S2034: Feature Fusion: The local geometric features extracted in step S2032, the location information features constructed in step S2033, and the symbolic distance field at the query point are fused to form the final input feature vector of the query point; Step S204: Wind field prediction: Input the input features of the query point constructed in step S203 into the trained wind field prediction model to realize the mapping from geometric information to wind field and predict the wind speed amplitude and turbulent kinetic energy. Step S205: Spatial wind field reconstruction: Based on the flow field physical quantities predicted by multiple query points, the spatial wind field is continuously reconstructed to generate continuously distributed flow field physical quantities within the computational domain.

2. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 1, characterized in that: By interpolating and fusing the predicted values ​​of discrete query points using the inverse distance weighting method, a continuous reconstruction of the spatial wind field is achieved, thereby generating continuously distributed flow field physical quantities and realizing a continuous characterization of the wind environment in the entire computational domain.

3. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 1, characterized in that: The symbolic distance field is SDF, which is used to characterize the signed distance from a spatial point to the building surface and its internal and external positional relationships.

4. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 1, characterized in that: The input features are further integrated with global operating parameters to characterize external inflow conditions.

5. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to any one of claims 1-4, characterized in that: Step S1 specifically includes the following steps: Step S101: Obtain architectural geometry data of typical urban building clusters to construct a benchmark sample of building clusters; Step S102: Based on the building cluster benchmark sample constructed in step S101, the computational fluid dynamics numerical simulation method is used to perform fluid simulation calculations to obtain the flow field physical quantities corresponding to each building cluster benchmark sample, and the flow field physical quantities are defined as CFD reference values. The flow field physical quantities specifically include wind speed amplitude and turbulent kinetic energy. Step S103, Construct training dataset: Associate the building geometry data obtained in step S101 with the corresponding flow field physical quantities obtained in step S102. Each set of associated building geometry data and flow field physical quantities constitutes a training sample, and multiple training samples constitute a benchmark dataset. Step S104: Input the benchmark dataset into the wind field prediction model to be trained, using building geometry data as input and flow field physical quantities as supervision labels for training, and finally obtain the trained wind field prediction model. The specific process of training the wind field prediction model is as follows: Step S1041, Forward Prediction and Supervision of the Model: Input the benchmark dataset into the wind field prediction model to be trained to obtain the predicted values ​​of the corresponding flow field physical quantities. Calculate the difference between the predicted values ​​and the corresponding CFD reference values ​​using an error metric function to obtain the model error. Inputting the benchmark dataset into the wind field prediction model to be trained specifically includes the following steps: Step 1: Extract the global geometric coding field: For each building cluster sample in the benchmark dataset, construct a symbolic distance field based on its building geometry data and extract multi-scale geometric features; by using an encoder-decoder structured feature fusion network to propagate and aggregate information across different spatial scales, a global geometric coding field covering the entire computational domain is generated. Step 2: Constructing query point input features and supervision signals: Select query points within the computational domain and construct a local computational template for each query point to extract the geometric features of the query point and its neighboring region from the global geometric coding field, as local geometric features; simultaneously, perform position encoding on the spatial coordinates of the query point to generate high-dimensional position information features; combine the local geometric features, high-dimensional position information features, and symbolic distance field to form the input features of the query point; Step S1042, Iterative optimization of model parameters: Based on the model error obtained in step S1041, a loss function is constructed, and an optimization algorithm is used to iteratively update the network parameters of the wind field prediction model; Step S1043, Model convergence determination: Repeat steps S1041 to S1042 until the model error converges or the preset number of training rounds is reached, and the trained wind field prediction model is obtained.

6. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 5, characterized in that: The error metric function is a volume-weighted mean square error loss function, which simultaneously imposes constraints on wind speed amplitude and turbulent kinetic energy.

7. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 5, characterized in that: The method for constructing the benchmark sample of the building group in step S101 is as follows: Constructing the benchmark sample of the ideal building group specifically involves: constructing a regular building array as the basic geometric model, and generating diverse ideal building group samples by deleting, adjusting the size or changing the direction of some buildings to enhance geometric diversity.

8. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 7, characterized in that: The method for deleting some buildings is to randomly delete approximately 30% of the buildings to change the floor plan density; The methods for adjusting the size or changing the direction are as follows: stretching approximately 30% of the building vertically to two or three times its original height to enhance height variability, and rotating approximately 30% of the building by 45° to break orthogonal symmetry.

9. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 5, characterized in that: The method for constructing benchmark samples of building clusters is as follows: construct benchmark samples of real urban building clusters.

10. The method for rapid prediction of urban multi-scale wind fields based on geometrically driven neural operators according to claim 9, characterized in that: Local building cluster samples are extracted from real urban data using a sliding window approach, with priority given to preserving geometrically intact building structures. Simultaneously, the extracted building clusters are oriented to further enrich the data distribution.

Images