Array fan wind field control method and system based on deep learning

Abstract

The present invention belongs to the field of wind tunnel flow field control and discloses a method and system for controlling a fan array wind field based on deep learning. The method comprises: collecting the fan array speed distribution and the test area speed distribution under different working conditions, normalizing the data and constructing an RGB image data set, designing and training a deep convolutional neural network, predicting the speed distribution given a target speed distribution, and executing speed control on the fan array. The present invention can avoid directly modeling and analyzing the flow field coupling effect between multiple fans, fully utilizing the nonlinear fitting capability of deep convolutional neural networks for gridded data, converting target flow characteristic generation into image generation, establishing a mapping model of target wind field-fan speed, and ultimately achieving flow characteristic simulation of complex wind fields.

Description

Technical Field

[0001] The present invention relates to the field of array fan wind tunnel flow field control, and in particular to an array fan wind field control method and system based on deep learning. Background Art

[0002] Currently, the commonly used wind tunnel simulation technologies include active simulation wind tunnel technology and passive simulation wind tunnel technology. Active simulation wind tunnel technology has advantages over passive simulation wind tunnel technology, such as high turbulence, large integral scale, and simple adjustment. As one of the methods to achieve active simulation of the target wind field, array fan wind tunnel is gradually becoming the research frontier of active simulation wind tunnel technology. The control of array fans is the core factor for the array fan wind tunnel to generate high-quality wind fields, which determines whether the generated wind field meets the index requirements. At present, the most widely used wind field simulation technology at home and abroad is passive simulation wind tunnel technology, while the most widely used active simulation wind tunnel technology is the active control grid method. There is relatively little research on the technology of array fan control.

[0003] By independently controlling the speed of each fan in the array, any target wind field can be created within the wind tunnel test area. Taking a shear wind field as an example, the array fan speed is controlled in layers, with fans at different heights running at different speeds, while fans at the same height run at the same speed. This creates a stepped distribution of fan outlet speeds at different heights. This stepped velocity distribution continuously mixes as it moves downstream, ultimately forming a shear wind field with a velocity gradient close to the target.

[0004] Traditional array fan control methods are mostly based on model predictive control strategies. However, this method cannot effectively solve the mutual coupling effect between fans, makes it difficult to obtain an ideal wind speed field, and requires a huge parameter optimization cost to obtain the accurate control parameters of each fan.

[0005] In view of this, the present invention is proposed. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems in the related art to a certain extent, and provides an array fan wind field control method and system based on deep learning. The present invention utilizes the powerful nonlinear fitting ability of deep neural networks to well capture the complex fluid dynamics under the operation of array fans and realize the refined control of the velocity field. It can not only obtain an ideal uniform wind speed field, but also realize the simulation of complex wind speed fields.

[0007] To achieve the above object, the first technical solution adopted by the present invention is:

[0008] The deep learning-based array fan wind field control method includes:

[0009] Collect the speed distribution data of the fan array under different working conditions and the speed distribution data of the test area, wherein the speed distribution data and the speed distribution data of each working condition are collected simultaneously and correspond one to one to form multiple data pairs;

[0010] Converting the rotation speed distribution data in each data pair into a first RGB image and the speed distribution data into a second RGB image;

[0011] Training a deep convolutional neural network with the second RGB image as input and the first RGB image as output to establish an end-to-end mapping model;

[0012] Convert the target speed distribution into a target second RGB image, input the mapping model, and output a predicted speed distribution RGB image;

[0013] The pixel values ​​of the predicted speed distribution RGB image are parsed to generate fan control instructions to perform the target wind field simulation.

[0014] Preferably, the rotational speed distribution data in each data pair is converted into a first RGB image, and the speed distribution data is converted into a second RGB image by normalizing the rotational speed distribution data and the speed distribution data respectively, interpolating the normalized data matrices to preset sizes as single-channel data, and copying their respective single-channel data to form three-channel RGB images.

[0015] Preferably, the rotation speed distribution data is normalized as follows: , is the original speed of each fan, is the upper speed limit, is the normalized value of the speed;

[0016] The velocity distribution data is normalized as follows: , is the original wind speed at each measuring point in the test area, is the upper limit of wind speed, is the normalized value of wind speed.

[0017] Preferably, the first RGB image is generated by rotation speed distribution data and has a size of 3×224×224;

[0018] The second RGB image is generated from velocity distribution data and has a size of 3×224×224.

[0019] Preferably, the deep convolutional neural network is a ResNet50 architecture, and the training loss function is: , where MSE stands for mean square error, is the number of samples, y is the predicted speed distribution, is the actual speed distribution.

[0020] Preferably, analyzing the pixel values ​​of the predicted rotational speed distribution RGB image includes: extracting pixel values ​​from the predicted rotational speed distribution RGB image according to the physical coordinates of the fan, and calculating the target rotational speed.

[0021] Preferably, the target speed of the fan with physical coordinates (j, k) is The calculation formula is:

[0022] ,in The coordinates in the image representing the predicted speed distribution are The pixel value of the array is a, a is the number of array fans in the horizontal direction, b is the number of array fans in the vertical direction, The upper limit of the speed.

[0023] The second technical solution adopted in the present invention is:

[0024] The array fan wind farm control system based on deep learning includes:

[0025] An acquisition module is used to collect the speed distribution data of the fan array and the speed distribution data of the test area under different working conditions, wherein the speed distribution data and the speed distribution data of each working condition are collected simultaneously and correspond one to one to form multiple data pairs;

[0026] a conversion module, configured to convert the rotation speed distribution data in each data pair into a first RGB image, and the speed distribution data into a second RGB image;

[0027] Establishing a module for training a deep convolutional neural network, using the second RGB image as input and the first RGB image as output, to establish an end-to-end mapping model;

[0028] an output module, configured to convert the target speed distribution into a target second RGB image, input the mapping model, and output a predicted speed distribution RGB image;

[0029] The generation module is used to parse the pixel values ​​of the predicted speed distribution RGB image and generate fan control instructions to perform the target wind field simulation.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention can avoid directly modeling and analyzing the flow field coupling effect between multiple fans, fully utilize the nonlinear fitting ability of deep convolutional neural networks for gridded data, transform the target flow characteristics generation into image generation, establish a mapping model of target wind field-fan speed, and ultimately realize the flow characteristics simulation of complex wind fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic flow chart of a deep learning-based array fan wind field control method provided in accordance with one embodiment of the present invention;

[0033] Figure 2 A schematic diagram of a fan array provided in accordance with an embodiment of the present invention;

[0034] Figure 3 A schematic structural diagram of a deep learning-based array fan wind farm control system provided in accordance with one embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the implementation of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the drawings in the embodiments of this application. The described embodiments are part of the embodiments of this application, rather than all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and should not be understood as limiting this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0036] refer to Figure 1 The first embodiment of the present invention provides a deep learning-based array fan wind field control method, comprising the following steps:

[0037] S101 , collecting speed distribution data of the fan array and speed distribution data of the test area under different working conditions, wherein the speed distribution data and speed distribution data of each working condition are collected simultaneously and correspond one to one to form multiple data pairs.

[0038] As a specific example, Figure 2 As shown in the figure, the power system of a fan array wind tunnel consists of a x b fixed fan arrays (a fans arranged horizontally and b fans arranged vertically). Each fan is an axial flow fan with a diameter of d (e.g., 1 meter). By independently controlling the speed of these a x b fans, any desired wind field can be generated within the wind tunnel test area.

[0039] This step is to collect multi-dimensional data and structure the data pairs. Specifically, the system constructs different working conditions by adjusting the speed distribution of a×b fans, and uses a hot wire anemometer probe rack to perform mobile measurements in the test area. Each working condition collects m×m total m data on the test area cross section. 2 Each experimental condition generates a set of data pairs of speed distribution (a×b matrix) and wind speed distribution (m×m matrix), forming the basic data set for system training.

[0040] S102 , converting the rotation speed distribution data in each data pair into a first RGB image, and converting the speed distribution data into a second RGB image.

[0041] This step converts the collected data into an RGB image. Before conversion, the collected data must be preprocessed, including normalization and spatial interpolation and expansion. This step utilizes a technical chain of normalization, interpolation, and channel replication to convert the physical quantity matrix into a standard image format, solving the problem of adapting non-visual data to the CV network.

[0042] Specifically, the method for converting the rotational speed distribution data in each data pair into the first RGB image and the velocity distribution data into the second RGB image is as follows: normalize the rotational speed distribution data and the velocity distribution data respectively, interpolate the normalized data matrices to a preset size as single-channel data, and copy each single-channel data into a 3-channel RGB format (size is 3×224×224) to form an RGB image, so that the physical field data can adapt to the standard image processing network structure.

[0043] To adapt the collected data to image processing requirements, the speed distribution data and wind speed distribution data are linearly normalized and mapped to the range of 0-255 before being converted into RGB images. For each fan speed, the following normalization operation is performed: , is the original speed of each fan, is the upper speed limit, is the value after the speed is normalized; for the speed of each measuring point in the test area, the following normalization operation is performed: , is the original wind speed at each measuring point in the test area, is the upper limit of wind speed, is the normalized wind speed value; then, the normalized a×b rotation speed distribution and m×m speed distribution are expanded to 224×224 by interpolation, and then converted into RGB images of size 3×224×224 by copying.

[0044] S103: Train a deep convolutional neural network, using the second RGB image as input and the first RGB image as output, to establish an end-to-end mapping model.

[0045] In some preferred embodiments, the deep convolutional neural network uses the ResNet50 architecture. This invention creatively transforms the ResNet50 image classification network into a physical field generator, leveraging its powerful feature extraction capabilities to establish a wind speed-rotation speed mapping relationship. Its residual connection structure addresses the vanishing gradient problem in deep networks, ensuring effective learning of high-dimensional nonlinear mappings. An end-to-end image generation model is then established, using an RGB image of the wind speed distribution in the test area as input and an RGB image of the fan array rotation speed distribution as output. The model is trained using the MSE loss function to measure the difference between the predicted and actual rotation speeds, and the ADAM optimizer (learning rate 0.0001, batch size 4) is used.

[0046] In one embodiment, the training loss function is: , where MSE stands for mean square error, is the number of samples, y is the predicted speed distribution, is the actual speed distribution.

[0047] S104 , converting the target speed distribution into a target second RGB image, inputting the image into the mapping model, and outputting a predicted speed distribution RGB image.

[0048] Specifically, given the target speed distribution data of the test area, it is converted into a 3×224×224 target RGB image according to the method in S102, input into the deep neural network obtained in S103, and output is the predicted speed distribution RGB image corresponding to the target speed distribution.

[0049] S105 , analyzing the pixel values ​​of the predicted rotation speed distribution RGB image, and generating fan control instructions to execute target wind field simulation.

[0050] This step uses the coordinate mapping formula to achieve accurate correspondence between the image pixel position and the physical fan position, ensuring the feasibility of the prediction results.

[0051] Parsing the pixel values ​​of the predicted speed distribution RGB image includes extracting pixel values ​​from the predicted speed distribution RGB image according to the physical coordinates of the fan, and calculating the target speed.

[0052] As an example, for Figure 2 a*b fans in the middle, Figure 2 Take the fan at the physical coordinate (j, k) as an example, the target speed to be executed Calculated according to the following formula: ,in The coordinates in the image representing the predicted speed distribution are The pixel value of the array is a, a is the number of array fans in the horizontal direction, b is the number of array fans in the vertical direction, is the upper limit of the speed; get the target speed Then adjust the corresponding fan to the speed.

[0053] This embodiment of the present invention transforms the physical field mapping problem into an image generation task for the first time, leveraging the three-channel nature of RGB images to encode fluid dynamics parameters. This overcomes the limitation of traditional control methods that require explicit physical models. This transformation naturally aligns the gridded data characteristics (rotation speed matrix, wind speed matrix) with the spatial feature extraction capabilities of convolutional neural networks, avoiding the need to solve complex coupled flow field equations.

[0054] The embodiments of the present invention can avoid directly modeling and analyzing the flow field coupling effects between multiple fans, fully utilize the nonlinear fitting ability of deep convolutional neural networks for gridded data, transform the target flow characteristics generation into image generation, establish a mapping model of target wind field-fan speed, and ultimately realize the flow characteristics simulation of complex wind fields.

[0055] refer to Figure 3 The second embodiment of the present invention provides a deep learning-based array fan wind farm control system 300, which includes: an acquisition module 301, a conversion module 302, an establishment module 303, an output module 304, and a generation module 305. The specific functions of each module are as follows:

[0056] An acquisition module 301 is used to acquire speed distribution data of the fan array and speed distribution data of the test area under different working conditions, wherein the speed distribution data and speed distribution data of each working condition are acquired simultaneously and correspond one to one, forming multiple data pairs;

[0057] A conversion module 302 is configured to convert the rotation speed distribution data in each data pair into a first RGB image and the speed distribution data into a second RGB image;

[0058] Establishing module 303 for training a deep convolutional neural network, using the second RGB image as input and the first RGB image as output, to establish an end-to-end mapping model;

[0059] An output module 304 is configured to convert the target speed distribution into a target second RGB image, input the mapping model, and output a predicted speed distribution RGB image;

[0060] The generation module 305 is used to analyze the pixel values ​​of the predicted rotation speed distribution RGB image and generate fan control instructions to perform target wind field simulation.

[0061] It should be noted that the above explanation of the method embodiment is also applicable to the system of this embodiment and will not be repeated here.

[0062] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0063] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A fan array wind field control method based on deep learning, characterized in that: Include: Collect the speed distribution data of the fan array under different working conditions and the speed distribution data of the test area, wherein the speed distribution data and the speed distribution data of each working condition are collected simultaneously and correspond one to one to form multiple data pairs; Converting the rotation speed distribution data in each data pair into a first RGB image and the speed distribution data into a second RGB image; Training a deep convolutional neural network with the second RGB image as input and the first RGB image as output to establish an end-to-end mapping model; Convert the target speed distribution into a target second RGB image, input the mapping model, and output a predicted speed distribution RGB image; The pixel values ​​of the predicted speed distribution RGB image are parsed to generate fan control instructions to perform the target wind field simulation.

2. The deep learning-based array fan wind field control method according to claim 1, characterized in that: The rotation speed distribution data in each data pair is converted into a first RGB image, and the speed distribution data is converted into a second RGB image by normalizing the rotation speed distribution data and the speed distribution data respectively, interpolating the normalized data matrices to a preset size as single-channel data, and copying each single-channel data to form a three-channel RGB image.

3. The deep learning-based array fan wind field control method according to claim 2, characterized in that: The speed distribution data is normalized as follows: , is the original speed of each fan, is the upper speed limit, is the normalized value of the speed; The velocity distribution data is normalized as follows: , is the original wind speed at each measuring point in the test area, is the upper limit of wind speed, is the normalized value of wind speed.

4. The deep learning-based array fan wind field control method according to claim 1, characterized in that: The first RGB image is generated by the rotation speed distribution data and has a size of 3×224×224; The second RGB image is generated from velocity distribution data and has a size of 3×224×224.

5. The deep learning-based array fan wind field control method according to claim 1, characterized in that: The deep convolutional neural network is a ResNet50 architecture, and the training loss function is: , where MSE stands for mean square error, is the number of samples, y is the predicted speed distribution, is the actual speed distribution.

6. The deep learning-based array fan wind field control method according to claim 1, characterized in that: Parsing the pixel values ​​of the predicted speed distribution RGB image includes extracting pixel values ​​from the predicted speed distribution RGB image according to the physical coordinates of the fan, and calculating the target speed.

7. The deep learning-based array fan wind field control method according to claim 6, characterized in that: The target speed of the fan at physical coordinates (j, k) The calculation formula is: ,in The coordinates in the image representing the predicted speed distribution are The pixel value of the array is a, a is the number of array fans in the horizontal direction, b is the number of array fans in the vertical direction, The upper limit of the speed.

8. Array fan wind farm control system based on deep learning, characterized by: Include: An acquisition module is used to collect the speed distribution data of the fan array and the speed distribution data of the test area under different working conditions, wherein the speed distribution data and the speed distribution data of each working condition are collected simultaneously and correspond one to one to form multiple data pairs; a conversion module, configured to convert the rotation speed distribution data in each data pair into a first RGB image, and the speed distribution data into a second RGB image; Establishing a module for training a deep convolutional neural network, using the second RGB image as input and the first RGB image as output, to establish an end-to-end mapping model; an output module, configured to convert the target speed distribution into a target second RGB image, input the mapping model, and output a predicted speed distribution RGB image; The generation module is used to parse the pixel values ​​of the predicted speed distribution RGB image and generate fan control instructions to perform the target wind field simulation.

Images