A high-precision reconstruction method of wind field around a square cross-section building

Abstract

The present application relates to a kind of high-precision reconstruction methods of square section building around wind field, the method includes obtaining the wind speed measured point value of spatially sparse collected by sensor, based on the WGAN-GP of embedding gradient penalty mechanism of optimized training, the wind speed measured point value of spatially sparse is reconstructed as the global wind speed distribution around building, the present application improves the prediction accuracy of reconstructing high-dimensional global wind field of blunt body based on sparse spatial wind speed measured point, successfully reproduces the high-frequency turbulent component under real wind field condition, with good generalization performance, and is expected to be applied to high-precision reproduction of urban wind field and accurate prediction of urban building wind pressure.

Description

Technical Field

[0001] The present invention relates to a high-precision reconstruction method for the wind field around a building with a square cross-section, specifically to a method for reconstructing the high-dimensional and high Reynolds number global wind field distribution around a building with a square cross-section by using a generative adversarial network based on spatially sparse wind speed measurement point information, belonging to the technical field of high-precision prediction of building wind fields in structural wind engineering. Background Art

[0002] In the past decade, the development and application of data reconstruction technologies for physical fields have received certain attention and development in some engineering fields, such as geophysics, atmospheric science, fluid dynamics, etc. In fluid engineering applications, the distribution of physical quantities in the global flow field (such as velocity or pressure field) usually needs to be determined by time-consuming and expensive high-fidelity wind tunnel experiments or numerical simulation methods. On the one hand, in actual measurements, the available sensors are usually limited; on the other hand, in actual engineering applications, more attention is paid to the global distribution of physical quantities. Therefore, it is of great engineering significance to reconstruct the high-dimensional global wind field by using limited local sensor data.

[0003] Wind field reconstruction is defined as the mapping process from a low-dimensional vector s ∈ R p to a high-dimensional vector x ∈ R m (p << m), where s represents the sparse sensor measurement value and x represents the global wind field. Recent studies have shown that machine learning or deep learning methods are expected to help in the reconstruction of high-dimensional flow fields based on sparse sensor data, achieving a balance between reconstruction accuracy and computational cost. For the flow field around a bluff body, most deep learning studies have only focused on the reconstruction of the wake of the structure under laminar conditions. Erichson et al. introduced a shallow perceptron and successfully predicted the two-dimensional laminar vorticity field of a cylinder under the condition of Re = 100. Fukami et al. systematically analyzed the accuracy and robustness of many deep learning methods, including multi-layer perceptrons, random forests, support vector regression, extreme learning machines, etc., and at the same time found that convolutional neural networks (CNNs) have certain advantages in extracting the distribution characteristics of high-dimensional data.

[0004] In structural engineering, square-section column structures are commonly used in the design of high-rise urban buildings. These buildings often experience more complex wake effects, including separation flow from bluff bodies, laminar-turbulent transition in shear layers, and the generation and propagation of von Kármán vortex streets. However, for the reconstruction of building wind fields at high Reynolds numbers, the sparse distribution of wind speed measurement points around the building obtained from actual measurements makes traditional numerical simulation methods prohibitively expensive. Existing deep learning techniques, on the other hand, suffer from insufficient reconstruction accuracy. For example, Diop et al. attempted to reconstruct complex turbulent fields using artificial neural networks based on limited velocity measurement points within the wake and pressure measurement points on the building surface, but this failed to achieve reconstruction results that meet engineering application requirements; its prediction accuracy was even lower than that of linear regression methods. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-precision prediction method for reconstructing the wind field around a square-section building based on generative adversarial networks.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A high-precision reconstruction method for the wind field around a square-section building includes acquiring spatially sparse wind speed measurement points collected by sensors, and reconstructing the spatially sparse wind speed measurement points into a global wind speed distribution around the building based on an optimized and trained Wasserstein generative adversarial network (WGAN-GP) with an embedded gradient penalty mechanism. The optimization training process of the Wasserstein generative adversarial network with the embedded gradient penalty mechanism includes the following steps:

[0008] (1) Obtain wind speed measurement data of a building wind field under a certain Reynolds condition, and process the wind speed measurement data using the direct data simulation method to generate training set data of the real wind field wind speed distribution.

[0009] (2) Construct a Wasserstein generative adversarial network based on an embedded gradient penalty mechanism;

[0010] (3) The Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is trained using training set data of real wind field wind speed distribution.

[0011] (4) The Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is optimized and trained using transfer learning techniques.

[0012] Furthermore, step (1) includes:

[0013] 1a) Obtain wind speed measurement data of a building wind field under a certain Reynolds condition;

[0014] 1b) The wind speed measurement data of the building wind field is processed by direct numerical simulation to generate wind speed distribution data of multiple horizontal sections along the building height.

[0015] 1c) The bicubic interpolation technique is used to interpolate the wind speed distribution data of each horizontal section to the grid positions that are uniformly distributed throughout the entire area, thereby generating global wind speed distribution data.

[0016] 1d) The k-means clustering technique was used to determine the grid points with large wind speed temporal fluctuations in the building shear layer and building leeward zone of the global building wind field, and to extract the spatially sparse wind speed measurement points.

[0017] 1e) Determine the location code of the sparse wind speed measurement point value based on the extracted sparse wind speed measurement point value;

[0018] 1f) The location code is added to the global wind speed distribution data and combined with the spatially sparse wind speed measurement points to form a training set of real wind field wind speed distribution data.

[0019] Further, step 1f) includes storing global wind speed distribution data using a first-layer channel, storing the location codes corresponding to sparse wind speed measurement points in the global wind speed distribution data using a second-layer channel, and constructing global wind speed distribution data with a matrix dimension of N×M×L×2 in the real wind field wind speed distribution; storing sparse wind speed measurement point values ​​using C sparse channels, and constructing sparse wind speed measurement point values ​​with a matrix dimension of N×1×1×C in the real wind field wind speed distribution, where N represents the number of training samples; M×L represents the pressure measurement point dimension of the real spatial accuracy, and C represents the number of sparse wind speed measurement points.

[0020] Furthermore, in step (2), the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism comprises a generator and a discriminator. The generator is used to generate spatial data distributions that approximate the real wind speed field. The generator adopts a deep fully convolutional architecture, containing m consecutive convolutional blocks. Each convolutional block contains a deconvolutional layer, a batch normalization layer, and a ReLU layer as the activation function. The deconvolutional layer contains a K×K convolutional kernel, used to upsample low-dimensional spatial features. The discriminator is used to distinguish between the real data spatial distribution and the data output by the generator module. The discriminator adopts a deep convolutional architecture, containing n consecutive convolutional blocks. Each convolutional block contains a convolutional layer and a LeakyReLU activation function module. The convolutional layer contains a K×K convolutional kernel, used to extract low-dimensional features from the high-dimensional data distribution. A fully connected layer is concatenated after the consecutive convolutional blocks to calculate the probability value that the generated data distribution is the real distribution.

[0021] Furthermore, in step (3), training the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism includes:

[0022] The weights and biases of each layer of the network are initialized. For the generator module of the network, the mean square error loss function is minimized for iterative optimization. At the same time, it is trained in conjunction with the discriminator module. Through multiple training iterations, the discriminator's ability to detect falsehoods is improved, and the optimization of each weight and bias of the generator module is promoted.

[0023] Furthermore, in step (4), optimizing the training of the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism includes: obtaining the Wasserstein generative adversarial network trained under a certain Reynolds condition; obtaining wind speed measurement data of the building wind field under different Reynolds conditions, generating migration data of the real wind field wind speed distribution; and using the migration data of the real wind field wind speed distribution and training steps to optimize the training of the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism.

[0024] Compared with the prior art, the present invention has significant advantages and beneficial effects, specifically reflected in the following aspects:

[0025] ① This invention employs partitioned k-means clustering technology to determine the location of sparse wind speed points by considering the differences in wind field characteristics in different areas around a building; based on location encoding preprocessing, it focuses on the spatial radiation effect of each wind speed measurement point location on the global wind speed distribution; and utilizes Wasserstein generation pairs (WGAN-GP) with embedded gradient penalty mechanism to reconstruct spatially sparse wind speed measurements into spatially dense wind speed distributions to predict the global wind field around a square-section building.

[0026] ② This invention overcomes the shortcomings of high cost and technical difficulty in measuring the wind field across a square building, and improves the prediction efficiency of high-dimensional, high Reynolds number wind field distribution. Compared with deep learning reconstruction methods based on multilayer perceptrons and convolutional neural networks, the wind field reconstruction accuracy generated by WGAN-GP in this invention is significantly improved, and it can more accurately reflect the physical characteristics of the real wind field.

[0027] ③ This invention fully considers the flow pattern differences of urban building wind fields under different Reynolds numbers. Based on the transfer learning method, it quickly updates the parameters of the WGAN-GP model to achieve high-precision reconstruction of the wake of square buildings under different Reynolds numbers, and has good generalization performance. The reconstruction accuracy of this invention is less affected by the number and location of sparse wind speed measurement points, so the wind field prediction is more robust.

[0028] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing specific embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of the reconstruction method according to an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the reconstruction method according to a preferred embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the square building wind field area division used in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram verifying the wind speed dataset of the direct numerical simulation of the wind field of a high-dimensional square building in space according to an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the generative adversarial network architecture according to an embodiment of the present invention;

[0035] Figure 6 This is a schematic diagram of the structure of the generative adversarial network generator module in an embodiment of the present invention;

[0036] Figure 7 This is a schematic diagram of the structure of the generative adversarial network discriminator module in an embodiment of the present invention;

[0037] Figure 8 This is a comparative cloud map based on the existing multilayer perceptron method and the method of this invention for reconstructing the instantaneous wind speed field distribution over the entire area of ​​a square building;

[0038] Figure 9 This invention reconstructs cloud maps of the global instantaneous wind speed field of a square building under different Reynolds numbers based on existing multilayer perceptron methods and the method of this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0040] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, directional and ordinal terms are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0041] To address the limitations of sparse wind speed measurement points around buildings in actual measurements, the high cost of traditional numerical simulation methods under high Reynolds numbers, and the low accuracy of existing traditional deep learning methods in reconstructing the global wind field around buildings, this invention utilizes a Wasserstein generative adversarial network (WGAN-GP) with an embedded gradient penalty mechanism to reconstruct a high-dimensional wind speed field based on spatially sparse wind speed measurement point data. This improves the accuracy of global wind field prediction around square buildings and enhances the robustness of the prediction, meeting the needs of practical engineering applications.

[0042] like Figure 1 As shown, this invention provides a high-precision reconstruction method for the wind field around a square-section building. It acquires sparse wind speed measurement points collected by sensors and, based on an optimized and trained Wasserstein generative adversarial network (WGAN-GP) with an embedded gradient penalty mechanism, reconstructs the sparse wind speed measurement points into a global wind speed distribution around the building. The optimization training process of the Wasserstein generative adversarial network with the embedded gradient penalty mechanism includes the following steps:

[0043] (1) Obtain wind speed measurement data of a building wind field under a certain Reynolds condition, and process the wind speed measurement data using the direct data simulation method to generate training set data of the real wind field wind speed distribution.

[0044] In a preferred embodiment of the present invention, such as Figure 2As shown, this embodiment can optimize the training of the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism using training set data and test set data. This embodiment obtains the training and validation sample sets of the real wind field wind speed distribution based on the direct numerical simulation results of the wind field of the global square building with high spatial dimension. The obtained training samples are preprocessed to obtain training and validation sample pairs of real and sparse spatial distribution.

[0045] In this embodiment of the invention, since it is necessary to reconstruct the global wind speed distribution around the building using spatially sparse wind speed measurement points, this embodiment needs to consider how to select reasonable sparse wind speed measurement points as training data. The extraction of sparse input takes into account the differences in wind field distribution characteristics in different areas around the building, such as... Figure 3 As shown, the entire building wind field is divided into two blocks: the building shear zone (near the two sides of the building) and the building leeward zone (wake region). K-means clustering is used to identify locations with significant wind speed temporal fluctuations in each zone. Spatially sparse wind speed measurement points are extracted and used as the sparse input to WGAN-GP, with dimensions N×1×1×C, where the number of channels C is the number of sparse measurement points (C < 1 / C). <M×L)。

[0046] In this embodiment of the invention, in addition to considering sparse wind speed measurement points, complete global wind speed distribution data is also needed to help optimize the reconstruction process. Based on this, the matrix dimension of the real wind speed distribution data is N×M×L×2, where N represents the number of training samples, that is, the training set contains wind speed distributions for N consecutive time steps, M×L represents the pressure measurement point dimension of the real spatial accuracy, and "2" refers to the number of data channels. The first layer of channels is used to store the real flow velocity value data, and the second layer of channels is the position encoding, which is used to store the position encoding corresponding to the sparse wind speed measurement point values ​​in the global wind speed distribution data. The matrix point corresponding to the sparse measurement point position is marked with data "1", and "0" is used to represent the position where no sparse measurement points are arranged. At the corresponding time step, the real wind speed distribution data is used as the high-dimensional label of the sparse input.

[0047] Based on the above embodiments, step (1) in this embodiment includes:

[0048] 1a) Obtain wind speed measurement data of a building wind field under a certain Reynolds condition;

[0049] 1b) The wind speed measurement data of the building wind field is processed by direct numerical simulation to generate wind speed distribution data of multiple horizontal sections along the building height.

[0050] 1c) The bicubic interpolation technique is used to interpolate the wind speed distribution data of each horizontal section to the grid positions that are uniformly distributed throughout the entire area, thereby generating global wind speed distribution data.

[0051] 1d) The k-means clustering technique was used to determine the grid points with large wind speed temporal fluctuations in the building shear layer and building leeward zone of the global building wind field, and to extract the spatially sparse wind speed measurement points.

[0052] In this embodiment of the invention, k-means clustering technology is used, with wind speed standard deviation as the criterion. Two and three grid points with the largest wind speed time fluctuations are selected in the building shear zone and the building leeward zone, respectively. Based on these five measuring points, sparse wind speed values ​​are extracted.

[0053] Specifically, this embodiment uses k-means clustering technology to first select N clustering regions and their central grid points from a 100×100 grid area. Then, based on the wind speed data of each sample, the variance of these N clustering regions, i.e., the pulsation value, is calculated. The central grid points of the five regions with the largest pulsations are then selected as the actual sensor placement locations, and the wind speed test value at these central grid points is used as the sparse wind speed measurement point values ​​in that space. In this embodiment, compared to the general global k-means clustering method, the wind field partitioning approach reduces the interference of multi-source wind speed noise on the clustering effect.

[0054] 1e) Determine the location code of the sparse wind speed measurement point value based on the extracted sparse wind speed measurement point value;

[0055] 1f) The location code is added to the global wind speed distribution data and combined with the spatially sparse wind speed measurement points to form a training set of real wind field wind speed distribution data.

[0056] In another embodiment of the invention, such as Figure 4 As shown, taking the high Reynolds number conditions of Re=22000 and Re=5000 as examples, a direct numerical simulation database of wind field for a high-dimensional square building in space is created. Figure 4(a) and (b) respectively show the wind pressure distribution of a square building obtained by direct numerical simulation under Re=22000 and Re=5000 conditions. Both are consistent with previous experimental and simulation results, proving the reliability of the database. Based on this database, training and validation sample sets of wind speed distribution are obtained. The obtained training samples are preprocessed to obtain training sample pairs of real and sparse wind speed distributions. To train the WGAN-GP network model and achieve high-precision reconstruction of the wind field around a square cross-section building, this embodiment directly uses a high-dimensional spatial square column flow field database obtained by direct numerical simulation using Nektar++ software as the training and testing set for WGAN-GP. The database records wind speed distribution data for 10 horizontal cross-sections along the building height direction. The database covers 2163 time steps, collects 21630 instantaneous cross-sectional data, and divides the training and validation sets in a 7:3 ratio. The first 10 data points are selected from the second 10 data points. The training set was generated from 1500 time points (15,000 instantaneous cross-sectional data points), and the validation set was generated from 663 time points (6,630 instantaneous cross-sectional data points). During preprocessing, bicubic interpolation was used to interpolate the original direct numerical simulation wind speed data of each cross-section to a globally distributed 100×100 grid, which served as the first channel of the WGAN-GP model output label. For each real instantaneous cross-section with a dimension of 100×100×2, partitioned k-means clustering was used to determine the locations with large wind speed temporal fluctuations in the shear zone and leeward zone, extracting the measurement values ​​of 5 grid points to form sparse wind speed measurement point data with a dimension of 1×1×5, which served as the sparse input of WGAN-GP. A 100×100 zero matrix was generated, and the matrix points corresponding to the sparse measurement point locations were replaced with the value "1" instead of "0", which served as the second channel of the WGAN-GP model output label.

[0057] (2) Construct a Wasserstein generative adversarial network based on an embedded gradient penalty mechanism;

[0058] In embodiments of the present invention, such as Figure 5 As shown, the adversarial generative network consists of a generator G and a discriminator D. The generator is used to generate spatial data distributions that approximate the real wind speed field, such as... Figure 6 As shown, the generator employs a deep fully convolutional architecture, containing m consecutive convolutional blocks. Each convolutional block includes a deconvolutional layer, a batch normalization layer, and a ReLU layer as the activation function. The deconvolutional layer contains a K×K convolutional kernel used to upsample low-dimensional spatial features. The discriminator distinguishes the spatial distribution of the real data from the output of the generator module, such as... Figure 7As shown, the discriminator adopts a deep convolutional architecture, which contains m consecutive convolutional blocks. Each convolutional block contains a convolutional layer and a LeakyReLU activation function module. The convolutional layer contains a convolutional kernel of size K×K, which is used to extract low-dimensional features from the high-dimensional data distribution. The fully connected layer is concatenated after the consecutive convolutional blocks and is used to calculate the probability value that the generated data distribution is the true distribution.

[0059] In this embodiment of the invention, considering that the number of convolutional blocks and the size of the convolutional kernel determine the scale of WGAN-GP, the larger the number of convolutional blocks and the larger the size of the convolutional kernel, the stronger the nonlinear learning ability of WGAN-GP, which will lead to too many network parameters and the model is prone to overfitting; the smaller the number of convolutional blocks and the smaller the size of the convolutional kernel, the weaker the nonlinear learning ability and the underfitting of the model. Through experimental adjustment, the number of convolutional block layers of the generator and the discriminator is determined to be m=16 and n=16, the kernel size of the first convolutional block of the generator is determined to be K×K=5×5, and K×K=4×4 in the remaining convolutional blocks, the kernel size of the last convolutional block of the discriminator is determined to be K×K=5×5, and K×K=4×4 in the remaining convolutional blocks.

[0060] (3) The Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is trained using training set data of real wind field wind speed distribution.

[0061] In this embodiment of the invention, in the training step (2), an adversarial network is generated, the weights and biases of each layer of the network are initialized, and the generator module of the network is iteratively optimized by minimizing the mean square error loss function. At the same time, it is trained in conjunction with the discriminator module. Through multiple training iterations, the discriminator's ability to detect falsehoods is continuously improved, and the optimization of each weight and bias of the generator module is promoted to ensure the high-quality output of the WGAN-GP model. Using the trained WGAN-GP model, the sparse wind speed measurement points around the square building are reconstructed into a spatially dense global wind speed distribution.

[0062] The mean squared error loss function is expressed as:

[0063]

[0064] Among them, y i x represents the actual wind speed distribution i For sparse wind speed measurement points, G(x) i ( ) represents the wind speed distribution reconstructed by the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism at the corresponding time step.

[0065] In this embodiment of the invention, a learning rate of 2×10 can be given. -4The batch size is 16, and the ADAM optimizer is used to minimize the training loss function. In each training step, the discriminator and generator are trained together. The discriminator parameters are updated every 5 times, and the generator parameters are updated once. Through multiple training iterations, the discriminator's ability to detect false positives is continuously improved, and the optimization of each weight and bias of the generator module is promoted to ensure the high-quality output of the WGAN-GP model. Training stops when the generator's loss function reaches convergence.

[0066] In a preferred embodiment of the present invention, this embodiment uses the WGAN-GP model trained under the corresponding Reynolds number conditions to reconstruct the sparse wind speed measurements at each time step in the validation set into a spatially dense global wind speed distribution around buildings, and verifies the reconstruction accuracy of the model based on the real wind speed distribution in the validation set.

[0067] (4) The Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is optimized and trained using transfer learning techniques.

[0068] In this embodiment of the invention, a Wasserstein generative adversarial network based on an embedded gradient penalty mechanism is obtained after training under a certain Reynolds condition; wind speed measurement data of building wind fields under different Reynolds conditions are obtained to generate transfer data of the real wind field wind speed distribution; the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is optimized and trained using the transfer data of the real wind field wind speed distribution and the training steps.

[0069] For example, taking the training set data of the high Reynolds number condition Re=22000 in the above embodiment as an example, the Wasserstein generative adversarial network WGAN-GP with embedded gradient penalty mechanism trained under this Reynolds number condition can reconstruct the wind field around the building under the high Reynolds number condition Re=22000, but it cannot effectively reconstruct the wind field around the building under other conditions. Therefore, this invention adopts the transfer learning approach. With the help of transfer learning technology, based on a small number of wind speed distribution training set samples under different Reynolds number conditions, a small number of training steps are used to quickly update the parameters of the parent WGAN-GP model, so as to achieve fast and accurate reconstruction and prediction of wind fields under different Reynolds number conditions.

[0070] To better illustrate the effects of the present invention, in this embodiment, the training environment for the WGAN-GP model is PyTorch-1.12.1-GPU.

[0071] Test content: Based on the WGAN-GP model trained under the condition of Re=22000, evaluate the reconstruction performance of the global wind field around the square pillar under different Reynolds number scenarios. The global wind field prediction is performed using the method of this invention and the reconstruction method based on the multilayer perceptron model, and compared with the real high-precision numerical simulation results. Figure 8a represents the instantaneous global wind speed field of a square building reconstructed by the method of this invention under the condition of Re = 22000. Figure 8 b represents the global instantaneous wind speed field of a square building under the condition of Re=22000, reconstructed based on the multilayer perceptron method. Figure 8 c represents the high-precision numerical simulation results of the wind field around the square building; Figure 9 a is a cloud map of the instantaneous wind speed field over the entire area of ​​a square building under the condition of Re=5000, reconstructed based on the existing multilayer perceptron; Figure 9 b is a cloud map of the instantaneous wind speed field of the entire square building under the condition of Re=5000, reconstructed based on the method of the present invention. Figure 9 c represents the high-precision numerical simulation result of the wind field of a square building under the condition Re = 5000.

[0072] Depend on Figure 8 It can be observed that, based on a certain number of sparse wind speed measurement points (5 points), both the method of this invention and the reconstruction method based on a multilayer perceptron can successfully capture the basic characteristics of the Karman vortex street in the wake of a square pillar. However, the flow field reconstructed by the multilayer perceptron is too smooth, missing many small-scale vortex structures. Compared with the multilayer perceptron reconstruction method, the method of this invention exhibits a stronger ability to reconstruct small-scale vortex structures not only in the wake region but also in the shear layer, and to a certain extent, it can predict the Kelvin-Helmholtz instability of the shear layer.

[0073] Depend on Figure 9 The results show that transfer learning techniques enable models to adapt to specific characteristics under different incoming wind conditions while retaining knowledge from the original model's training scenario. Considering different Reynolds number conditions, the method of this invention can still reproduce small-scale eddy information that closely matches the real high-precision flow field, demonstrating strong generalization performance.

[0074] Table 1 shows the errors in reconstructing the instantaneous wind speed field of the square column using existing multilayer perceptron-based reconstruction methods, existing convolutional neural network reconstruction methods, and the method of this invention under different Reynolds numbers. The error indices used are Normalized Mean Square Error (NMSE) and L2 error (L2), which can be defined as follows:

[0075]

[0076]

[0077] Where x represents the actual global wind field. Global wind field reconstructed for deep learning models The wind field is the time-averaged value.

[0078] As observed in Table 1, compared with the multilayer perceptron and convolutional neural network reconstruction methods, the method of the present invention significantly reduces the error indices of the instantaneous wind speed field of the entire area of ​​a square building, indicating that the present invention can improve the reconstruction accuracy of the wind field around the building.

[0079] Table 1:

[0080]

[0081] The method of this invention can improve the accuracy of traditional reconstruction techniques and more accurately reflect the physical properties of the actual building wind field. This invention achieves high-precision reconstruction of the wake of a square building under different Reynolds numbers and has good generalization performance. The reconstruction accuracy of this invention is less affected by the number and location of sparse wind speed measurement points, thus the wind field prediction is more robust.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of protection of the invention. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0083] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0084] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0085] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include ROM, RAM, disk, or optical disk, etc.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of high-precision reconstruction of a wind field around a square-section building, characterized in that: This involves acquiring spatially sparse wind speed measurement points collected by sensors, and then reconstructing the spatially sparse wind speed measurement points into a global wind speed distribution around the building based on an optimized Wasserstein generative adversarial network (WGAN-GP) with an embedded gradient penalty mechanism. The optimization training process of the Wasserstein generative adversarial network with the embedded gradient penalty mechanism includes the following steps: (1) Obtain wind speed measurement data of a building wind field under a certain Reynolds condition, and process the wind speed measurement data using the direct data simulation method to generate training set data of the real wind field wind speed distribution. (2) Construct a Wasserstein generative adversarial network based on an embedded gradient penalty mechanism; (3) The Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is trained using training set data of real wind field wind speed distribution. (4) The Wasserstein generative adversarial network based on the embedded gradient penalty mechanism is optimized and trained using transfer learning techniques.

2. The method according to claim 1, characterized in that: Step (1) includes: 1a) Obtain wind speed measurement data of a building wind field under a certain Reynolds condition; 1b) The wind speed measurement data of the building wind field is processed by direct numerical simulation to generate wind speed distribution data of multiple horizontal sections along the building height. 1c) The bicubic interpolation technique is used to interpolate the wind speed distribution data of each horizontal section to the grid positions that are uniformly distributed throughout the entire area, thereby generating global wind speed distribution data. 1d) The k-means clustering technique was used to determine the grid points with large wind speed temporal fluctuations in the building shear layer and building leeward zone of the global building wind field, and to extract the spatially sparse wind speed measurement points. 1e) Determine the location code of the sparse wind speed measurement point value based on the extracted sparse wind speed measurement point value; 1f) The location code is added to the global wind speed distribution data and combined with the spatially sparse wind speed measurement points to form a training set of real wind field wind speed distribution data.

3. The method for high-precision reconstruction of the wind field around a square cross-section building according to claim 2, characterized in that: Step 1d) includes dividing the entire building wind field into a building shear zone and a building leeward zone; using k-means clustering technology, with wind speed standard deviation as the criterion, clustering grid points with similar wind speed standard deviations, selecting two grid points with the largest wind speed time fluctuations in the building shear zone, and selecting three grid points with the largest wind speed time fluctuations in the building leeward zone, and extracting sparse wind speed values ​​based on the wind speed measurement data of the five grid points.

4. The method for high-precision reconstruction of wind field around a square cross-section building according to claim 2, characterized in that: Step 1f) includes storing global wind speed distribution data using a first-layer channel, storing the location codes corresponding to sparse wind speed measurement points in the global wind speed distribution data using a second-layer channel, and constructing global wind speed distribution data with a matrix dimension of N×M×L×2 in the real wind field wind speed distribution; storing sparse wind speed measurement point values ​​using C sparse channels, and constructing sparse wind speed measurement point values ​​with a matrix dimension of N×1×1×C in the real wind field wind speed distribution, where N represents the number of training samples; M×L represents the pressure measurement point dimension of the real spatial accuracy, and C represents the number of sparse wind speed measurement points.

5. The method according to claim 1, wherein: In step (2), the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism includes: The system consists of a generator and a discriminator. The generator generates spatial data distributions that approximate the real wind speed field. It employs a deep fully convolutional architecture, containing m consecutive convolutional blocks. Each convolutional block includes a deconvolutional layer, a batch normalization layer, and a ReLU (Revised Linear Unit) layer as the activation function. The deconvolutional layer contains K×K convolutional kernels for upsampling low-dimensional spatial features. The discriminator distinguishes the real data spatial distribution from the generator's output. It also employs a deep convolutional architecture, containing n consecutive convolutional blocks. Each convolutional block contains a convolutional layer and a LeakyReLU activation function module. The convolutional layer contains K×K convolutional kernels for extracting low-dimensional features from the high-dimensional data distribution. A fully connected layer is concatenated after the consecutive convolutional blocks to calculate the probability that the generated data distribution is the real distribution.

6. The method according to claim 1, wherein: In step (3), training the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism includes: The weights and biases of each layer of the network are initialized. For the generator module of the network, the mean square error loss function is minimized for iterative optimization. At the same time, it is trained in conjunction with the discriminator module. Through multiple training iterations, the discriminator's ability to detect falsehoods is improved, and the optimization of each weight and bias of the generator module is promoted.

7. A method of high-precision reconstruction of the wind field around a building with square cross-section according to claim 6, characterized in that: The mean squared error loss function is expressed as follows: Among them, y i x represents the actual wind speed distribution i For sparse wind speed measurement points, G(x) i ( ) represents the wind speed distribution reconstructed by the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism at the corresponding time step.

8. The method according to claim 1, wherein: In step (4), the optimization training of the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism includes: Obtain a Wasserstein generative adversarial network trained under a certain Reynolds condition and based on an embedded gradient penalty mechanism; obtain wind speed measurement data of building wind fields under different Reynolds conditions and generate transfer data of the real wind field wind speed distribution; use the transfer data of the real wind field wind speed distribution and training steps to optimize and train the Wasserstein generative adversarial network based on the embedded gradient penalty mechanism.

Images