A deep learning-based urban low-altitude wind field prediction method and related device

Abstract

The embodiment of the application provides a kind of based on deep learning's city low-altitude wind field prediction method and related equipment, belong to wind field prediction technical field.The method comprises: obtaining the three-dimensional building contour data and wind speed and direction meteorological data of city target area;According to three-dimensional building contour data, the single or multiple city low-altitude wind field prediction target points in 0-120 meter height range are respectively carried out three-dimensional cylindrical building geometry feature reading convolution calculation, and the feature vector of each prediction target point is obtained;Based on wind speed and direction meteorological data, adjust feature vector order, and input to the trained deep learning wind field prediction model to carry out wind field prediction, and obtain the initial wind field data output of single or multiple prediction target points;Multiple initial wind field data outputs are synthesized, and the three-dimensional wind field prediction of city target area is obtained.The embodiment of the application can improve the efficiency of city low-altitude wind field prediction.

Description

Technical Field

[0001] This application relates to the field of wind field prediction technology, and in particular to a deep learning-based method and related equipment for predicting urban low-altitude wind fields. Background Technology

[0002] Accurate prediction of high-resolution three-dimensional wind speed and turbulence distribution in urban built environments is crucial for fields such as urban ventilation assessment, building energy consumption simulation, and pollutant diffusion control. Computational fluid dynamics (CFD) methods, such as Reynolds-averaged simulation (RANS) and large eddy simulation (LES), can provide relatively accurate results, but their huge computational costs and time overhead make them difficult to apply to scenarios requiring rapid response, such as urban planning and wind disaster early warning.

[0003] To overcome the efficiency bottleneck of CFD methods, data-driven deep learning models are considered an effective way to achieve rapid urban wind field prediction. The underlying idea of ​​these technologies is to directly map wind speed distribution onto the model by inputting feature data representing urban building morphology. One such approach is a model based on manually designed parameters: using methods such as k-nearest neighbor regression, relying on a series of predefined geometric parameters (such as distances to neighboring buildings, street width, etc.) for wind speed prediction.

[0004] Models based on manually designed geometric parameters are highly dependent on the rationality of the selected parameters and the accuracy of the calculations. In complex urban environments, these parameters are often vaguely defined and computationally complex, introducing significant uncertainties and errors, which severely restricts the efficiency of wind field prediction.

[0005] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0006] The main objective of this application is to propose a deep learning-based method and related equipment for predicting urban low-altitude wind fields, which can improve the efficiency of wind field prediction.

[0007] To achieve the above objectives, one aspect of this application proposes a deep learning-based method for predicting urban low-altitude wind fields, the method comprising:

[0008] Acquire 3D building outline data and wind speed, direction and meteorological data for the target urban area;

[0009] Based on the three-dimensional building outline data, three-dimensional cylindrical building geometric feature reading convolution calculation is performed on one or more predefined prediction target points to obtain the feature vector of each prediction target point;

[0010] The feature vector is adjusted based on the wind speed and wind direction meteorological data to obtain the adjusted feature vector;

[0011] The adjusted feature vector is input into the trained urban wind field prediction model to predict the wind field, and the initial wind field data of one or more prediction target points is output.

[0012] The initial wind field data outputs are combined to obtain a three-dimensional wind field prediction for the target urban area.

[0013] In some embodiments, the step of performing three-dimensional cylindrical building geometric feature reading convolution calculations on predefined single or multiple prediction target points based on the three-dimensional building outline data to obtain the feature vector of each prediction target point includes:

[0014] Using the predicted target point as the spatial origin of the three-dimensional building outline data, a cylindrical coordinate system is constructed in the surrounding space;

[0015] The surrounding space is divided into multiple partitions based on the three dimensions of the cylindrical coordinate system.

[0016] The feature values ​​of the partition are obtained by calculating the feature values ​​of the partition.

[0017] The feature vector is obtained by constructing a feature vector from multiple feature values.

[0018] In some embodiments, the spatial division of the surrounding space according to the three dimensions of the cylindrical coordinate system to obtain multiple partitions includes:

[0019] The surrounding space is divided according to the radial direction of the cylindrical coordinate system to obtain multiple annular spaces;

[0020] The annular space is divided into multiple first spaces by sectoring the cylindrical coordinate system according to the azimuth angle.

[0021] The first space is divided into multiple partitions based on the vertical height of the cylindrical coordinate system.

[0022] In some embodiments, the step of calculating feature values ​​for the partition to obtain feature values ​​for the partition includes:

[0023] The volume of the partition is calculated to obtain the total volume;

[0024] The volume of the building space occupied by the building within the partition is calculated to obtain the volume of the building space occupied by the partition.

[0025] The building volume occupancy rate is calculated based on the total volume and the building's occupied space volume.

[0026] The building volume occupancy rate is used as the characteristic value.

[0027] In some embodiments, constructing a feature vector from the plurality of feature values ​​to obtain the feature vector includes:

[0028] The feature vector is obtained by sorting the multiple feature values ​​according to the vertical height index, the radius direction index, and the azimuth index.

[0029] In some embodiments, adjusting the feature vector based on the wind speed and direction meteorological data to obtain the adjusted feature vector includes:

[0030] The azimuth index order of the feature vector is adjusted based on the wind direction data in the wind speed and wind direction meteorological data to obtain a new feature value sequence;

[0031] The positions of the feature values ​​in the feature vector are adjusted according to the feature value sequence to obtain the adjusted feature vector.

[0032] In some embodiments, the trained deep learning urban wind field prediction model is obtained through the following steps:

[0033] Obtain datasets of wind speed components and turbulence parameters from large eddy simulations in different urban simulation areas;

[0034] Data samples were selected from the large eddy simulation wind speed components and turbulence parameter dataset.

[0035] The feature vector of the data sample is calculated by convolution based on the geometric features of a three-dimensional cylindrical building.

[0036] The feature vector is input into the deep learning urban wind field prediction model for prediction, and the wind speed component and turbulence parameters are output.

[0037] Based on the simulation data of the large eddy simulation wind speed component and turbulence parameter dataset, the error is calculated with the wind speed component and turbulence parameter respectively to obtain their respective mean square errors;

[0038] With the goal of minimizing the mean square error, the deep learning urban wind field prediction model is trained and optimized to obtain the trained deep learning urban wind field prediction model.

[0039] To achieve the above objectives, another aspect of this application proposes a deep learning-based urban low-altitude wind field prediction device, the device comprising:

[0040] The acquisition module is used to acquire 3D building outline data and wind speed, wind direction and meteorological data of the target urban area;

[0041] The calculation module is used to perform three-dimensional cylindrical building geometric feature reading convolution calculation on one or more predefined prediction target points according to the three-dimensional building outline data, so as to obtain the feature vector of each prediction target point;

[0042] The adjustment module adjusts the feature vector based on the wind speed and direction meteorological data to obtain the adjusted feature vector;

[0043] The prediction module is used to input the adjusted feature vector into the trained deep learning urban wind field prediction model to predict the wind field and output the initial wind field data of one or more prediction target points.

[0044] The synthesis module is used to synthesize multiple initial wind field data outputs to obtain a three-dimensional wind field prediction for the target urban area.

[0045] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0046] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0047] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0048] The embodiments of this application include at least the following beneficial effects: This application provides a method, device, electronic device, storage medium, and program product for predicting urban low-altitude wind fields based on deep learning. This solution is based on three-dimensional building contour data and performs three-dimensional cylindrical building geometric feature reading convolution calculation on multiple prediction target points to obtain the feature vector of each prediction target point. The three-dimensional cylindrical building geometric feature reading convolution can extract the building morphology features in three-dimensional space more directly and naturally, realize feature extraction without relying on manual design parameters, improve the generalization ability of subsequent wind field prediction models in different urban environments through more essential geometric representation, reduce the errors and uncertainties introduced by manual design parameters, and thus improve the efficiency of wind field prediction. Attached Figure Description

[0049] Figure 1 This is a flowchart of the deep learning-based urban low-altitude wind field prediction method provided in the embodiments of this application;

[0050] Figure 2This is a schematic diagram of the structure of the convolution for reading the geometric features of a three-dimensional cylindrical building provided in this application embodiment;

[0051] Figure 3 This is a schematic diagram of the structure of the deep learning urban wind field prediction model provided in the embodiments of this application;

[0052] Figure 4 This is a complete flowchart of the deep learning-based urban low-altitude wind field prediction method provided in the embodiments of this application;

[0053] Figure 5 This is a comparison chart and scatter plot of the flow direction and wind speed results provided in the embodiments of this application, showing the prediction results of the deep learning model and the results of the large eddy simulation.

[0054] Figure 6 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0057] Accurate prediction of high-resolution three-dimensional wind speed and turbulence distribution in urban built environments is crucial for fields such as urban ventilation assessment, building energy consumption simulation, and pollutant diffusion control. Computational fluid dynamics (CFD) methods, such as Reynolds-averaged simulation (RANS) and large eddy simulation (LES), can provide relatively accurate results, but their huge computational costs and time overhead make them difficult to apply to scenarios requiring rapid response, such as urban planning and wind disaster early warning.

[0058] To overcome the efficiency bottleneck of CFD methods, data-driven deep learning models are considered an effective way to achieve rapid prediction of urban wind fields. The underlying idea of ​​these technologies is to directly map the wind speed distribution in the city onto the input feature data representing urban building morphology, primarily following the following two technical paths:

[0059] Two-dimensional grid-based model: Using image segmentation model architectures such as U-Net, it uses two-dimensional features based on Cartesian grids, such as building distance field and height field, as input to predict the two-dimensional wind speed distribution at a specific height plane.

[0060] Models based on manually designed parameters: These models use methods such as k-nearest neighbor regression to predict wind speed at specific locations based on a series of predefined geometric parameters (such as distances to nearby buildings, street widths, etc.).

[0061] While the aforementioned deep learning solutions have shown potential in rapidly predicting urban wind fields, their inherent limitations at the data representation level result in significant deficiencies in predictive power, efficiency, and generalization.

[0062] i. Insufficient 3D modeling capability: Models based on 2D image segmentation architectures such as U-Net essentially output a 2D planar distribution, which cannot directly generate physically consistent 3D wind fields end-to-end. The method of piecing together 3D results by training independent models for different heights is inefficient, and due to the lack of coordination between models, it is difficult to guarantee the continuity of the wind field in the vertical direction.

[0063] ii. Rigid data architecture and poor wind direction adaptability: Models based on fixed Cartesian grids have input data that is strongly coupled with specific grid orientation and size. This necessitates time-consuming and tedious preprocessing such as grid rotation and clipping when wind direction changes, resulting in extremely low efficiency for long-term forecasts or multi-scenario analysis, and failing to meet the application requirements for rapid response.

[0064] iii. Reliance on manually designed parameters and weak generalization ability: The performance of models based on manually designed geometric parameters is highly dependent on the rationality of the selected parameters and the accuracy of the calculation. These parameters are often vaguely defined and computationally complex in complex urban environments, introducing significant uncertainties and errors, which limit the model's generalization ability and robustness in new environments, and severely restrict the efficiency and accuracy of wind field prediction.

[0065] In view of this, this application provides a deep learning-based method and related equipment for predicting urban low-altitude wind fields. This method performs convolutional calculations on predefined single or multiple prediction target points based on 3D building outline (including height) data, using 3D cylindrical building geometric feature reading, to obtain feature vectors for each prediction target point. The 3D cylindrical building geometric feature reading convolution can more directly and naturally extract building morphological features in 3D space, achieving feature extraction independent of manual design parameters. Through more fundamental geometric representation, it improves the generalization ability of subsequent wind field prediction models in different urban environments, reduces errors and uncertainties introduced by manual design parameters, and thus improves the efficiency and accuracy of wind field prediction.

[0066] The deep learning-based urban low-altitude wind field prediction method provided in this application relates to the field of wind field prediction technology. This deep learning-based urban low-altitude wind field prediction method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, or desktop computer, etc., but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the deep learning-based urban low-altitude wind field prediction method, etc., but is not limited to the above forms.

[0067] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0068] Figure 1 This is an optional flowchart of a deep learning-based urban low-altitude wind field prediction method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S105.

[0069] Step S101: Obtain three-dimensional building outlines (including height) and wind speed and direction meteorological data for the target urban area;

[0070] Step S102: Based on the 3D building outline data, perform 3D cylindrical building geometric feature reading and convolution calculation on one or more predefined prediction target points to obtain the feature vector of each prediction target point;

[0071] Step S103: Adjust the feature vector according to the meteorological data of wind speed and direction to obtain the adjusted feature vector;

[0072] Step S104: Input the adjusted feature vector into the trained deep learning urban wind field prediction model to predict the wind field and obtain the initial wind field data output for one or more prediction target points.

[0073] Step S105: Combine multiple initial wind field data outputs to obtain a three-dimensional wind field prediction for the target urban area.

[0074] Steps S101 to S105 of this embodiment, through the Geometry Reading Filter (GRF), can more directly and naturally extract the morphological features of buildings in three-dimensional space. This achieves feature extraction independent of manual design parameters, improves the generalization ability of the subsequent wind field prediction model in different urban environments through more fundamental geometric representation, reduces the errors and uncertainties introduced by manual design parameters, and improves the efficiency of urban wind field prediction. The wind field prediction model uses a single point as the prediction unit and can be applied to any new urban area without retraining, breaking through the limitations of fixed grids. Utilizing the rotational symmetry of the cylindrical coordinate system, only the GRF feature vector needs to be adjusted to adapt to different wind directions, thereby reducing the time-consuming grid reconstruction process and improving the efficiency of long-term forecasting and multi-scenario analysis by orders of magnitude.

[0075] In step S102 of some embodiments, please refer to Figure 2 The structure of convolution is read based on the geometric features of a three-dimensional cylindrical building, such as... Figure 2 As shown in (a) and (b), only one GRF calculation is needed for a specific urban target area. First, the 3D building outline data of the urban target area is input (e.g., a common format such as Shapefile). Then, the 3D spatial target point set P_i for wind speed prediction within this area is defined. The target point set P_i can be regularly distributed (e.g., a 3D Cartesian grid) or arbitrarily discrete, with its height range preferably covering 0 to 120 meters and a spatial resolution up to 5 meters. Next, the target point set is traversed, and the corresponding 3D cylindrical GRF feature vector is calculated for each target point P_i. After the calculation is complete, the GRF feature vectors of all target points are stored in a database for repeated use by subsequent wind field prediction models.

[0076] In some embodiments, step S102 may include, but is not limited to, steps S201 to S204:

[0077] Step S201: Using the predicted target point as the spatial origin of the 3D building outline data, construct a cylindrical coordinate system to the surrounding space;

[0078] Step S202: Divide the surrounding space according to the three dimensions of the cylindrical coordinate system to obtain multiple partitions;

[0079] Step S203: Calculate the feature values ​​of the partitions to obtain the feature values ​​of the partitions;

[0080] Step S204: Construct eigenvectors from multiple eigenvalues ​​to obtain eigenvectors.

[0081] In step S201 of some embodiments, GRF constructs a local cylindrical coordinate system with P_i as the spatial origin, and divides the space around the point along three dimensions to facilitate subsequent partitioning.

[0082] In step S202 of some embodiments, the three dimensions of the cylindrical coordinate system include the radial direction (R direction), the azimuth angle (θ direction), and the vertical height (Z direction).

[0083] In some embodiments, step S202 may include, but is not limited to, steps S221 to S223:

[0084] Step S221: Divide the surrounding space according to the radial direction of the cylindrical coordinate system to obtain multiple annular spaces;

[0085] Step S222: Divide the annular space into sectors based on the azimuth angle of the cylindrical coordinate system to obtain multiple first spaces;

[0086] Step S223: Divide the first space into multiple partitions based on the vertical height of the cylindrical coordinate system.

[0087] In step S221 of some embodiments, the space is divided into N_r non-overlapping annular spaces along the radial direction. For example, N_r is set to 10, and the boundaries of the intervals are [0, 20], [20, 40], ..., [340, 400] (unit: meters).

[0088] In step S222 of some embodiments, for the initially divided annular space, the 360-degree range is evenly divided into N_a sector regions on the horizontal plane, thereby obtaining multiple first spaces. For example, N_a=16, and each sector covers 22.5 degrees.

[0089] In step S223 of some embodiments, the first space is divided into N_z layers along the height direction, thereby obtaining multiple partitions. For example, N_z is set to 10, and its layer boundaries are [-70, -40], [-40, -20], [-20, -10], [-10, -5], [-5, 0], [0, 5], [5, 10], [10, 20], [20, 40], [40, 70] (unit: meters).

[0090] Based on the above division, GRF divides the space around the predicted point into a total of N = N_r × N_a × N_z partitions C_ijk (where i, j, and k are the indices of the radial, azimuth, and height directions, respectively). Preferably, the total number of partitions N = 1600. Partitioning facilitates the subsequent calculation of feature values, and the indexes allow for the sorting of feature values, thereby constructing the feature vector.

[0091] In step S203 of some embodiments, for each partition C_ijk, its eigenvalue is calculated to achieve refined calculation of the eigenvalue.

[0092] In some embodiments, step S203 may include, but is not limited to, steps S231 to S234:

[0093] Step S231: Calculate the volume of the partition to obtain the total volume;

[0094] Step S232: Calculate the volume of the building space occupied within the partition to obtain the volume of the building space occupied by the partition.

[0095] Step S233: Calculate the volume occupancy rate based on the total volume and the building's occupied space volume to obtain the building volume occupancy rate;

[0096] Step S234: Use the building volume occupancy rate as a feature value.

[0097] In steps S231 to S234 of some embodiments, for each partition C_ijk, its Building Volume Fraction (BVF) is calculated as a characteristic value of that partition. The formula for calculating BVF is defined as follows:

[0098] BVF_ijk=V_building,ijk / V_total,ijk;

[0099] Where V_building,ijk is the building volume of partition C_ijk, and V_total,ijk is the total volume of partition C_ijk. Therefore, BVF_ijk is a dimensionless number between 0 and 1.

[0100] When C_ijk is completely inside the building, BVF_ijk=1;

[0101] When C_ijk is entirely in outdoor space, BVF_ijk=0;

[0102] When C_ijk contains both buildings and air, 0 <BVF_ijk<1。

[0103] In some embodiments, step S204 may include, but is not limited to, step S401:

[0104] Step S401: Sort multiple feature values ​​according to the vertical height index, radius direction index, and azimuth index to obtain the feature vector.

[0105] In step S401 of some embodiments, after calculating the BVF_ijk values ​​for all N partitions, these values ​​are arranged in a certain order (e.g., first fixing the height index k and radial index i, then traversing all azimuth indices j) to generate an N-dimensional feature vector. This feature vector is a digital, fixed-dimensional representation of the three-dimensional building morphology around the target point P_i.

[0106] The Geometric Feature Reading Convolution (GRF) for 3D cylindrical buildings can efficiently extract morphological features directly from 3D building models without relying on specific wind directions and coordinate systems. GRF extracts features based on the basic geometric volume occupancy rate and does not rely on any artificially preset, potentially uncertain, complex parameters. It has stronger adaptability and robustness to different urban forms and layouts, laying the foundation for achieving end-to-end true 3D wind field prediction.

[0107] In some embodiments, step S103 may include, but is not limited to, steps S501 to S502:

[0108] Step S501: Adjust the azimuth index order of the feature vectors according to the wind direction data in the wind speed and wind direction meteorological data to obtain a new feature value sequence.

[0109] Step S502: Adjust the position of the eigenvalues ​​in the eigenvector according to the eigenvalue sequence to obtain the adjusted eigenvector.

[0110] In steps S501 to S502 of some embodiments, the wind speed and direction meteorological data includes the wind speed and direction at a known altitude (e.g., 120 meters). This data can be provided by observational data or by a mesoscale weather forecasting model (e.g., WRF). Subsequently, for each target point P_i defined in the pre-calculation stage, a loop operation is performed: First, the pre-stored GRF feature vector for that target point is read from a preset database. This feature vector is divided into 16 uniform regions along the azimuth (θ direction), corresponding to 16 different wind directions. For example, if the pre-stored GRF(r_i, θ_j, z_k) is constructed based on westerly winds by default, then when the wind direction data of the input wind speed and direction meteorological data is easterly, the azimuth order of the feature vector needs to be adjusted accordingly, i.e., the mapping GRF(r_i, θ_j, z_k) → GRF(r_i, θ_j', z_k) is performed. Specifically, if the original azimuth index is θ_j=[1, 2, 3, ..., 16], it is adjusted to θ_j'=[9, 10, 11, ..., 16, 1, 2, ..., 8] under easterly conditions to achieve wind direction alignment. In addition, r_i and z_k are the indexes of the radius direction and the vertical height, respectively.

[0111] Thanks to the rotational symmetry of the cylindrical coordinate system, this embodiment eliminates the need for time-consuming preprocessing such as repeated rotation and cropping of the original building outline data when dealing with wind direction changes. It only requires adjusting the order of the pre-stored GRF feature vectors to quickly complete the prediction. This embodiment boasts excellent scalability and repeatability, allowing for continuous prediction of the next time step based on actual needs, or switching to different meteorological scenarios for multiple simulation analyses, significantly improving the efficiency of long-term forecasting and multi-scenario analysis.

[0112] In step S104 of some embodiments, please refer to Figure 3 A residual network is used as the urban wind field prediction model. The model architecture includes:

[0113] Input layer: The number of neurons is consistent with the dimension N of the GRF feature vector (e.g., 1600).

[0114] Hidden layers: consist of multiple residual blocks. Each residual block contains a fully connected layer, an activation function (such as the tanh function), and a skip connection. The output of the fully connected layer is the activation function (input × weight matrix + bias vector). For example, the model contains four residual blocks with the following numbers of neurons: 150, 125, 75, and 40, respectively.

[0115] Output layer: The number of neurons corresponds to the number of predicted targets, typically including three normalized wind speed components (flow velocity u / U).120m spanwise velocity v / U 120m Vertical velocity w / U 120m ), and can be extended to include turbulence parameters.

[0116] The adjusted feature vector is input into a pre-trained deep learning model, which outputs the normalized wind speed component (e.g., u / U) for the predicted target point P_i. 120m ,v / U 120m w / U 120m ) and turbulence parameters, i.e., initial wind field data.

[0117] Using a single point as the prediction unit, it can be directly applied to any new urban area without retraining, breaking through the limitations of fixed grids. From the adjusted feature vector to the prediction of standardized wind speed components and turbulence parameters, it achieves end-to-end true 3D wind field prediction, reducing inconsistencies caused by model patching.

[0118] In some embodiments, the trained deep learning urban wind field prediction model can be obtained through steps S601 to S606:

[0119] Step S601: Obtain the large eddy simulation wind speed components and turbulence parameter datasets for different urban simulation areas;

[0120] Step S602: Select data samples from the large eddy simulation wind speed components and turbulence parameter dataset;

[0121] Step S603: Read the feature vector of the convolution calculation data sample based on the geometric features of the three-dimensional cylindrical building;

[0122] Step S604: Input the feature vector into the deep learning urban wind field prediction model for prediction, and obtain the wind speed component and turbulence parameter output;

[0123] Step S605: Based on the simulation data of the wind speed component and turbulence parameter dataset from the large eddy simulation, the error is calculated separately for the wind speed component and turbulence parameter to obtain their respective mean square errors.

[0124] Step S606: With the goal of minimizing the mean square error, the deep learning urban wind field prediction model is trained and optimized to obtain a trained deep learning urban wind field prediction model.

[0125] In step S601 of some embodiments, training data for the wind field prediction model is generated using a high-precision Large Eddy Simulation (LES) method validated through wind tunnel experiments. This training data is generated for multiple (e.g., three) urban areas with different building layouts (approximately 3-4 km² in area). 2Large eddy simulations (LES) with a spatial resolution of 5 meters were performed under various inflow winds. The time-averaged normalized flow fields obtained from the simulations were used as the ground truth data, namely, the large eddy simulation wind speed components and turbulence parameter dataset.

[0126] In steps S602 to S606 of some embodiments, a large number of sample points are selected from the LES dataset. The feature vector of each sample point is extracted using the GRF algorithm of this embodiment as the model input, and the corresponding LES wind speed / turbulence value is used as the target output. The stochastic gradient descent algorithm is used to calculate the mean square error (predicted value - calculated LES value) for the three wind speed components and different turbulence parameters. 2 The parameters of the neural network are optimized as a loss function until the model converges, resulting in a well-trained wind field prediction model.

[0127] This embodiment significantly reduces data preparation costs and training barriers, while GRF's essential representation of building morphology greatly improves the model's learning efficiency. Practice shows that only a small number (e.g., 12) of representative high-precision simulation cases are needed to train a generalized model with strong generalization capabilities, which greatly saves computing resources and time costs compared to related technologies that require hundreds of cases.

[0128] In step S105 of some embodiments, when the deep learning urban wind field prediction model completes the prediction of all prediction target points and obtains the initial wind field data, the initial wind field data is integrated to obtain the three-dimensional wind field prediction of the corresponding urban target area under different wind directions.

[0129] In some embodiments, please refer to Figure 4 The overall process of the urban low-altitude wind field prediction method based on three-dimensional cylindrical building geometric feature reading convolution (GRF) and deep learning provided in this application embodiment includes two core stages:

[0130] (1) GRF feature pre-calculation stage (independent of meteorological conditions):

[0131] This stage only needs to be executed once for a specific urban area. First, input the 3D building outline data of the target urban area (e.g., a common format such as Shapefile). Then, define the 3D spatial target point set P_i within this area for wind speed prediction. The target point set P_i can be regularly distributed (e.g., a 3D Cartesian grid) or arbitrarily discrete, with its height range preferably covering 0 to 120 meters, and a spatial resolution up to 5 meters. Next, traverse the target point set, calculating the corresponding 3D cylindrical GRF feature vector for each target point P_i. After calculation, store the GRF feature vectors of all target points in a database for repeated use in the prediction stage.

[0132] (2) Wind field forecasting stage (related to meteorological conditions):

[0133] This stage can be executed multiple times depending on different meteorological conditions. When it is necessary to predict specific meteorological conditions, the meteorological parameters (including wind speed and direction at a known altitude, such as 120 meters) are input. Subsequently, for each target point P_i defined in the pre-calculation stage, a loop operation is performed: First, the pre-stored GRF feature vector of the target point is read from the preset database. This feature vector is divided into 16 uniform regions in the azimuth (θ direction), corresponding to the 16 different wind directions considered. Since the pre-stored GRF(r_i, θ_j, z_k) is constructed based on the westerly wind direction by default, when the input wind direction is easterly, the azimuth order of the feature vector needs to be adjusted accordingly, that is, the mapping GRF(r_i, θ_j, z_k) → GRF(r_i, θ_j', z_k) is executed. Specifically, if the original azimuth index is θ_j=[1, 2, 3, ..., 16], it is adjusted to θ_j'=[9, 10, 11, ..., 16, 1, 2, ..., 8] under easterly conditions to achieve wind direction alignment. The adjusted feature vector is input into a pre-trained deep learning model, and the model outputs the standardized wind speed component (e.g., u / U) of the target point P_i. 120m ,v / U 120m w / U 120m The data includes turbulence parameters. After all target points have been processed, a three-dimensional wind field prediction for the entire target area is synthesized.

[0134] Please refer to Figure 5 To verify the prediction performance of the deep learning (DL) model in this embodiment, Figure 5 (a) is a comparison of the wind direction and speed results between the DL model prediction and the LES simulation results, and (b) is a scatter plot (wind field within 0-120 meters). The background wind directions are (i): +x, (ii): -x, (iii): +y, and (iv): -y. RSME is the root mean square error, PE is the prediction error, and NMB is the normalized mean deviation.

[0135] The deep learning-based urban low-altitude wind field prediction method provided in this embodiment can be specifically productized as a rapid urban low-altitude wind field prediction system based on three-dimensional cylindrical building geometric feature reading convolution (GRF). This product, by deploying the proposed three-dimensional cylindrical GRF module and corresponding prediction model, can access standard urban three-dimensional building data and achieve three-dimensional wind field distribution prediction within a height range of 0-120 meters. This product possesses true three-dimensional prediction capabilities and instantaneous wind direction adaptation characteristics, supporting long-term forecasting and multi-scenario analysis. It can quickly simulate meteorological conditions such as different wind directions and speeds after completing GRF feature calculations in one go, significantly outperforming solutions requiring repeated data preprocessing. Simultaneously, the system has good generalization ability; the trained model can be directly applied to different urban areas for rapid evaluation. This product is applicable to multiple professional fields such as urban ventilation assessment, wind safety assessment, and planning and design optimization, providing efficient and reliable decision support for urban management and planning.

[0136] This application also provides a deep learning-based urban low-altitude wind field prediction device, which can implement the above-described method. The device includes:

[0137] The acquisition module is used to acquire 3D building outline data and wind speed, wind direction and meteorological data of the target urban area;

[0138] The calculation module is used to perform three-dimensional cylindrical building geometric feature reading and convolution calculation on one or more predefined prediction target points based on the three-dimensional building outline data, so as to obtain the feature vector of each prediction target point;

[0139] The adjustment module adjusts the feature vector based on meteorological data of wind speed and direction to obtain the adjusted feature vector;

[0140] The prediction module is used to input the adjusted feature vector into the trained deep learning urban wind field prediction model to predict the wind field and obtain the initial wind field data output for one or more prediction target points.

[0141] The synthesis module is used to synthesize multiple initial wind field data outputs to obtain a three-dimensional wind field prediction for the target urban area.

[0142] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0143] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0144] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0145] Please see Figure 6 , Figure 6 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0146] The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0147] The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 using the methods described in the embodiments of this application.

[0148] The input / output interface 903 is used to implement information input and output;

[0149] The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0150] Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904);

[0151] The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0152] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0153] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0154] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0155] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0156] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0157] The urban low-altitude wind field prediction method, device, electronic device, storage medium, and program product based on deep learning provided in this application have at least the following beneficial effects:

[0158] 1. True 3D representation capability: The cylindrical GRF can directly and naturally extract the architectural features in 3D space, overcoming the fundamental limitation of the existing 2D mesh method in terms of insufficient representation capability in the vertical dimension, and laying the foundation for achieving end-to-end true 3D wind field prediction.

[0159] 2. High wind direction adaptability and computational efficiency: Thanks to the rotational symmetry of the cylindrical coordinate system, the model can instantly complete predictions when dealing with wind direction changes without requiring time-consuming preprocessing such as repeated rotation and cropping of the original building model data. This gives it an unparalleled efficiency advantage in long-term forecasting and multi-scenario analysis.

[0160] 3. Strong generalization ability and robustness: GRF extracts features based on basic geometric volume occupancy, without relying on any artificially preset complex parameters that may have uncertainties, making this embodiment more adaptable and robust to different urban forms and layouts.

[0161] 4. Significantly reduced data preparation costs and training threshold: GRF's essential representation of building morphology greatly improves the model's learning efficiency. Practice shows that only a small number (e.g., 12) of representative high-precision simulation cases are needed to train a generalized model with strong generalization capabilities, which greatly saves computing resources and time costs compared to existing technologies that require hundreds of cases.

[0162] In summary, this embodiment effectively solves the core bottlenecks of related technologies in terms of 3D modeling, wind direction adaptability, and data efficiency, and provides a practical and feasible technical solution for achieving efficient and universal urban wind field prediction.

[0163] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0164] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0165] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0166] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0167] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0168] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0170] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A deep learning-based method for predicting urban low-altitude wind fields, characterized in that, The method includes the following steps: Acquire 3D building outline data and wind speed, direction and meteorological data for the target urban area; Based on the three-dimensional building outline data, three-dimensional cylindrical building geometric feature reading convolution calculation is performed on one or more predefined prediction target points to obtain the feature vector of each prediction target point; The feature vector is adjusted based on the wind speed and wind direction meteorological data to obtain the adjusted feature vector; The adjusted feature vector is input into the trained deep learning urban wind field prediction model to predict the wind field, and the initial wind field data output of one or more prediction target points is obtained. The multiple initial wind field data outputs are combined to obtain a three-dimensional wind field prediction for the target urban area. The step of performing convolution calculations on predefined single or multiple prediction target points based on the three-dimensional building contour data to obtain the feature vector of each prediction target point includes: Using the predicted target point as the spatial origin of the three-dimensional building outline data, a cylindrical coordinate system is constructed in the surrounding space; The surrounding space is divided into multiple partitions based on the three dimensions of the cylindrical coordinate system. The feature values ​​of the partition are obtained by calculating the feature values ​​of the partition. The feature vector is obtained by constructing a feature vector from multiple feature values; The surrounding space is divided according to the three dimensions of the cylindrical coordinate system to obtain multiple partitions, including: The surrounding space is divided according to the radial direction of the cylindrical coordinate system to obtain multiple annular spaces; The annular space is divided into multiple first spaces by sectoring the cylindrical coordinate system according to the azimuth angle. The first space is divided into multiple partitions based on the vertical height of the cylindrical coordinate system.

2. The method according to claim 1, characterized in that, The step of calculating the feature values ​​of the partition to obtain the feature values ​​of the partition includes: The volume of the partition is calculated to obtain the total volume; The volume of the building space occupied by the building within the partition is calculated to obtain the volume of the building space occupied by the partition. The building volume occupancy rate is calculated based on the total volume and the building's occupied space volume. The building volume occupancy rate is used as the characteristic value.

3. The method according to claim 1, characterized in that, The step of constructing a feature vector from multiple feature values ​​to obtain the feature vector includes: The feature vector is obtained by sorting the multiple feature values ​​according to the vertical height index, the radius direction index, and the azimuth index.

4. The method according to claim 1, characterized in that, The step of adjusting the feature vector based on the wind speed and wind direction meteorological data to obtain the adjusted feature vector includes: The azimuth index order of the feature vector is adjusted based on the wind direction data in the wind speed and wind direction meteorological data to obtain a new feature value sequence; The positions of the feature values ​​in the feature vector are adjusted according to the feature value sequence to obtain the adjusted feature vector.

5. The method according to claim 1, characterized in that, The trained deep learning urban wind field prediction model is obtained through the following steps: Obtain datasets of wind speed components and turbulence parameters from large eddy simulations in different urban simulation areas; Data samples were selected from the large eddy simulation wind speed components and turbulence parameter dataset. The feature vector of the data sample is calculated by convolution based on the geometric features of a three-dimensional cylindrical building. The feature vector is input into the deep learning urban wind field prediction model for prediction, and the wind speed component and turbulence parameters are output. Based on the simulation data of the large eddy simulation wind speed component and turbulence parameter dataset, the error is calculated with the wind speed component and turbulence parameter respectively to obtain their respective mean square errors; With the goal of minimizing the mean square error, the deep learning urban wind field prediction model is trained and optimized to obtain the trained deep learning urban wind field prediction model.

6. A deep learning-based urban low-altitude wind field prediction device, characterized in that, The device includes: The acquisition module is used to acquire 3D building outline data and wind speed, wind direction and meteorological data of the target urban area; The calculation module is used to perform three-dimensional cylindrical building geometric feature reading convolution calculation on one or more predefined prediction target points according to the three-dimensional building outline data, so as to obtain the feature vector of each prediction target point; The adjustment module adjusts the feature vector based on the wind speed and direction meteorological data to obtain the adjusted feature vector; The prediction module is used to input the adjusted feature vector into the trained deep learning urban wind field prediction model to predict the wind field and output the initial wind field data of one or more prediction target points. The synthesis module is used to synthesize multiple initial wind field data outputs to obtain a three-dimensional wind field prediction of the urban target area. The calculation module is specifically used to perform the following steps: Using the predicted target point as the spatial origin of the three-dimensional building outline data, a cylindrical coordinate system is constructed in the surrounding space; The surrounding space is divided into multiple partitions based on the three dimensions of the cylindrical coordinate system. The feature values ​​of the partition are obtained by calculating the feature values ​​of the partition. The feature vector is obtained by constructing a feature vector from multiple feature values; The surrounding space is divided according to the three dimensions of the cylindrical coordinate system to obtain multiple partitions, including: The surrounding space is divided according to the radial direction of the cylindrical coordinate system to obtain multiple annular spaces; The annular space is divided into multiple first spaces by sectoring the cylindrical coordinate system according to the azimuth angle. The first space is divided into multiple partitions based on the vertical height of the cylindrical coordinate system.

7. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1 to 5.

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

Images