A method for identifying ground roughness in building wind farms based on multi-scale input neural network

Abstract

The present invention discloses a method for identifying the ground roughness of a building wind farm based on a multi-scale input neural network, comprising the following steps: S1. Divide at least two scales according to the size of the surrounding area of ​​the building, and collect satellite images; S2. Use the ESDU method to calibrate the ground roughness of the image; S3. Construct a multi-scale input convolutional neural network model, use independent convolutional neural networks to respectively train the sector graphs of each scale of the input, and output the results to the SoftMax layer; S4. In the SoftMax layer, the training output results of the sector graphs of multiple scales are integrated to calculate the Loss function; S5. Compare the convergence error of the Loss function value to obtain a classification model; S6. Input the image to be tested into the classification model for identification and classification, and obtain the classification result. The present invention is used to improve the accuracy and efficiency of the building wind farm test simulation.

Description

Technical Field

[0001] The present invention relates to the technical field of building structure wind field analysis and image recognition, and in particular to a method for identifying ground roughness of a building wind field based on a multi-scale input neural network. Background Art

[0002] The simulation of wind field in wind tunnel test of building structure needs to determine the ground roughness category in advance, and then determine the corresponding wind profile according to parameters such as basic wind pressure. Conventional methods for determining ground roughness category include code qualitative identification method, ESDU digital wind model method and terrain simulation test method. Among them, code qualitative identification method mainly judges the surrounding environment category of the structure to be tested based on the qualitative description of the site conditions in the code. The subjective influence factors are large, and the reliability of the judgment result of this method is low when there are many surrounding environmental elements. The terrain simulation test method involves model making, measurement point arrangement, etc., the test cycle is long, the efficiency is low, and it is difficult to implement in the case of many actual projects. The ESDU method determines the structural wind field category by calculating the wind force within a given range. This method has high reliability, but it involves complex calculation process, the overall implementation efficiency needs to be improved, and it is difficult for general technicians to fully master it. Therefore, how to identify the ground roughness of building wind field efficiently and accurately is an urgent problem to be solved in the simulation of structural wind engineering test.

[0003] With the development and application of computer vision technology, image recognition and classification methods based on deep learning have received widespread attention. Most of the existing image recognition and classification methods are applied to the recognition of single-scale images. Because the environmental elements around buildings include buildings, roads, rivers, grasslands, coasts, trees, etc., and the environmental elements within different radius have different effects on ground roughness, it is difficult for single-scale images to fully reflect the ground characteristics of building wind farms and meet the requirements for ground roughness recognition of building wind farms. At present, there is a lack of methods for identifying the ground roughness of building wind farms through images. Summary of the invention

[0004] The present invention provides a method for identifying ground roughness of a building wind farm based on a multi-scale input neural network, so as to overcome the shortcomings of low recognition accuracy and poor execution efficiency of conventional ground roughness judgment methods, and to improve the accuracy and efficiency of building wind farm test simulation.

[0005] The method for identifying the ground roughness of a building wind farm based on a multi-scale input neural network of the present invention comprises the following steps:

[0006] S1. Divide the surrounding area of ​​the building into at least two scales according to the size of the surrounding area, collect satellite images of different scales around the building to obtain sample images, preprocess the sample images into circles, and divide the circular sample images into at least 8 sectors;

[0007] S2. calibrating the surface roughness of the area within the sector diagram using the ESDU method;

[0008] S3. Construct a multi-scale input convolutional neural network model, use independent convolutional neural networks to train the fan graphs of each scale of the input respectively, and output the results to the SoftMax layer;

[0009] S4. The training output results of the fan-shaped graphs of various scales are fused in the SoftMax layer, and weights are assigned according to the scales during fusion. The final classification result is obtained after weighted calculation, and the Loss function is calculated according to the classification result and the ground roughness calibration label of the ESDU method;

[0010] S5. Compare the convergence error of the Loss function value. When the error is less than 0.5%, the model training is completed and the classification model is obtained; otherwise, the network weights are optimized by back propagation until the model Loss function converges to obtain the classification model;

[0011] S6. Make the scale division and the number and size of the fan-shaped diagram of the uncalibrated image to be tested consistent with the trained sample image, input the image to be tested into the classification model for recognition and classification, and obtain the classification result.

[0012] The present invention can quickly obtain satellite images of the surrounding area of ​​the building through existing technologies such as Baidu Maps, Amap, and Google Maps. Generally, the screenshot satellite images are rectangular. Based on the wind direction selection, the present invention processes the sample image into a circle, and divides each sample image into at least 8 sectors. Each sector of the same sample image represents a partition of the 360-degree wind direction, ensuring the number of wind directions and meeting the requirements of wind field simulation. The collected images include different scales, and the sample images of each scale are input into the convolutional neural network model for independent training. After training, the results are output at the SoftMax layer, and the output results of the multi-scale sample images are fused at the SoftMax layer to obtain a classification model.

[0013] The present invention also has the following preferred design:

[0014] The sample images are divided into three scales: 10Km, 5Km, and 2Km.

[0015] When the training output results of the fan-shaped graphs of various scales are integrated in the SoftMax layer, the following optimal weight combination can be adopted: the weight coefficient of the 10Km scale is 0.5, the weight coefficient of the 5Km scale is 0.3, and the weight coefficient of the 2Km scale is 0.2.

[0016] The sample image of the present invention is equally divided into 8 to 36 sectors, which meets the requirements of wind field simulation.

[0017] The ground roughness calibration tag of the ESDU method includes:

[0018] Label A: offshore seas and islands, coasts, lake shores and desert areas;

[0019] Label B: fields, countryside, jungle, hills, and sparsely populated towns and city suburbs;

[0020] Label C: Urban areas with densely built-up areas and low-rise buildings;

[0021] Label D: Urban areas with densely built-up areas and taller houses.

[0022] The multi-scale input convolutional neural network model of the present invention is an AlexNet convolutional neural network, including 5 convolutional layers and 2 fully connected layers, and the activation function adopts the ReLU function; the Loss function of the SoftMax layer adopts the cross entropy loss function. Other network models with the same effect can also be used for training.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The present invention uses a multi-scale input convolutional neural network model, and the training image samples can be quickly intercepted from the satellite map. The sample image is processed into a fan-shaped diagram. The influence of wind direction is fully considered during training, and the landform characteristics of the site around the building are fully expressed, which improves the robustness of the training model and makes the ground roughness classification results of the building wind farm more accurate and reliable. The trained classification model is used to identify the image of the site to be tested, and the classification efficiency is higher, which is more conducive to the development of building wind farm test simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a method for identifying ground roughness of a building wind farm based on a multi-scale convolutional neural network of the present invention;

[0026] Figure 2 is a flow chart of preprocessing of satellite images of the surrounding area of ​​a building in an embodiment of the present invention;

[0027] Figure 3 It is a multi-scale input convolutional neural network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments so that those skilled in the art can better understand and implement the technical solution of the present invention.

[0029] The ground roughness recognition method of building wind farm based on multi-scale input neural network, such as Figures 1 to 3 As shown in the figure, it mainly includes building a data set, training a classification model, and testing a model, which specifically includes the following steps:

[0030] S1. Divide the surrounding area of ​​the building into three scales according to the size of the surrounding area. In this embodiment, the sample image is divided into three scales of 10Km, 5Km, and 2Km. Collect satellite images of different scales around the building to obtain sample images. The sample images are preprocessed into circles, and the circular sample images are equally divided into 36 sectors. The satellite images can be directly intercepted in commercial map software such as Baidu Maps, Amap or Google Maps. The intercepted images should ensure feature integrity. It is recommended that the image size is not less than 227mm*227mm, and then the images are cropped into circles as sample images;

[0031] S2. calibrating the surface roughness of the area within the sector diagram using the ESDU method;

[0032] Specifically, the ground roughness calibration tags of the ESDU method include:

[0033] Label A: offshore seas and islands, coasts, lake shores and desert areas;

[0034] Label B: fields, countryside, jungle, hills, and sparsely populated towns and city suburbs;

[0035] Label C: Urban areas with densely built-up areas and low-rise buildings;

[0036] Label D: Urban areas with densely built-up areas and taller houses;

[0037] S3. Construct a multi-scale input convolutional neural network model, use independent convolutional neural networks to train the fan graphs of each scale of the input respectively, and output the results to the SoftMax layer;

[0038] S4. The training output results of the fan-shaped graphs of various scales are fused in the SoftMax layer. When fusion is performed, weights are assigned according to the scales. The final classification result is obtained after weighted calculation. The Loss function is calculated according to the classification result and the ground roughness calibration label of the ESDU method.

[0039] In this example, the following preferred weight combination is adopted: the weight coefficient of 10Km scale is 0.5, the weight coefficient of 5Km scale is 0.3, and the weight coefficient of 2Km scale is 0.2;

[0040] S5. Compare the convergence error of the Loss function value. When the error is less than 0.5%, the model training is completed and the classification model is obtained; otherwise, the network weights are optimized by back propagation until the model Loss function converges to obtain the classification model;

[0041] In this embodiment, Figure 3As shown in the figure, the multi-scale input convolutional neural network model is an AlexNet convolutional neural network, including 5 convolutional layers and 2 fully connected layers, and the activation function adopts the ReLU function; the Loss function of the SoftMax layer adopts the cross entropy loss function, and the ReLU function is the linear rectification function (Rectified Linear Unit), also known as the rectified linear unit, which is a commonly used activation function in artificial neural networks. The specific expression is as follows:

[0042] ReLU(x)=max(0,x), where

[0043] The cross entropy loss function is expressed as follows:

[0044]

[0045] Where N is the number of samples, M is the number of categories, which is 4; y ic is a sign function, if the true category of sample i is equal to c, it takes 1, otherwise it takes 0; is the network output distribution, that is, the predicted probability that the observed sample i belongs to category c,

[0046] In other embodiments, other network models with the same effect may also be used for training;

[0047] S6. Make the scale division and the number and size of the fan-shaped diagram of the uncalibrated image to be tested consistent with the trained sample image, input the image to be tested into the classification model for recognition and classification, and obtain the classification result.

[0048] By adopting the technical solution provided by the embodiment of the present invention, the geomorphic characteristics of the site around the building can be fully expressed through the multi-scale input convolutional neural network model, the robustness of the training model can be improved, and the ground roughness classification results of the building wind farm can be more accurate and reliable, with a classification accuracy rate of up to 91%. The trained classification model is used to identify the image of the site to be tested, and the classification efficiency is higher, which is more conducive to the development of building wind farm test simulation.

[0049] In other embodiments, the size of the building surrounding area can also be divided into 2 scales or more than 3 scales, and the sample image can be preprocessed into 8 to 36 equally divided fan-shaped images that meet the requirements of wind field simulation. Other network models with the same effect as the AlexNet convolutional neural network can also be used to train and obtain a classification model.

[0050] The above embodiments are only preferred embodiments of the present invention, but cannot be used as limitations of the invention. Any modifications and improvements based on the concept of the present invention should fall within the protection scope of the present invention. The specific protection scope shall be subject to the claims.

Claims

1. A method for identifying ground roughness of building wind farms based on a multi-scale input neural network, characterized in that: The steps include: S1. Divide the surrounding area of ​​the building into at least two scales according to the size of the surrounding area, collect satellite images of different scales around the building to obtain sample images, preprocess the sample images into circles, and divide the circular sample images into at least 8 sectors; S2. calibrating the surface roughness of the area within the sector diagram using the ESDU method; S3. Construct a multi-scale input convolutional neural network model. The multi-scale input convolutional neural network model is an AlexNet convolutional neural network, which includes 5 convolutional layers and 2 fully connected layers. The activation function uses the ReLU function. The Loss function of the SoftMax layer uses the cross entropy loss function. Independent convolutional neural networks are used to train the fan graphs of each scale of the input respectively, and the SoftMax layer outputs the results. S4. The training output results of the fan-shaped graphs of various scales are fused in the SoftMax layer, and weights are assigned according to the scales during fusion. The final classification result is obtained after weighted calculation, and the Loss function is calculated according to the classification result and the ground roughness calibration label of the ESDU method; S5. Compare the convergence error of the Loss function value. When the error is less than 0.5%, the model training is completed and the classification model is obtained; otherwise, the network weights are optimized by back propagation until the model Loss function converges to obtain the classification model; S6. Make the scale division and the number and size of the fan-shaped diagram of the uncalibrated image to be tested consistent with the trained sample image, input the image to be tested into the classification model for recognition and classification, and obtain the classification result.

2. The method for identifying ground roughness of a building wind farm based on a multi-scale input neural network according to claim 1 is characterized in that: The sample images are divided into three scales: 10Km, 5Km, and 2Km.

3. The method for identifying ground roughness of a building wind farm based on a multi-scale input neural network according to claim 2 is characterized in that: When the training output results of the fan-shaped graphs of multiple scales are fused in the SoftMax layer, the weight coefficient of the 10Km scale is 0.5, the weight coefficient of the 5Km scale is 0.3, and the weight coefficient of the 2Km scale is 0.

2.

4. The method for identifying ground roughness of a building wind farm based on a multi-scale input neural network according to claim 3 is characterized in that: The sample image is equally divided into 8 to 36 sectors.

5. The method for identifying ground roughness of a building wind farm based on a multi-scale input neural network according to claim 4 is characterized in that: The ground roughness calibration tag of the ESDU method includes: Label A: offshore seas and islands, coasts, lake shores and desert areas; Label B: fields, countryside, jungle, hills, and sparsely populated towns and city suburbs; Label C: Urban areas with densely built-up areas and low-rise buildings; Label D: Urban areas with densely built-up areas and taller houses.

Images