Forecasting system for wind speed / wind blast

Abstract

To enable the prediction of a spatial distribution of wind information in any block.SOLUTION: A forecasting system includes: a preprocessing unit that uses a numeric fluid analysis model to create computational data from wind information corresponding to at least one of wind speed and wind blast in a spatial distribution and creates a teacher data set that divides the calculated data into predetermined sections for input into a model using a neural network; a learning unit that uses the teacher data set and takes the divided unit as an image file unit for each input / output information, to generate a learned model with an image file showing an incoming wind condition and an image file showing building height information as inputs, and an image file of calculation results of the wind information as an output; and a prediction unit that inputs the incoming wind conditions in a block containing a planned building and the height information including the height of the planned building to the learned model, and predicts the wind information for the planned building based on the output of the learned model.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a forecasting system. [Background technology]

[0002] As an example of such a method, there is a technology related to a device capable of estimating wind speed distribution and wind pressure distribution (see Patent Document 1). With this technology, wind speed information corresponding to each of a plurality of mutually orthogonal axial directions (X, Y, Z) is stored for the wind speed distribution. Also, voxel data having a three-dimensional structure is used for the building information data, wind direction data, and wind speed distribution data.

[0003] There is also a technology for estimating wind speed distribution around a building (see Patent Document 2). In this technology, a height image, a shape image representing the cross-sectional shape information of the building, and a wind direction image representing wind direction information are used as inputs to a neural network to estimate the wind speed distribution around the building. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2023-009313 A [Patent Document 2] JP 2018-004568 A Summary of the Invention [Problem to be solved by the invention]

[0005] In the technology of Patent Document 1, the input is treated as numerical data, and as parameters resolved in the XYZ directions. However, when the data is treated as numerical data and voxel data, the amount of data increases, and there is a problem that learning and estimation are costly. In addition, while the technology of Patent Document 2 makes it possible to estimate wind information around a building, the estimation is limited to the building, and it cannot be applied to estimating wind information over a wide range, such as the terrain around the building or the entire city block.

[0006] In consideration of the above, an object of the present invention is to make it possible to predict the spatial distribution of wind information in any city block. [Means for solving the problem]

[0007] In order to achieve the above object, the prediction system of the present invention includes a pre-processing unit that uses a computational fluid analysis model to create calculation data including inflow wind conditions, height information including building height and height due to topographical relief, and the calculation results of the wind information from wind information corresponding to at least one of wind speed and wind pressure in spatial distribution, and creates a teacher dataset by dividing the calculation data into predetermined sections to input to a model using a neural network, a learning unit that uses the teacher dataset to generate a trained model in which the divided units are used as image file units for each of the input and output information, and which takes as input an image file showing the inflow wind conditions and an image file showing the height information of the building, and outputs an image file of the calculation result of the wind information, and a prediction unit that inputs the inflow wind conditions in a block including a planned building and the height information including the height of the planned building into the trained model, and predicts the wind information for the planned building using the output of the trained model, thereby making it possible to predict the spatial distribution of wind information in any block. Effect of the Invention

[0008] According to the present invention, it is possible to obtain an effect of making it possible to predict the spatial distribution of wind information in any city block. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a block diagram showing the configuration of a prediction system. [Diagram 2] FIG. 2 is a diagram illustrating an example of division of calculation data. [Diagram 3] FIG. 3 is a diagram showing an example of an image file indicating inflow wind conditions and an image file indicating height information. [Figure 4] FIG. 4 is a schematic diagram showing the model learning flow. [Diagram 5] FIG. 5 is a flowchart showing the learning process in the prediction system. [Figure 6] FIG. 6 is a flowchart showing the prediction process in the prediction system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] [Embodiments of the present invention] Hereinafter, a prediction system according to an embodiment of the present invention will be described with reference to the drawings.

[0011] FIG. 1 is a block diagram showing the configuration of a prediction system 100. As shown in FIG. 1, the prediction system 100 includes a model storage unit 102, a preprocessing unit 110, a learning unit 112, and a prediction unit 114. The prediction system 100 is configured as an apparatus, and is realized by a computer including a CPU (Central Processing Unit), a ROM (Read Only Memory) storing programs for implementing each processing routine, a RAM (Random Access Memory) for temporarily storing data, a memory as a storage means, and a network interface. Note that a GPGPU or an accelerator may be used instead of a CPU depending on suitability for each process, and an arithmetic unit may be used appropriately depending on the process. In particular, it is preferable to use a GPGPU or an accelerator for processes related to learning and inference. In addition, the prediction system may be configured by separating the learning unit 112 and the prediction unit 114 into a learning device and a prediction device, respectively.

[0012] The model storage unit 102 stores a trained model trained by the training unit 112. The trained model receives an image file indicating the inflow wind conditions and an image file indicating height information as input, and outputs an image file of the calculation result of the wind information. Details of the various image files will be described later.

[0013] The preprocessing unit 110 (1) first creates calculation data using a computational fluid analysis model, and (2) secondly divides the calculation data to create a teacher dataset. For example, CFD (Computational Fluid Dynamics) is used as the computational fluid analysis model. FIG. 2 is a diagram showing an example of division of the calculation data. (a1) in FIG. 2 is image data depicting a region of spatial distribution, and wind speed distribution is shown by a scalar value of each pixel. The region includes a city block, and wind speed distribution in the city block from above is depicted. The range of (a1) is approximately several kilometers square. (a2) is image data in which the calculation data is divided into several city blocks of about several hundred meters. The division unit is an example. Note that since the purpose is to create a teacher dataset for model learning for the space to be predicted, calculation data can be created for the space to be predicted.

[0014] The creation of the calculation data will be described. The pre-processing unit 110 creates the calculation data by performing a large number of unsteady calculations of wind information such as wind speed and wind pressure in the spatial distribution using CFD. In the CFD calculation, the inflow boundary surface corresponding to each region of the spatial distribution is set at a grid interval, and the inflow boundary data is input. The inflow boundary data is a value that reproduces the natural wind disturbance due to the unevenness near the ground surface based on the map data, and is a value that changes over time. Then, the spatial distribution is analyzed at a predetermined grid interval to obtain the calculation result of the wind information. The calculation data includes the inflow wind conditions and height information used in the CFD calculation, and the calculation result of the wind information. The height information is information including the height of the building in the spatial distribution and the height due to the undulations of the terrain. Here, the information of the existing map data, etc. may be referred to for the height of the building during learning. Note that the planned building to be predicted, which will be described later, is unknown data during learning and is not included.

[0015] In addition, multiple calculations are performed so that wind direction patterns corresponding to both wind speed and wind pressure are included. When the purpose is to predict wind speed on the ground surface, calculations are performed so that at least 16 wind direction patterns are included. When the purpose is to predict wind pressure on the surface of a building, calculations are performed so that at least 72 wind direction patterns are included. Of the 72 wind pressure directions, those that are common to some of the 16 wind speed direction patterns may be omitted. The pre-processing unit 110 incorporates the results of calculations performed corresponding to each of these wind directions into the teacher dataset.

[0016] The calculation results can also include various wind information that can be calculated using CFD. For example, the calculation results can include the distribution of wind pressure and wind load values ​​on the exterior surfaces of buildings within the block area. The calculation results can also include the distribution of wind speed values ​​and wind pressure distribution near the ground surface of the block as spatial distribution output, and the wind pressure distribution can include time history or average values.

[0017] The creation of a teacher data set by dividing the calculation data will be described. The preprocessing unit 110 creates a teacher data set by dividing the calculation data into predetermined sections in order to input the calculation data into a neural network model (NN model). An example of the section of the division unit is a block of several hundred meters square. Here, the image file showing the inflow wind conditions and the image file showing the height information are also divided into similar division units. FIG. 3 is a diagram showing an example of an image file showing the inflow wind conditions and an image file showing the height information. (b1) is an image file showing the inflow wind conditions, and (b2) is an image file showing the height information. Here, the image file showing the inflow wind conditions does not show the wind direction, and the wind speed value at each pixel of the division unit is shown as a scalar, which is one of the features of the method of this embodiment. In the example of (b1), the wind speed value in the Y direction on the left side shows the wind speed value of the entire area of ​​the division unit including the X direction and the Y direction. Note that the left side may correspond to any wind direction. Also, the arrangement on the left side is an example, and the right side, upper side, and lower side may be set arbitrarily. In other words, rather than preparing input conditions for each wind direction for a target block for which learning or prediction is to be performed, the wind direction is fixed (only one wind direction input condition) and the target block is rotated as appropriate to perform learning and prediction under multiple wind direction conditions. The target block can be rotated by rotating the height information divided for each section in learning and prediction. The rotation pattern is the wind direction pattern described above. In this way, for a predetermined wind direction pattern corresponding to each of the wind speed and wind pressure, the pre-processing unit 110 rotates the height information including the height of the planned building according to the wind direction of the wind direction pattern, calculates wind information for the building and wind information for the block, and obtains the calculation results. The scalar representation indicating the wind speed value and height information is an example of color information in this disclosure.

[0018] The learning unit 112 uses the teacher dataset to perform machine learning of a model using a neural network, and generates a trained model. In model learning, the divided units are used as units of image files for each piece of input and output information. Of the teacher dataset, an image file showing inflow wind conditions and an image file showing building height information are used as input to the neural network, and an image file of the calculation results of wind information is used as output from the neural network. The generated trained model is stored in the model storage unit 102, and the prediction unit 114 reads it out as appropriate and executes prediction processing.

[0019] Figure 4 is a schematic diagram showing the flow of model learning. The correct value of (b3) is the calculation result of wind information. For inputs of (b1) and (b2), the neural network error calculation is performed so that the model output approaches the correct value of (b3), and the model parameters are updated to learn the model. The update is repeated until the conditions (number of times, threshold value, etc.) are satisfied, and a learned model is generated.

[0020] The prediction unit 114 inputs the inflow wind conditions in the block including the planned building and height information including the height of the planned building into the trained model, and predicts wind information for the planned building by outputting the trained model. Here, the prediction unit 114 predicts wind information for the building and wind information for the block as wind information for the planned building. An example of wind information for a building is the distribution of wind pressure values ​​and wind load values ​​on the outdoor surface of the building within the area of ​​the block. An example of wind information for a block is the distribution of wind speed values ​​and wind pressure distribution near the ground surface of the block. In addition, in the prediction, for a predetermined wind direction pattern corresponding to each of the wind speed and wind pressure, height information including the height of the planned building is rotated according to the wind direction of the wind direction pattern, and inputted into the trained model, and wind information is predicted for each wind direction pattern.

[0021] Next, the operation of the embodiment of the present invention will be described. Fig. 5 is a flowchart showing the learning process in the prediction system 100. Fig. 6 is a flowchart showing the prediction process in the prediction system 100. The arithmetic unit of the CPU, GPGPU, or accelerator of the prediction system 100 reads out and executes programs and various data from the ROM, and the arithmetic unit executes the processing as each part of the prediction system 100. The learning process and the prediction process may be executed as one unit.

[0022] The learning process will now be described. In step S100, the pre-processing unit 110 creates calculation data by performing a large number of unsteady calculations of wind information, such as wind speed and wind pressure in a spatial distribution, using CFD.

[0023] In step S102, the preprocessing unit 110 creates a teacher data set by dividing the calculation data into predetermined partitions in order to input the data to a neural network model (NN model).

[0024] In step S104, the learning unit 112 uses the teacher data set to perform machine learning of a model using a neural network to generate a trained model. The trained model is stored in the model storage unit 102.

[0025] The prediction process will now be described. In step S110, the prediction unit 114 acquires the inflow wind conditions in the city block including the planned building and height information including the height of the planned building.

[0026] In step S112, the prediction unit 114 inputs the inflow wind conditions in the block including the acquired planned building and height information including the height of the planned building into the trained model, and predicts wind information regarding the planned building based on the output of the trained model.

[0027] In step S114, the prediction unit 114 outputs the prediction result of the wind information regarding the planned building to a predetermined display interface or the like.

[0028] As described above, the prediction system 100 according to the embodiment of the present invention makes it possible to predict the spatial distribution of wind information in any city block.

[0029] The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the spirit and scope of the present invention.

[0030] The method of this embodiment has various advantages, as listed below. For example, the method of this embodiment does not use a physical model for prediction, so that the calculation cost and calculation time are significantly smaller. In addition, by learning a model using the results of wind information calculated using CFD, it is possible to predict wind speed and wind pressure for various surrounding block shapes and building shapes. In addition, by using the results of unsteady calculations as learning data, it is possible to predict the peak values ​​of wind speed and wind pressure, and the range of applications is wide. In addition, it is possible to pinpoint forecasts of planned buildings (each work site at a construction site), and if strong winds are predicted, measures can be taken in advance, which leads to the reduction of disasters. In addition, by utilizing the predicted wind pressure in wind-resistant design, rational design is possible, which leads to cost reduction. In addition, since wind speed and wind pressure can be predicted in a short time, it is possible to quickly respond to changes in the shape of a building. In addition, it is possible to remove time and budgetary constraints, and to consider a large number of designs at the basic design stage, maximizing the expressive power of architectural design. [Explanation of symbols]

[0031] 100 Prediction System 102 Model memory section 110 Pretreatment section 112 Learning Department 114 Prediction Department

Claims

1. Using a computational fluid analysis model, calculation data is created from wind information corresponding to at least one of wind speed and wind pressure in a spatial distribution, the calculation data including inflow wind conditions, height information including building height and height due to topographical relief, and the calculation results of the wind information; a pre-processing unit that divides the calculation data into predetermined blocks to create a teacher data set for input to a model using a neural network; a learning unit that uses the teacher dataset to generate a trained model in which the divided units are units of image files of each piece of input and output information, the image file indicating the inflow wind conditions and the image file indicating the building height information are input, and the image file of the calculation result of the wind information is output; a prediction unit that inputs the inflow wind conditions in a block including a planned building and the height information including the height of the planned building into the trained model and predicts the wind information regarding the planned building based on the output of the trained model; A prediction system including:

2. Regarding the wind information, the wind information for a building is defined as the distribution of wind pressure and wind load values ​​on the outdoor surface, and the wind information for a block is defined as the distribution of wind speed values ​​and wind pressure distribution near the ground surface of the block, The prediction system according to claim 1 , wherein the prediction unit predicts wind information related to the building and wind information related to the block as the wind information related to the planned building.

3. The prediction system of claim 2, wherein the prediction unit rotates the height information including the height of the planned building according to the wind direction of a predetermined wind direction pattern corresponding to each of wind speed and wind pressure, inputs the rotated height information into the trained model, and predicts the wind information for each wind direction pattern.

4. the image file showing the inflow wind conditions shows a wind speed value using color information; The prediction system according to claim 1 , wherein the image file indicating the height information indicates heights of buildings and terrain using color information.

Images