A high-precision building environment prediction method and terminal based on point-by-point calculation

Abstract

The present invention discloses a high-precision building environment prediction method and terminal based on point-by-point calculation, the method comprising: obtaining a building space geometric model, characterizing the building space geometric model to obtain a characterization model; slicing the characterization model with each grid point of the building space geometric model as the center, obtaining a plurality of characterization model slices centered on each grid point; sequentially inputting each characterization model slice into a neural network model, the neural network model sequentially outputting the building performance field reference value of the grid points within a set range centered on the grid point corresponding to the input characterization model slice; calculating the average value of all building performance field reference values ​​of each grid point as the building performance field value of the grid point. The present invention can smooth out the jump of the building performance field value caused by the change of the slice by averaging the building performance field values ​​of the overlapping parts, and has higher accuracy.

Description

Technical Field

[0001] The present invention relates to the field of building technology science and technology, and in particular to a high-precision building environment prediction method and terminal based on point-by-point calculation. Background Art

[0002] High-density cities bring many environmental challenges. Take the wind environment as an example. High-density cities have problems such as urban heat islands and air pollution. Natural ventilation can play a good role in alleviating these problems. In order to improve the natural ventilation of cities, the relationship between the wind environment and the architectural design scheme is usually analyzed through the calculation of fluid dynamics equations (hereinafter referred to as CFD, computational fluid dynamics). In the prior art, the use of numerically discretized Navier-Stokes equations to simulate flow problems has made significant progress. Compared with wind tunnel tests and field measurements, computational fluid dynamics shows advantages such as low cost and high precision. Therefore, it has also been widely used in related research on building wind environment.

[0003] However, this cost advantage is only relative, and it usually takes hours or days to obtain results. On the other hand, CFD calculations are different from the working logic of designers, and require tedious physical parameter settings. Therefore, existing airflow simulation tools are rarely intuitively integrated into the tools commonly used by designers.

[0004] To this end, the prior art has proposed a variety of algorithms, such as a technology for quickly solving the incompressible NS equations, namely Fast Fluid Dynamics simulation, which has been used for fluid visualization in video games. Its application on GPU can realize real-time informative airflow simulation for a small building, which is 500-1500 times faster than the traditional CFD method based on CPU calculation.

[0005] For example, the porous media model (PMM) is applied to the wind environment simulation of building complexes. The basic idea is to simplify the urban geometry and divide the urban street canyons into interconnected areas, and then evaluate the ventilation conditions and pollutant diffusion between urban blocks.

[0006] For example, the boundary-lattice Boltzmann method based on forced boundary conditions studies the influence of particles on the wall boundary layer. It is based on the collision and propagation of particles and the numerical simulation of a discrete Boltzmann equation.

[0007] However, all of the above methods are based on mathematical models or empirical formulas. Such methods require a large number of iterative calculations and data inputs, and complex theoretical assumptions are required before calculation. The technical difficulty is relatively high, there are high technical barriers, and it is difficult to accurately predict complex architectural forms.

[0008] The prior art proposes a method using a GAN adversarial generative network. This method encodes the building geometry into the network in the form of a two-dimensional picture, thereby obtaining the wind environment calculation result of the building space. Although this method overcomes the shortcomings of the method based on mathematical models or empirical formulas, it is easy to implement and easy to operate. However, its structure is still relatively complex, requiring a large number of building wind environment CFD calculation cases (more than 10,000) for model training, and the error is high, and the prediction problem robustness is low.

[0009] To this end, on this basis, the existing technology adopts a point-by-point calculation method to transform the perspective from the relationship between fields to a point-based perspective, and explores the relationship between the field where the point is located and the performance field corresponding to the point. The data input to the neural network model is only related to the surrounding environment of the grid point. The model can input any number of grid points in an area of ​​any shape or size, thereby greatly reducing the complexity of the input dimension and the output dimension, and greatly reducing the number of training sets required for the model. Only a small number of calculation samples are required. The applicable scenarios of model calculation will be more diversified, and the accuracy and robustness of the acquisition will also be greatly improved.

[0010] However, the existing point-by-point calculation method uses the calculation point as the perspective to explore the relationship between the model slice centered on the point and the architectural environmental performance field of the point. As the exploration grid moves, the model slice also changes, which may cause jumps in the environmental performance field values ​​between adjacent points. In addition, since the model slice contains less information than the entire model, there is still room for further improvement in accuracy. Summary of the invention

[0011] The purpose of the present invention is to provide a high-precision building environment prediction method and terminal based on point-by-point calculation, aiming to solve the problems of jump and insufficient precision caused by model slicing problems in the prior art.

[0012] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0013] The present invention provides a high-precision building environment prediction method based on point-by-point calculation, and the high-precision building environment prediction method based on point-by-point calculation includes:

[0014] Acquire a building space geometric model, and characterize the building space geometric model to obtain a characterization model;

[0015] Slicing the representation model with each grid point of the building space geometric model as the center to obtain a plurality of representation model slices with each grid point as the center;

[0016] Inputting each of the characterization model slices into the neural network model in sequence, and the neural network model outputs the building performance field values ​​of the grid points in a set range centered on the grid point corresponding to the input characterization model slice in sequence;

[0017] The average value of all building performance field values ​​at each grid point is calculated as the building performance field value of the grid point.

[0018] Furthermore, the step of acquiring the building space geometric model and characterizing the building space geometric model to obtain a characterization model specifically includes:

[0019] Acquire a building space geometric model and a building space three-dimensional model, and discretize the building space geometric model to obtain a plurality of grid points;

[0020] The characterization values ​​of the respective grid points are calculated according to the architectural space geometric model to form a characterization value matrix, and the characterization value matrix is ​​used as a characterization model.

[0021] Furthermore, the characterization value of each grid point is calculated according to the building space geometric model to form a characterization value matrix, which specifically includes one or both of the following:

[0022] Calculate the binary representation value of each grid point according to the building space geometric model to form a binary representation matrix;

[0023] The signed distance field representation value of each grid point is calculated according to the building space geometric model to form a signed distance field representation matrix.

[0024] Furthermore, the step of calculating the binary representation value of each grid point according to the building space geometric model to form a binary representation matrix specifically includes:

[0025] If the grid point is where the obstacle is located, the binary number representation value of the grid point is 1;

[0026] If the grid point is not the grid point where the obstacle is located, the binary number representation value of the grid point is 0.

[0027] Furthermore, the step of calculating the signed distance field representation value of each grid point according to the building space geometric model to form a signed distance field representation matrix specifically includes:

[0028] If the grid point is where the obstacle is located, the value of the signed distance field representation of the grid point is 0;

[0029] If the grid point is not the grid point where the obstacle is located, the data of the minimum distance from the grid point to the nearest obstacle is calculated as the signed distance field representation value of the grid point.

[0030] Furthermore, slicing the representation model with each grid point as the center to obtain the representation model slice with each grid point as the center specifically includes:

[0031] Extending the representation model by the size of the slice radius all around to obtain an extended representation model;

[0032] The extended representation model is sliced ​​with each of the grid points as the center to obtain representation model slices with each of the grid points as the center.

[0033] Furthermore, when training the neural network model, the following steps are specifically included:

[0034] Acquire the architectural space geometric model of the training case, and characterize the architectural space geometric model of the training case to obtain the characterization model of the training case;

[0035] Sampling points in the representation model of the training case to obtain a plurality of sampling grid points, and obtaining a representation model slice of the training case centered at each sampling grid point;

[0036] Train the neural network model based on the representation model slices of the training cases.

[0037] Further, the building performance field is specifically a wind field, and the building performance field value is specifically a wind speed vector value;

[0038] The closer the sampling grid points are to the ground, the denser they are.

[0039] Furthermore, the high-precision building environment prediction method based on point-by-point calculation also includes:

[0040] A visualization view of the building performance field is drawn according to the building performance field values ​​of each of the grid points.

[0041] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, which includes: a memory, a processor, and a high-precision building environment prediction program based on point-by-point calculation stored in the memory and executable on the processor, wherein the high-precision building environment prediction program based on point-by-point calculation, when executed by the processor, controls the terminal to implement the steps of the high-precision building environment prediction method based on point-by-point calculation as described above.

[0042] The present invention adopts the above technical solution to achieve the following effects:

[0043] The present invention allows the neural network to output the numerical values ​​of multiple points centered on the representation model slice. For the input of multiple representation model slices, the building performance field values ​​output by the neural network model overlap. The building performance field values ​​of the overlapping parts are averaged to smooth out the jumps in the building performance field values ​​caused by the changes in the slices. At the same time, since different representation model slices contain different ranges of information, the building performance field value finally obtained can contain a larger range of information than the building performance field value of a single slice, and therefore can have higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flowchart of the steps of a high-precision building environment prediction method based on point-by-point calculation in a preferred embodiment of the present invention;

[0045] Figure 2 is a schematic diagram of dome shielding for calculating sky visibility factor in a preferred embodiment of the present invention;

[0046] Figure 3 1 is a schematic diagram of a plane projection of a dome occlusion for calculating a sky visibility factor in a preferred embodiment of the present invention;

[0047] Figure 4 is a schematic diagram of an example of calculating a sky visibility factor in a preferred embodiment of the present invention;

[0048] Figure 5 is another exemplary schematic diagram of calculating the sky visibility factor in a preferred embodiment of the present invention;

[0049] Figure 6 Schematic diagram of the steps for calculating the frontal area ratio in a preferred embodiment of the present invention;

[0050] Figure 7 is a schematic diagram representing a geometric model of a building space in a preferred embodiment of the present invention;

[0051] Figure 8 is a schematic diagram of a partial representation model slice in a preferred embodiment of the present invention;

[0052] Fig. 9 is a schematic diagram of a partial structure of a neural network model in a preferred embodiment of the present invention;

[0053] Fig.10 is a schematic diagram of establishing sampling grid points in a preferred embodiment of the present invention;

[0054] Fig.11 is an overlapping schematic diagram of the output of the neural network model of the preferred embodiment of the terminal of the present invention;

[0055] Fig.12 It is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0057] Embodiment 1

[0058] See also Figure 1 Embodiment 1 of the present application is a high-precision building environment prediction method based on point-by-point calculation. In this embodiment, it is taken as an example in a wind field, but it does not mean that the method is only applicable to wind field environments. It can also be applied to the calculation and acquisition of other building environments, such as light fields, thermal fields or sound fields.

[0059] S1. Obtain a building space geometric model, and characterize the building space geometric model to obtain a characterization model.

[0060] Specifically, the architectural space geometry model obtained in this embodiment may be an architectural space geometry model of a block section, or an architectural space geometry model of an urban plane building.

[0061] After obtaining the architectural space geometric model, the architectural space geometric model needs to be represented in a format that can be understood by the neural network model.

[0062] In this embodiment, obtaining a building space geometric model and characterizing the building space geometric model to obtain a characterization model specifically includes:

[0063] A building space geometric model and a building space three-dimensional model are obtained, and the building space geometric model is discretized to obtain a plurality of grid points.

[0064] The said meshing of the architectural space geometric model is to represent the architectural space by point cloud method (point-based model) or triangle mesh method (triangle mesh), but such methods have huge data volume and complex structure, which greatly increases the complexity of the neural network. Therefore, in this implementation, it is specifically represented by (x n ,yn ) is a two-dimensional matrix (x, y) formed by precision. Among them, x is the horizontal coordinate of the two-dimensional matrix, y is the vertical coordinate of the two-dimensional matrix, x n is the grid representation accuracy of the horizontal axis, y n is the grid representation accuracy of the ordinate, x n ,y n The closer it is to 0, the more the two-dimensional matrix (x, y) can fully reflect the environment of the building space. However, due to the limitation of computational cost, (x n ,y n ) takes the value of (1,1), which means that for any point in space, (1,1) is the smallest dividing unit.

[0065] The three-dimensional model information parameters of each of the grid points are calculated according to the three-dimensional model of the building space and a three-dimensional model information parameter matrix is ​​formed. The characterization value of each of the grid points is calculated according to the geometric model of the building space and a characterization value matrix is ​​formed. The three-dimensional model information parameter matrix and the characterization value matrix are combined into a characterization model.

[0066] Specifically, in this embodiment, the three-dimensional model information parameters of each grid point of the gridded building space geometric model are calculated according to the three-dimensional model of the building space to form a three-dimensional model information parameter matrix, which specifically includes one or both of the following:

[0067] Calculating the sky visible factor three-dimensional model information parameters of each grid point of the gridded building space geometric model according to the building space three-dimensional model to form a sky visible factor information matrix;

[0068] The windward area ratio three-dimensional model information parameters of each grid point of the gridded building space geometric model are calculated according to the building space three-dimensional model to form a windward area ratio information matrix.

[0069] The step of calculating the sky visibility factor of each grid point of the gridded building space geometric model according to the building space three-dimensional model specifically includes:

[0070] If the grid point is where the obstacle is located, the sky visibility factor three-dimensional model information parameter of the grid point is zero;

[0071] Please refer to Figure 2 If the grid point is not the grid point where the obstacle is located, a hemispherical dome is generated with the grid point as the center, the occlusion projection of each obstacle in the three-dimensional model of the building space on the dome is obtained, and the area ratio of the unobstructed dome in the dome is calculated as the sky visibility factor three-dimensional model information parameter of the grid point.

[0072] Please refer to Figure 4, calculate the area ratio of the unobstructed dome in the dome, and the obstruction projection of each obstacle on the dome can be further projected as a projection on a plane prototype with the grid as the prototype, and calculate the area of ​​the projection on the circle.

[0073] For example, see Figure 4 and Figure 5 As shown, Figure 4 is a schematic diagram of an example of calculating the sky visibility factor in a preferred embodiment of the present invention, Figure 5 is another exemplary schematic diagram of calculating the sky visibility factor in a preferred embodiment of the present invention, Figure 4 and Figure 5 In the figure, a represents the projection of the obstacle on the horizontal plane, b represents the three-dimensional model of the building space, and c represents the plane projection of the occlusion projection of each obstacle in the three-dimensional model of the building space on the dome.

[0074] Please refer to Figure 6 , calculating the windward area ratio three-dimensional model information parameters of each grid point of the gridded building space geometric model according to the building space three-dimensional model to form a windward area ratio information matrix, specifically including:

[0075] If the grid point is the grid point where the obstacle is located, the windward area ratio three-dimensional model information parameter of the grid point is zero;

[0076] If the grid point is not the grid point where the obstacle is located, the design wind direction of the building space is obtained, and a plane perpendicular to the design wind direction and passing through the grid point is established for each design wind direction. The height of the plane is twice the highest height of the three-dimensional model of the building space, and the width is the width of the three-dimensional model of the building space on the plane. The intersection area A of each plane and the obstacle is obtained. proj , calculate the intersection area A proj The area A of the corresponding plane T The ratio of the windward area to the plane corresponding to the wind direction is λ f (θ)_e:

[0077]

[0078] Among them, λ f (θ)_e represents the windward surface area ratio under a specific wind direction θ, A proj A represents the total area of ​​each obstacle projected onto the plane, that is, the total area of ​​the intersection of the plane and the obstacle. T Represents the plane area, and the weighted sum of the windward surface area ratios of all wind directions at each grid point is calculated as the three-dimensional model information parameter λ of the windward area ratio of the grid point f _e:

[0079]

[0080] λf _e is the weighted sum of the windward surface area ratios under N specific wind directions, where N is the number of specific wind directions, and B θ is a weighted value corresponding to a specific wind direction θ, which may be the frequency of the corresponding specific wind direction θ.

[0081] Calculating the representation value of each of the grid points according to the building space geometric model specifically includes:

[0082] Calculating the binary number representation value of each grid point according to the building space geometric model;

[0083] The signed distance field (SDF) representation value of each grid point is calculated according to the building space geometric model.

[0084] The binary number representation value of each grid point is calculated according to the building space geometric model in a Cartesian coordinate system with an accuracy of 1 meter. The grid points where obstacles are located are represented as 1, and the grid points without obstacles are represented as 0, forming a two-dimensional matrix with a matrix dimension of (x, y).

[0085] Please refer to Figure 7 , the directed distance field representation value of each grid point is calculated according to the building space geometric model, specifically in a Cartesian coordinate system with an accuracy of 1 meter, the grid point where the obstacle is located is represented as zero, and the value of the grid point without obstacle is represented as the data of the minimum distance between the point and the nearest obstacle, forming a two-dimensional matrix with a matrix dimension of (x, y).

[0086] The minimum distance is determined by the sum of the number of grid positions that the point moves in the horizontal direction or the vertical direction.

[0087] Through experiments and verification, the present invention represents the grid point where the obstacle is located as zero, and represents the value of the grid point without obstacle as the data of the minimum distance between the point and the nearest obstacle. It has global information and can reflect the distance and shape of the nearest obstacle in the entire space to a certain extent in the spatial slice of any point. Therefore, the prediction effect is more excellent.

[0088] S2. Slice the representation model with each grid point of the representation model as the center, and obtain a plurality of representation model slices with each grid point as the center.

[0089] In this embodiment, slicing the representation model with each grid point as the center specifically includes:

[0090] The slice radius will be extended all around the characterization model.

[0091] In this embodiment, the size of the original representation model is 360m in the horizontal direction and 120m in the vertical direction. The predetermined slice range in this embodiment is 120m, so in this embodiment, the upper, lower, left and right sides of the representation model are extended by 120 meters. The purpose of this is to still sample the environmental data around the sampling point at the edge of the representation model.

[0092] Please refer to Figure 8 , the extended representation model is sliced ​​with each grid point as the center to obtain multiple Figure 8 Slices shown.

[0093] In this embodiment, for any point (x, y) in space, the surrounding environment S (x,y) It can be expressed as the perimeter of the point (x m ,y m ) in the range of two-dimensional matrix (x,y), when x m ,y m As S approaches infinity, (x,y) The more it can fully reflect the surrounding environment of the point, but due to the limitation of computational cost, in this paper (x m ,y m ) takes the value of (120,120), which means that for any grid point in the space, a slice with a surrounding range of (120,120) is obtained, that is, a slice of size (241,241) is obtained, including the sampling grid point itself.

[0094] S3. Input each of the characterization model slices into the neural network model in turn, and the neural network model outputs the building performance field reference value of the grid points in a set range centered on the grid point corresponding to the input characterization model slice in turn.

[0095] Since the wind speed vector value V corresponding to the range centered at any point (x, y) in space is (x,y) Only with the surrounding environment S (x,y) Therefore, the building performance field value around the point can be predicted based on the representation model slice of its surrounding environment.

[0096] In this embodiment, the building performance field value specifically refers to the wind speed vector value. In other optional embodiments, it may also be light intensity or temperature, etc.

[0097] Please refer to Fig. 9 The neural network model specifically includes a Vgg-CFD-19 neural network model constructed based on the classic Vggnet deep convolutional neural network. Figure 6 A partial structure thereof is shown. In other optional embodiments, neural network models with other network structures such as ResNet and Inception may also be selected, but the present invention is not limited thereto.

[0098] Vgg-CFD-19 consists of 5 sequentially connected blocks and 4 fully connected layers. Each block consists of a sequentially connected convolution part, a pooling layer, a normalization layer and an activation function.

[0099] The convolution part of the first block of Vgg-CFD-19 includes two convolution layers connected in sequence, the convolution part of the second block of Vgg-CFD-19 includes two convolution layers connected in sequence; the convolution part of the third block of Vgg-CFD-19 includes four convolution layers connected in sequence; the convolution part of the fourth block of Vgg-CFD-19 includes four convolution layers connected in sequence; the convolution part of the fifth block of Vgg-CFD-19 includes four convolution layers connected in sequence.

[0100] It is worth noting that in the research process of the present invention, it was found that as the depth of the model increased, the prediction performance of the model gradually improved, but after the depth of the model reached a certain level, the effect decreased instead. On the one hand, it may be due to the fact that the training process of the model has a certain randomness, and this randomness may occasionally cause the accuracy of a deeper network to decrease instead. On the other hand, the negative effects brought about by the increase in the depth of the neural network, such as gradient explosion and other phenomena, can no longer offset the positive effects it brings. Vgg-CFD-19 is the best model obtained in the research process of the present invention.

[0101] The neural network model of the prior art establishes the relationship between fields. The neural network model specifies the size and dimension of the field through preset definitions. In the model calculation, the scene setting needs to follow this dimension, and its dimension cannot be changed. The data input by the neural network model of this embodiment is only related to the surrounding environment of the grid. The model can input any number of grid points in areas of any shape or size, and the applicable scenarios of the model calculation will be more diversified.

[0102] The training samples of the model are CFD calculation cases of more than 100 cases, among which the CFD calculation cases can be obtained by calculation using common CFD calculation software (such as OpenFOAM, Fluent, CFX, Start-CCM+, etc.). According to the trained model, the corresponding actual urban cases or ideal block models are selected as cases according to the type of object to be calculated.

[0103] For the calculated training cases and the corresponding calculation results,

[0104] For the architectural space geometric model of the training case, the architectural space geometric model is characterized to obtain a representation model;

[0105] The points in the representation model are sampled to obtain a plurality of sampling grid points, and a slice of the representation model centered at each grid point is obtained.

[0106] Please refer to Fig.10 Specifically, the sampling method can be uniform sampling at intervals of 1 meter, or non-uniform sampling with reference to CFD calculation. In this embodiment, because the wind field is studied, in the relevant research on the wind environment, the wind speed near the ground is considered to be more important. Therefore, in this embodiment, a non-uniform network is specifically used for sampling to enable the model to obtain a better training effect, specifically, including:

[0107] Establish horizontal sampling lines and vertical sampling lines;

[0108] The intersection of the horizontal sampling line and the vertical sampling line is used as the sampling grid point.

[0109] In this embodiment, the vertical sampling lines are evenly arranged in the characterization model before extension, with an interval of 2m, and the height of the horizontal sampling lines is rounded to an exponential multiple of 1.2, and finally 21 horizontal sampling lines and 180 sampling lines are obtained, obtaining 21×180=2160 sampling grid points.

[0110] After that, the 100 training cases are multiplied into a large number of training cases, which are divided into training sets and validation sets in proportion to train the model.

[0111] The loss function during model training is set to MSE (Mean Square Error):

[0112]

[0113] In the formula, m represents the number of samples, y i represents the true value, Represents the output value of the model.

[0114] The ADAM optimization method is used for training, and the maximum iteration number of the model is set to 100 epochs. In other embodiments, AdaGrad or RMSProp methods may also be used for optimization training, but the present invention is not limited thereto.

[0115] In this embodiment, the representation model slices centered at each grid point are input into a neural network model, and the building performance field value of each grid point is obtained according to the output of the neural network model, specifically including:

[0116] The representation model slices centered on each grid point are sequentially input into the neural network model, and the neural network model sequentially outputs the building performance field values ​​of the grid points within a set range centered on the grid point corresponding to the input representation model slice;

[0117] S4. Calculate the average value of all building performance field reference values ​​at each grid point as the building performance field value of the grid point.

[0118] Please refer to Fig.11 In this embodiment, by outputting the building performance field value of a set range centered on the center grid point of the slice in each calculation, for example, outputting the building performance field value of a grid point in the range of 3*3 or 4*4, the building performance field values ​​output by the neural network model overlap for the input of multiple characterization model slices. The building performance field values ​​of the overlapping parts are averaged, which can smooth the jumps in the building performance field values ​​caused by the changes in the slices, so as to achieve the smoothing of the building performance output value prediction of the entire building space geometric model and increase the prediction accuracy.

[0119] S5. Draw a visualization of the building performance field based on the building performance field values ​​of each grid point.

[0120] Specifically, in this embodiment, the wind speed vector data V_pre of each grid point is obtained. (x,y) , they will be rearranged according to the original coordinates (x, y) to obtain wind field vector data. These data will be plotted into wind speed cloud charts through matlibplot or connected to the Grasshopper platform through python scripts to obtain visualization results.

[0121] In addition, this embodiment uses the wind field as the performance field only as an illustrative distance, and does not mean that the technical solution of the present invention is only applicable to the wind field. For example, the method of the present invention can also be used in geothermal fields. As a neural network model widely used in image and video processing tasks, the convolutional neural network has good spatial resolution capabilities, which enables it to be applied to various performance fields. The local receptive field and weight sharing characteristics of the convolution kernel play an important role in parsing the mapping relationship between the building wind environment and the building space.

[0122] Embodiment 2

[0123] See also Fig.12 Based on the above method, the present invention also provides a terminal, which includes: a memory 10, a processor 20, and a high-precision building environment prediction program based on point-by-point calculation stored in the memory 10 and executable on the processor 20. When the high-precision building environment prediction program based on point-by-point calculation is executed by the processor 20, the terminal is controlled to implement the steps of the high-precision building environment prediction method based on point-by-point calculation as described above.

[0124] In some embodiments, the memory 10 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 10 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 10 may also include both an internal storage unit of the terminal and an external storage device. The memory 10 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 10 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a high-precision building environment prediction program based on point-by-point calculation is stored on the memory 10, and the high-precision building environment prediction program based on point-by-point calculation can be executed by the processor 20, thereby realizing the high-precision building environment prediction method based on point-by-point calculation in the present application.

[0125] In some embodiments, the processor 20 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 10, such as executing the high-precision building environment prediction method based on point-by-point calculation.

[0126] Embodiment 3

[0127] This embodiment provides a storage medium, which stores a high-precision building environment prediction program based on point-by-point calculation. When the high-precision building environment prediction program based on point-by-point calculation is executed by a processor, the steps of the high-precision building environment prediction method based on point-by-point calculation as described above are implemented.

[0128] In summary, the present invention allows the neural network to output the numerical values ​​of multiple points centered on the representation model slice. For the input of multiple representation model slices, the building performance field values ​​output by the neural network model overlap. The building performance field values ​​of the overlapping parts are averaged to smooth out the jumps in the building performance field values ​​caused by the slice changes. At the same time, since different representation model slices contain different ranges of information, the final building performance field value obtained can contain a larger range of information than the building performance field value of a single slice, and therefore can have higher accuracy.

[0129] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0130] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium, and the program can include the processes of the above-mentioned method embodiments when executed. The storage medium can be a memory, a disk, an optical disk, etc.

[0131] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A high-precision building environment prediction method based on point-by-point calculation, characterized in that: The high-precision building environment prediction method based on point-by-point calculation includes: Acquire a building space geometric model, and characterize the building space geometric model to obtain a characterization model; Slicing the representation model with each grid point of the building space geometric model as the center to obtain a plurality of representation model slices with each grid point as the center; Inputting each of the characterization model slices into the neural network model in sequence, and the neural network model outputs the building performance field reference value of the grid points in a set range centered on the grid point corresponding to the input characterization model slice in sequence; Calculate the average value of all building performance field reference values ​​at each grid point as the building performance field value at that grid point; The acquiring of the building space geometric model and characterizing the building space geometric model to obtain the characterization model specifically includes: Acquire a building space geometric model and a building space three-dimensional model, and discretize the building space geometric model to obtain a plurality of grid points; Calculate the representation value of each grid point according to the building space geometric model to form a representation value matrix, and use the representation value matrix as a representation model; Slicing the representation model with each grid point as the center to obtain the representation model slice with each grid point as the center specifically includes: Extending the representation model by the size of the slice radius all around to obtain an extended representation model; The extended representation model is sliced ​​with each of the grid points as the center to obtain representation model slices with each of the grid points as the center.

2. According to the high-precision building environment prediction method based on point-by-point calculation in claim 1, the characterization value of each grid point is calculated according to the building space geometric model to form a characterization value matrix, which specifically includes one or both of the following: Calculate the binary representation value of each grid point according to the building space geometric model to form a binary representation matrix; The signed distance field representation value of each grid point is calculated according to the building space geometric model to form a signed distance field representation matrix.

3. The high-precision building environment prediction method based on point-by-point calculation according to claim 2 is characterized in that: The step of calculating the binary number representation value of each grid point according to the building space geometric model to form a binary number representation matrix specifically includes: If the grid point is where the obstacle is located, the binary number representation value of the grid point is 1; If the grid point is not the grid point where the obstacle is located, the binary number representation value of the grid point is 0.

4. The high-precision building environment prediction method based on point-by-point calculation according to claim 2 is characterized in that: The step of calculating the signed distance field representation value of each grid point according to the building space geometric model to form a signed distance field representation matrix specifically includes: If the grid point is where the obstacle is located, the value of the signed distance field representation of the grid point is 0; If the grid point is not the grid point where the obstacle is located, the data of the minimum distance from the grid point to the nearest obstacle is calculated as the signed distance field representation value of the grid point.

5. The high-precision building environment prediction method based on point-by-point calculation according to claim 1 is characterized in that: The neural network model, when being trained, specifically includes: Acquire the architectural space geometric model of the training case, and characterize the architectural space geometric model of the training case to obtain the characterization model of the training case; Sampling points in the representation model of the training case to obtain a plurality of sampling grid points, and obtaining a representation model slice of the training case centered at each sampling grid point; Train the neural network model based on the representation model slices of the training cases.

6. The high-precision building environment prediction method based on point-by-point calculation according to claim 5 is characterized in that: The building performance field is specifically a wind field, and the building performance field value is specifically a wind speed vector value; The closer the sampling grid points are to the ground, the denser they are.

7. The high-precision building environment prediction method based on point-by-point calculation according to claim 1 is characterized in that: The high-precision building environment prediction method based on point-by-point calculation also includes: A visualization view of the building performance field is drawn according to the building performance field values ​​of each of the grid points.

8. A terminal, characterized in that: The terminal includes: a memory, a processor, and a high-precision building environment prediction program based on point-by-point calculation stored in the memory and executable on the processor. When the high-precision building environment prediction program based on point-by-point calculation is executed by the processor, the terminal is controlled to implement the steps of the high-precision building environment prediction method based on point-by-point calculation as described in any one of claims 1-7.

Images