A residential building indoor wind field real-time prediction method based on a generative adversarial network

Abstract

The application relates to a residential building indoor wind field real-time prediction method based on a generative adversarial network, which comprises the following steps: generating a residential building three-dimensional model; model coding is performed on the residential building three-dimensional model to generate a residential building coded image; indoor wind field simulation is performed on the residential building three-dimensional model to generate a wind field image; the residential building coded image is taken as the input of a Pix2pix model, the wind field image is taken as the output of the Pix2pix model, and the Pix2pix model after training is used for residential building wind field prediction. Compared with the prior art, the application has the advantages of fast prediction speed, high prediction accuracy and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of indoor wind environment prediction, and particularly relates to a residential building indoor wind field real-time prediction method based on a generative adversarial network. BACKGROUND

[0002] In the research of building environment prediction, many scholars have used data-driven surrogate models to improve the prediction speed while maintaining the prediction accuracy. For example, machine learning methods are used to predict building wind pressure; an artificial neural network-based surrogate model is established to predict indoor airflow patterns and temperature distribution. However, environmental prediction focusing on partial key coefficients is difficult to help solve local problems such as indoor space layout and floor plan design.

[0003] Image prediction is beneficial to intuitive feedback of building environment, and generative adversarial networks have promoted the development of image prediction. Some studies use Pix2pix to predict building wind pressure images, use the Pix2pix model to predict outdoor wind environment, comfort and solar radiation in real time, and accelerate environment performance-driven urban design.

[0004] In the field of indoor environment prediction, some studies use CNN and GAN as a planar daylight simulation surrogate model to predict indoor static, annual daylight metrics and spatial illuminance distribution. Some scholars have created a new boundary condition CGAN model for indoor two-dimensional airflow distribution image prediction. However, indoor wind field prediction research is mostly based on fixed boundary space or single space, and it is currently difficult to achieve global wind environment prediction in diverse indoor spaces. SUMMARY

[0005] The purpose of the present application is to provide a residential building indoor wind field real-time prediction method based on a generative adversarial network, which efficiently and accurately predicts the natural ventilation wind field of general residential indoor spaces with indefinite boundaries and multiple spaces.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A residential building indoor wind field real-time prediction method based on a generative adversarial network, comprising the following steps:

[0008] S1, generating a residential building three-dimensional model;

[0009] S2, model encoding the residential building three-dimensional model to generate a residential building encoded image;

[0010] S3, simulating the indoor wind field of the residential building three-dimensional model to generate a wind field image;

[0011] S4, taking the residential building coding image as the input of the Pix2pix model and taking the wind field image as the output of the Pix2pix model, and using the trained Pix2pix model to perform residential building wind field prediction.

[0012] The S1 comprises the following steps:

[0013] S11, generating a parameterized residential plan;

[0014] S12, obtaining a real residential plan;

[0015] S13, establishing a two-level form control index;

[0016] S14, processing the parameterized residential plan and the real residential plan based on the two-level form control index, respectively, to generate a residential building three-dimensional model based on parameterization and a residential building three-dimensional model based on a real plan.

[0017] The S11 is specifically: by randomly setting the parameters of the plan outer contour shape and size, the entrance door position, the window opening surface, etc., a plan with indefinite boundary, multiple rooms and diversified layout is parameterized generated.

[0018] The two-level form control index comprises floor height, door width, door height, window sill height, window width, window height and window-wall ratio.

[0019] In the S2, the geometric data of the residential building three-dimensional model is coded into a 256*256*3 three-order tensor.

[0020] In the S2, the distance from the indoor space to the wall is considered in the image coding, and the plan geometric information, height information and the distance from the indoor space to the wall are superimposed in the depth direction, wherein the first layer information of the coding is the plan geometric information, the height of the door hole lower edge from the ground, the height of the window hole lower edge from the ground and the floor height; the second layer information is the height of the door hole upper edge from the ground and the height of the window hole upper edge from the ground; and the third layer information is the distance from the indoor space to the entity wall.

[0021] The method for obtaining the distance from the indoor space to the entity wall is: a monitoring surface with a preset size is set at a preset height from the ground, and the monitoring surface is divided into 256*256 evenly distributed monitoring points, and the Manhattan distance from the monitoring points inside the room to the nearest wall is calculated as the distance from the indoor space to the entity wall.

[0022] In the S2, the height and distance data in each layer of information are normalized and mapped to integers in the range of 0-255, and the three layers of data are respectively corresponded to RGB three channel values to form a residential building coding image.

[0023] Specifically, S3 involves using the Butterfly plugin to simulate the indoor wind field of a 3D model of a residential building, setting an upper limit for the simulated wind speed data and normalizing it to an integer range of 0 to 255, and generating wind field image data.

[0024] The dataset, consisting of one-to-one correspondence of coded residential building images and wind field images, is divided into training, testing, and validation sets in a ratio of 0.70:0.15:0.15.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) This invention considers diverse multi-space residential buildings with uncertain boundaries and predicts the details of the global wind field in the room, which is more comprehensive and intuitive.

[0027] (2) This invention accelerates simulation through a proxy model. Compared with traditional numerical simulation methods, the model trained on a generative adversarial network has a shorter simulation time and is more efficient in predicting the natural wind environment of general residential buildings with uncertain boundaries and multiple spaces. It can provide rapid and effective feedback on the indoor wind field of residential buildings in the early stages of the design process, which is conducive to the design of healthy, comfortable, low-carbon and energy-saving residential buildings. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention;

[0029] Figure 2 A schematic diagram of the modeling process for parametric generative housing;

[0030] Figure 3 A schematic diagram of image encoding for a 3D model of a residential building;

[0031] Figure 4 This is a schematic diagram of the output result of the Pix2pix model in one embodiment. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0033] This embodiment provides a method for real-time prediction of indoor wind fields in residential buildings based on generative adversarial networks, such as... Figure 1 As shown, it includes the following steps:

[0034] S1 generates a 3D model of the residential building.

[0035] Specifically, S1 includes the following steps:

[0036] S11, generate parameterized residential plan.

[0037] Batch parameterized residential plans are generated in grasshopper according to the PlanFinder plug-in, which can generate multi-room residential plans according to input parameters such as external contour, entrance door position, and windowed surface. By setting the value range of these parameters and performing random sampling, plans with indefinite boundaries, multiple rooms, and diversified layouts are generated.

[0038] S12, obtain real residential plan.

[0039] To enhance the authenticity and rationality of the data set plan, some real residential plans are also collected in this embodiment.

[0040] S13, establish secondary form control index.

[0041] The secondary form control index is established with reference to the residential building design specification, energy saving specification, and commonly used sizes, including floor height, door width, door height, window sill height, window width, window height, and window-wall ratio.

[0042] S14, process the parameterized residential plan and the real residential plan based on the secondary form control index, respectively, to generate a residential building three-dimensional model based on the parameterized plan and a residential building three-dimensional model based on the real plan.

[0043] As shown in Figure 2 , three-dimensional models are generated based on the parameterized plan and the real residential plan using the secondary form control index, wherein the secondary form control index parameters of the parameterized plan are set as shown in Table 1, the window sill height and the window height of the real residential plan are set as uniform values of 0.9m and 1.5m, respectively, and the other parameters are set as in Table 1. In order to expand the data set and reduce overfitting, the window width of the real residential plan is changed to enhance the data of the real situation, and the window width is reduced by 20% and 30%, respectively.

[0044] Table 1

[0045]

[0046] S2, model coding is performed on the residential building three-dimensional model to generate a residential building coded image.

[0047] Since the input definition of the pix2pix algorithm is a 256*256*3 three-order tensor, the geometric data of the residential building three-dimensional model is coded into a 256*256*3 three-order tensor, as shown in Figure 3 (1).

[0048] Since the room wall surface and opening have a great influence on the indoor wind speed, the distance from the indoor space to the wall is considered in image coding. In order to maximize the geometric information contained in coding, the plane geometric information, height information and the distance from the indoor space to the wall are superimposed in the depth direction, wherein the first layer information of coding is the plane geometric information, the height of the lower edge of the door hole from the ground, the height of the lower edge of the window hole from the ground and the floor height; the second layer information is the height of the upper edge of the door hole from the ground and the height of the upper edge of the window hole from the ground; and the third layer information is the distance from the indoor space to the solid wall. In an embodiment, the method for obtaining the distance from the indoor space to the solid wall is as follows: a monitoring surface of 20m*20m is set at 1.5m from the ground, and the monitoring surface is divided into 256*256 evenly distributed monitoring points, the Manhattan distance from the monitoring points inside the room to the nearest wall is calculated as the distance from the indoor space to the solid wall, as shown in (2). Figure 3

[0049] The height and distance data in each layer of information are normalized and mapped to integers in the range of 0-255, and the three layers of data are respectively mapped to RGB three channel values to form a residential building coding image.

[0050] S3, indoor wind field simulation is performed on the residential building three-dimensional model to generate a wind field image.

[0051] Specifically, the indoor wind field simulation is performed on the residential building three-dimensional model using the butterfly plug-in, the upper limit of the simulated wind speed data is set and normalized to map to integers in the range of 0-255 to generate wind field image data.

[0052] S4, the residential building coding image is used as the input of the Pix2pix model, and the wind field image is used as the output of the Pix2pix model, and the trained Pix2pix model is used to predict the wind field of the residential building.

[0053] In this embodiment, the data set composed of one-to-one corresponding residential building coding image and wind field image is divided into training set, test set and validation set in the ratio of 0.70:0.15:0.15, and the Pix2pix model is trained, tested and verified.

[0054] This embodiment collects a comprehensive data set of 403 samples, including 319 generative residential data and 84 real plane based residential data. Figure 4 The prediction images of part of the test set generated after training of the Pix2pix model and the evaluation indexes are shown, and the prediction images of the model are basically consistent with the indoor wind field distribution information of the real images. The average values of MAPE, SSIM and R 2 of all test results are 2.6113%, 0.9798 and 0.9114 respectively, indicating that the prediction accuracy of the method is high.​

[0055] Regarding the computation time, the average prediction time for each residential sample using the traditional simulation software is about 1200 seconds, while the average prediction time for the same sample using the Pix2pix model of the present application is 0.8 seconds, which greatly reduces the overall time cost. These results demonstrate the effectiveness and efficiency of the present method in predicting the indoor natural ventilation wind field of a residence. In summary, the present method can efficiently and accurately predict the indoor natural ventilation wind field of a general residence with an indefinite boundary and multiple spaces, providing fast and effective feedback for the early stage of the residence design process.

[0056] The preferred embodiments of the present application are described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and changes without departing from the spirit and scope of the present application. Therefore, any technical solutions obtained by those of ordinary skill in the art on the basis of the concept of the present application through logical analysis, reasoning, or limited experiments, should be within the scope of protection determined by the claims.

Claims

1. A method for real-time prediction of indoor wind field in residential buildings based on generative adversarial networks, characterized in that, Includes the following steps: S1 generates a 3D model of the residential building; S2, encode the 3D model of the residential building to generate a coded image of the residential building; S3 simulates indoor wind fields in 3D models of residential buildings and generates wind field images. S4. Using coded images of residential buildings as input to the Pix2pix model and wind field images as output to the Pix2pix model, the trained Pix2pix model is used to predict wind fields in residential buildings. S1 includes the following steps: S11, Generate Parametric Residential Floor Plan: By randomly setting the shape and size of the outer contour of the floor plan, the location of the entrance door, and the parameters of the window opening, a floor plan with indefinite boundaries, multiple rooms, and diverse layouts is generated parametrically. S12, Obtain the actual residential floor plan; S13, establish secondary form control indicators, including floor height, door width, door height, windowsill height, window width, window height, and window-to-wall ratio; S14, based on the secondary morphological control index, the parametric residential plan and the real residential plan are processed respectively to generate a three-dimensional residential building model based on the parametric generation and a three-dimensional residential building model based on the real plan respectively. In step S2, the distance from the indoor space to the wall is considered in the image encoding, and planar geometric information, height information, and distance information from the space to the wall are superimposed in the depth direction. The first layer of encoded information is planar geometric information, the height of the bottom edge of the door opening from the ground, the height of the bottom edge of the window opening from the ground, and the floor height; the second layer of information is the height of the top edge of the door opening from the ground and the height of the top edge of the window opening from the ground; the third layer of information is the distance from the indoor space to the solid wall. The method for obtaining the distance from the indoor space to the solid wall is as follows: a monitoring surface of a preset size is set at a preset height above the ground, and the monitoring surface is divided into 256*256 evenly distributed monitoring points. The Manhattan distance from the monitoring point inside the room to the nearest wall is calculated as the distance from the indoor space to the solid wall. In step S2, the height and distance data in each layer of information are normalized and mapped to integers in the range of 0 to 255. The three layers of data are then mapped to the values ​​of the three RGB channels to form a residential building coded image.

2. The method for real-time prediction of indoor wind field in residential buildings based on generative adversarial networks according to claim 1, characterized in that, In step S2, the geometric data of the three-dimensional model of the residential building is encoded as a 256*256*3 third-order tensor.

3. The method for real-time prediction of indoor wind field in residential buildings based on generative adversarial networks according to claim 1, characterized in that, Specifically, S3 involves using the Butterfly plugin to simulate the indoor wind field of a 3D model of a residential building, setting an upper limit for the simulated wind speed data and normalizing it to an integer in the range of 0 to 255, and generating wind field image data.

4. The method for real-time prediction of indoor wind field in residential buildings based on generative adversarial networks according to claim 1, characterized in that, The dataset, consisting of one-to-one correspondence of coded residential building images and wind field images, is divided into training, testing, and validation sets in a ratio of 0.70:0.15:0.15.

Images