JP7374010B2 - Wind speed distribution estimation device and wind speed distribution estimation method - Google Patents

Description

The present invention relates to a wind speed distribution estimation device and a wind speed distribution estimation method.

BACKGROUND OF THE INVENTION Conventionally, wind environments such as building wind around high-rise buildings, for example, have been evaluated. The wind environment may be evaluated as a wind speed distribution, for example, by wind tunnel experiments or computational fluid analysis.
In wind tunnel experiments, the wind environment and wind speed distribution around the building are evaluated by applying wind to a model that reproduces the building to be evaluated and measuring the wind speed etc. around the model using sensors installed in the model. Ru. Patent Document 1 discloses a wind speed measuring device used during wind tunnel experiments.
In computational fluid analysis, a building to be evaluated is modeled on a computer, and an approximate solution for wind speed is calculated by applying discretized and modeled equations related to fluid motion. Patent Document 2 discloses a wind environment prediction method based on the results of fluid analysis.

Generally, the wind speed distribution formed around a building is affected not only by the shape of the building and the wind direction, but also by the group of buildings around the building. Therefore, the wider the area for evaluating the wind speed distribution, the more accurately the wind speed distribution can be evaluated.
However, in wind tunnel experiments, when the area to be evaluated becomes vast, a correspondingly large wind tunnel device and many building models are required.
In addition, in computational fluid analysis, analysis targeting an urban area generally involves large-scale three-dimensional calculations, which requires a considerable amount of calculation time. Therefore, when the area to be evaluated becomes large, a large amount of calculation time is required.
As described above, in wind tunnel experiments and computational fluid analysis, it is difficult to easily and accurately evaluate the wind speed distribution over a wide area in terms of cost, calculation time, and the like.

In addition to the above-mentioned wind tunnel experiments and computational fluid analysis, particularly in recent years, attempts have been made to evaluate wind speed distribution using neural networks.
For example, Patent Document 3 discloses a wind speed distribution estimation device around a building that involves evaluation of wind speed distribution using a neural network.
The wind speed distribution estimation device of Patent Document 3 includes an input means for inputting a height image representing height information to be estimated, a shape image representing cross-sectional shape information of a building, and a wind direction image representing wind direction information, and wind speed distribution image creation means for creating a wind speed distribution image that is an image representing the wind speed distribution. The wind speed distribution image creation means is composed of a convolutional neural network that uses a height image, a shape image, and a wind direction image as input images, and uses a wind speed distribution image as an output image.
When a neural network is used, an evaluation device can be realized using a general-purpose computer, and large-scale equipment such as wind tunnel experiments is not required. In addition, calculation time is generally faster than computational fluid analysis. Therefore, compared to wind tunnel experiments and computational fluid analysis, it may be possible to evaluate the wind speed distribution over a wide area more easily in terms of cost and calculation time.

However, when trying to evaluate the wind speed distribution in a wide area using a neural network, the learning data for training the machine learning device, that is, the experimental data that measured the wind speed distribution in a wide area (for example, measurement data, computational fluid It is difficult to prepare sufficient detailed analysis results. Even if the amount of experimental data is small, if the area to be evaluated is small, it is possible to increase the training data by, for example, cutting out the experimental data at an arbitrary location to match the size of the area. can. However, when the evaluation target area becomes wider, the number of training data generated by cutting out also decreases.
Therefore, when using a neural network, when evaluating wind speed distribution over a wide area, there is a possibility that the machine learning machine will not be able to learn sufficiently because there is not enough training data, and the evaluation accuracy will decrease. There is.

Theoretically, it is also possible to evaluate the wind speed distribution in a wide area by dividing a wide evaluation target area into multiple narrow individual areas, evaluating the wind speed distribution for each area, and combining the evaluation results. It is possible.
In this case, the wind speed distribution in each individual region is evaluated independently. If the evaluation target area is simply divided by the boundary line, each of the two points across the boundary line belongs to a different evaluation target area, so the difference in the evaluation value of the wind speed distribution between the points across the boundary line may become large. There is sex. Therefore, when combining the evaluation results, the wind speed distribution is not continuous between individual areas at the boundary line.
In particular, when evaluating the wind speed distribution of each individual area, information regarding buildings outside the boundary line, ie, the outer edge of the individual area, is not given at the time of evaluation and is therefore not taken into account. Therefore, the impact of these buildings is not reflected in the evaluation. For this reason, the accuracy of estimating the wind speed distribution tends to decrease near the outer edge of the individual region.
For these reasons, when the evaluation target area is simply divided by boundary lines, the evaluation accuracy in the vicinity of the boundary lines decreases.

JP2014-48120A Japanese Patent Application Publication No. 2018-165884 JP 2018-4568 Publication

The problem to be solved by the present invention is to provide a wind speed distribution estimation device and a wind speed distribution estimation method that can easily and accurately estimate the wind speed distribution in a wide area.

In order to solve the above problems, the present invention employs the following means. That is, the present invention is a wind speed distribution estimating device for estimating the wind speed distribution around a building, which divides the entire area to be estimated for the wind speed distribution into a plurality of individual areas so as to have overlapping parts that overlap with each other. a wind speed distribution estimator that estimates the wind speed distribution in the plurality of individual areas; and a wind speed distribution estimation unit that estimates the wind speed distribution in the plurality of individual areas; a weight setting unit in which weights are individually set; and an estimation method that calculates an estimation result in which the wind speed of the overlapped point is adjusted based on the estimation result of the weight and the wind speed distribution of the overlapped point. a result adjustment unit; and an estimation result that combines the estimation results of the wind speed distribution in the plurality of individual regions based on the adjusted estimation result to generate an estimation result of the wind speed distribution in the entire region. A wind speed distribution estimation device is provided, comprising: a coupling portion;
The present invention also provides a wind speed distribution estimation method for estimating a wind speed distribution around a building, in which the entire area to be estimated for the wind speed distribution is divided into a plurality of individual areas so as to have overlapping parts that overlap with each other. The wind speed distribution in a plurality of individual regions is estimated, and for each of the individual regions, individually set weights and the wind speed are calculated for each overlapping point that is a point included in the overlapping part. Based on the distribution estimation results, calculate an estimation result in which the wind speed at the overlapped point is adjusted, and based on the adjusted estimation result, combine the estimation results of the wind speed distribution in the plurality of individual regions. The present invention provides a wind speed distribution estimation method that generates an estimation result of the wind speed distribution in the entire area.
According to the above configuration, the entire area for which the wind speed distribution is estimated is divided into multiple individual areas, the wind speed distribution is estimated for each of these multiple individual areas, and the multiple estimation results are combined. Then, an estimation result of the wind speed distribution in the entire area is generated.
That is, since a wide area is divided into small individual areas and the wind speed distribution is estimated for each of these areas, it is possible to estimate the wind speed distribution over a wide area relatively easily.
Here, the entire area is divided into a plurality of individual areas so as to have overlapping parts that overlap with each other. That is, when dividing the entire area into individual areas, the area is not simply divided by one boundary line, but is divided so that adjacent individual areas have a common overlapping portion. In this state, when combining the estimation results of the individual regions as the estimation results of the entire region, the estimation results of each individual region are adjusted based on the weights set appropriately for each overlapping point for each individual region. be done. Therefore, discontinuity between individual areas is alleviated.
Furthermore, the overlapping portion is included in a plurality of adjacent individual regions from different directions. Therefore, even if the estimation accuracy of overlapping parts located near the outer edge of each individual area is reduced due to lack of information on buildings located outside the outer edge, these buildings are adjacent to each other across the overlapping part. The missing information is reflected in the estimation of the wind speed distribution in this adjacent individual area. Therefore, when combining the estimation results of the individual areas as the estimation results of the entire area, information regarding all the buildings located around the overlapping area is reflected and adjusted in the estimation result of the overlapping area. This makes it possible to suppress a decrease in estimation accuracy near the outer edge of the individual area, compared to a case where the evaluation target area is simply divided by a boundary line without providing an overlapping part.
Due to the above reasons, the accuracy of estimating the wind speed distribution can be improved over almost the entire wide area.

In one aspect of the present invention, the estimation result adjustment unit calculates, for each of the overlapping points, the sum of the weights set to the overlapping points, and calculates the sum of the weights set to the overlapping points, and Then, by dividing the weight by the sum of the weights and then multiplying by the estimation result at the overlap point, and calculating the sum of the multiplication values, the adjusted estimation result at the overlap point is calculated. Calculate.
According to the above configuration, it is possible to appropriately realize the above wind speed distribution estimation device.

According to the present invention, it is possible to provide a wind speed distribution estimation device and a wind speed distribution estimation method that can easily and accurately estimate the wind speed distribution in a wide area.

FIG. 1 is a block diagram of a wind speed distribution estimation device in an embodiment of the present invention. It is a typical explanatory diagram of the machine learning machine in the above-mentioned wind speed distribution estimation device. FIG. 2 is an explanatory diagram of building information data that is input to the machine learning device. FIG. 3 is an explanatory diagram of the entire area dividing section in the wind speed distribution estimation device. FIG. 2 is a schematic explanatory diagram of a learning model that has been trained by the machine learning device. FIG. 3 is an explanatory diagram of weights set in a weight setting section in the wind speed distribution estimating device. It is a flowchart of the wind speed distribution estimation method using the above-mentioned wind speed distribution estimation device. It is a wind speed estimation result by the said wind speed distribution estimation device.

The present invention is an estimation device for an urban wind environment including wind speed information and wind direction information, and an estimation method thereof, which targets an urban area including a plurality of building groups. Specifically, detailed wind speed distribution including wind speed information and wind direction information is obtained for each divided area that forms the urban area, and by partially overlapping each divided area, the detailed wind speed distribution in the urban area is estimated.
Embodiments of the present invention will be described in detail below with reference to the drawings.
FIG. 1 is a block diagram of a wind speed distribution estimation device in this embodiment.
As described above, the wind speed distribution estimation device 1 in this embodiment is a device that estimates the wind speed distribution around a building in an arbitrary building. The wind speed distribution estimation device 1 is, for example, an information processing device such as a personal computer. The wind speed distribution estimation device 1 includes a learning section 20, a wind speed distribution estimation section 21, and a learning model parameter storage section 22. The wind speed distribution estimating device 1 further includes a whole area dividing section 50, a weight setting section 51, an estimation result adjusting section 52, and an estimation result combining section 53.
Among these components of the wind speed distribution estimating device 1, the learning section 20, the wind speed distribution estimating section 21, the whole area dividing section 50, the estimation result adjusting section 52, and the estimation result combining section 53 are, for example, a CPU in the information processing device. It may be software or a program executed by. Further, the learning model parameter storage unit 22 and the weight setting unit 51 may be realized by a storage device such as a semiconductor memory or a magnetic disk provided inside or outside the information processing device. The learning unit 20 is processed by, for example, a GPU (Graphics Processing Unit).

As will be described later, when the wind speed distribution estimating unit 21 receives building information data 31 including building shape information and wind direction data 32, it estimates a wind speed distribution corresponding to the building information data 31 and wind direction data 32. In order to effectively perform this estimation, the wind speed distribution estimating unit 21 is equipped with a learning model 25 generated by machine learning using a machine learning device 24 provided in the learning unit 20, especially in this embodiment. There is. More specifically, the learning unit 20 inputs the learning data 2 into the machine learning device 24 to perform machine learning, and generates learning model parameters related to estimation of wind speed distribution.
That is, the wind speed distribution estimating device 1 performs two types of operations: learning the wind speed distribution and estimating the wind speed distribution. In order to simplify the explanation, each component of the wind speed distribution estimating device 1 will be described below at the time of learning the wind speed distribution, and then the behavior of each component at the time of estimating the wind speed distribution will be described.
Note that the wind speed distribution estimating device 1 estimates the wind speed distribution in a wide area, for example, one side is 250 m or more, as will be explained later. In order to accurately estimate the wind speed distribution in such a wide area, the wind speed distribution estimation device 1 divides the wide area (the entire area) into a plurality of smaller individual areas when estimating the wind speed distribution, and in each individual area, Estimate the wind speed distribution and combine these estimation results. That is, when estimating the wind speed distribution, a wide overall area is input to the wind speed distribution estimating device 1, but divided individual areas are input to the learning model 25 that actually estimates the wind speed distribution. Therefore, information corresponding to, for example, an area smaller than 250 m on a side, which will be explained later as an individual area, is input to the machine learning device 24 as learning data 2.

When learning the wind speed distribution, as described above, the learning unit 20 causes the machine learning device 24 to perform machine learning based on the learning data 2, which is information corresponding to an area having a side smaller than 250 m, for example. The machine learning device 24 performs deep learning to generate a learning model 25. FIG. 2 is a schematic explanatory diagram of the machine learning device 24. The learning data 2 includes building information data 3, wind direction data 4, and wind speed distribution data 5 as teacher data.
The building information data 3 is data having information regarding the external shape of a building, and includes a building shape image 6 and a building height image 7. FIGS. 3A and 3B are explanatory diagrams of a building shape image 6 and a building height image 7, respectively.
The building shape image 6 is an image expressing the shape of a building in a bird's-eye view of a given area. More specifically, in the building shape image 6, pixels in a building portion 6a that corresponds to a building and pixels in a non-building portion 6b that does not correspond to a building have different pixel values. is set to . In this embodiment, the building shape image 6 is a one-channel image, and the building portion 6a and the non-building portion 6b are respectively black and a color corresponding to the maximum pixel value such as white, for example, a pixel value (luminance When the value) is expressed in 8 bits, it is expressed as 0 and 255.
Hereinafter, in each image in this embodiment including this building shape image 6, the axial direction extending horizontally is the X direction, and the axial direction extending vertically perpendicular to the X direction in the horizontal plane is the Y direction. The axial direction perpendicular to the horizontal plane is called the Z direction.

The building height image 7 is an image representing the height of a building in a bird's-eye view of the same area as that represented in the building shape image 6. More specifically, in the building height image 7, when a non-building portion 7b, which is a portion not corresponding to a building such as a road or the ground surface, is set as a reference plane with a height of 0 m, a building portion 7a corresponding to a building is set as a reference plane with a height of 0 m. A pixel within is set to have a pixel value corresponding to the height of the building in which the pixel is included. In this embodiment, the building height image 7 is a one-channel image, and the non-building portion 7b is expressed in black, for example, 0. Furthermore, the building portion 7a is expressed in a neutral color, for example, with a pixel value of 1 or more. In particular, in this embodiment, the pixel value of the pixel within the building portion 7a is set to be the same value as the height of the building in which the pixel is included. For example, when a pixel value is expressed in 8 bits, a pixel in the building part 7a corresponding to a building with a height of 10 m has a pixel value of 10, and a pixel in the building part 7a corresponding to a building with a height of 35 m has a pixel value of 10. The pixel value is 35. The pixels in the building portion 7a corresponding to a building with a height of 255 m have the maximum pixel value, that is, 255.
Since the building height image 7 is an image representing the same area as that expressed in the building shape image 6 as described above, it has the same resolution as the building shape image 6.
As a target for estimating the wind speed distribution in the wind speed distribution estimation device 1, for example, when pixel values are expressed in 8 bits and it is assumed that there are many high-rise buildings exceeding 255 m, for example, from 0 to the tallest building. The height value range may be converted into a pixel value range, that is, a range from 0 to 255, and this may be used as the pixel value of the building height image 7.

The wind direction data 4 is data having information regarding the wind direction. More specifically, the wind direction data 4 is data representing the direction of wind blowing from outside the area into a building or a group of buildings in the area expressed in the building information data 3 described above. The wind direction is uniquely determined at all points within the region, regardless of the location within the region in the building information data 3.
The wind direction data 4 expresses the wind direction as described above, and since the value is unique at all points, it can be regarded as one vector. The wind direction data 4 consists of a wind direction A wind direction Y component image 9 is provided in which component values are expressed as a one-channel image. For example, the wind direction X component image 8 can be an image in which the pixel values of all pixels are set to the pixel values corresponding to the component values in the X direction obtained by decomposing the vector into components. Further, the wind direction Y component image 9 can be an image in which the pixel values of all pixels are set to the pixel values corresponding to the component values in the Y direction obtained by decomposing the vector into components.

For example, in a two-dimensional coordinate system, if we consider the wind direction as the direction in which the wind blows downstream when the origin is upstream, that is, the vector, then the component values in each axis direction obtained by decomposing this into components. can take a negative value depending on the direction of the vector. Therefore, when expressing the wind direction in the wind direction For example, if a pixel value is expressed in 8 bits, it is necessary to convert it into a pixel value range of 0 to 255. In this embodiment, when the component value is 0, the pixel value is set to the center value of the range of possible values of the pixel value, for example 128, and when the component value is a positive value, the pixel value is set within the range from 129 to 255. In the case of a negative value, the pixel value is set to a value within the range from 0 to 127.
For example, in the coordinate system formed by the X direction and Y direction in the building information data 3, if the wind is from the west, if the origin is set upstream, the wind direction will be in the positive direction on the X axis. The component value on the top becomes a positive value, and the component value on the Y direction becomes 0. Therefore, the wind direction X component image 8 can be, for example, a solid white image in which the pixel values of all pixels are set to the maximum pixel value, for example 255. Further, the wind direction Y component image 9 can be a specific neutral color image in which the pixel values of all pixels are set to the center value of the range of possible pixel values, for example, 128. Conversely, in the case of wind from the east, if the origin is upstream, the wind direction will be in the negative direction on the X axis, so the component value in the X direction will be a negative value, and the component value in the Y direction will be becomes 0. Therefore, the wind direction The image may be a specific neutral color image.

The wind speed distribution data 5 is information regarding the distribution of wind speed in the area expressed in the building information data 3, that is, data regarding the wind speed at each point around the building. The wind speed distribution data 5 is data obtained corresponding to the building information data 3 and the wind direction data 4, and includes information regarding the shape and height of the building and information regarding the wind blowing into the area expressed in the building information data 3. When given, it is uniquely determined based on these.
Particularly in this embodiment, the wind speed distribution data 5 also includes information regarding the wind direction at each point. For this purpose, the wind speed distribution data 5 includes a wind speed distribution X component image 10, a wind speed distribution Y component image 11, and a wind speed distribution A distribution Z component image 12 is provided. The wind speed distribution This is an image set to a value corresponding to the component value in the X direction when the components are decomposed into each of the Y direction and the Z direction. Similarly, the wind speed distribution Y component image 11 and the wind speed distribution Z component image 12 are images in which the pixel value of each pixel is set to a value corresponding to the component value in each of the Y direction and the Z direction. When the point corresponding to each pixel is a building, the pixel value of the pixel is set to a value corresponding to no wind.

As mentioned above, since the wind speed distribution data 5 includes information regarding the wind direction, similarly to the case of the wind direction data 4, the component values of the wind speed in each axis direction are not necessarily positive values, and depending on the wind direction, If the wind blows in the direction opposite to the direction in which the value of , the component value will be a negative value. In order to express such positive and negative values in the image, in this embodiment, the maximum positive value of the maximum wind speed that can be assumed is set to the maximum value of the pixel value, for example, 255 when the pixel value is expressed in 8 bits. , the case of no wind, that is, wind speed of zero, is made to correspond to the center value of the range of values that the pixel value can take, for example, 128, and the negative value of the maximum wind speed is made to correspond to a value that is 1 larger than the minimum value of the pixel value, for example, 1. We are making it compatible. The pixels corresponding to the building portion 6a of the building shape image 6 are set to the minimum pixel value, for example 0. Then, the wind speed within the range from the negative value of the maximum wind speed to the positive value of the maximum wind speed is associated with the value from the minimum value of the range of possible values of the pixel value by 1 to the maximum value, and this correspondence Based on the relationship, the pixel value of each pixel is set to a value corresponding to the wind speed that should be held as information at that pixel.
In this embodiment, each image of the learning data 2 is actually input from 0 to 1 by dividing the value of each pixel by the maximum value of the pixel value, for example, 255. It is used scaled to a range of values.

When the machine learning device 24 performs machine learning, the building information data 3 and the wind direction data 4 are used as input to the machine learning device 24. The learning estimation result 15, which is the output of the machine learning device 24 at this time, is compared with the wind speed distribution data 5 corresponding to the building information data 3 and the wind direction data 4, and the machine learning device 24 performs machine learning based on this comparison result. be done. In this way, the wind speed distribution data 5 is used as the correct value in machine learning, that is, as teacher data, and the machine learning device 24 is trained to output the learning estimation result 15 that is close to the wind speed distribution data 5.
As already explained, each of the building information data 3, the wind direction data 4, and the wind speed distribution data 5 as teacher data is an image. Therefore, the learning estimation result 15 to be compared with the wind speed distribution data 5 is also an image. Therefore, in this embodiment, the machine learning device 24 is realized by a fully convolutional network (hereinafter referred to as FCN) that is compatible with processing when images are used as input and output. As explained below, the FCN does not include a fully connected layer, and the feature map processed and generated in the convolutional layer is directly input to the transposed convolutional layer.

As shown in FIG. 2, the machine learning device 24 includes a convolution processing section 27 and a transposed convolution processing section 28. The convolution processing unit 27 includes a plurality of convolution layers 27a, 27b, and 27c connected in series. The transposed convolution processing section 28 includes a plurality of transposed convolution layers 28c, 28b, and 28a, which are similarly connected in series. In the schematic diagram of FIG. 2, the number of convolutional layers and transposed convolutional layers are each three, but the number of these convolutional layers and transposed convolutional layers is not limited to three.
The learning unit 20 inputs the building information data 3 and the wind direction data 4 to the first stage convolution layer 27a. As already explained, in this embodiment, each of the building shape image 6 and building height image 7 of the building information data 3 and the wind direction X component image 8 and the wind direction Y component image 9 of the wind direction data 4 is one channel. It is realized as an image. Therefore, the building shape image 6, the building height image 7, the wind direction The data is input to the machine learning device 24 as data 14.

The convolution layer 27a includes a predetermined number of filters. The machine learning device 24 positions each filter on the learning input data 14, and sets the pixel value of each pixel of the learning input data 14 in the filter corresponding to the pixel position in the filter. Convolution filtering is performed by calculating the sum with the given weights. Thereby, the pixel value of one pixel in the convolutional layer 27a is calculated. The machine learning device 24 calculates a plurality of pixel values by executing such convolution filter processing while moving the filter on the learning input data 14 at predetermined resolution increments, and arranges the pixel values to correspond to the filter. generate a single image.
The machine learning device 24 executes this process for all filters and generates feature maps according to the number of filters.
Batch normalization processing, pooling processing, and activation functions are performed on the feature map as necessary.
The feature map generated in the convolutional layer 27a becomes an input image for the next convolutional layer 27b.

In the convolution layer 27b, convolution filter processing is performed on the feature map generated in the convolution layer 27a, similarly to the convolution layer 27a. The convolution layer 27b includes a predetermined number of filters, and performs convolution filter processing using these filters, and further performs batch normalization processing and pooling processing as necessary. Generate a predetermined number of feature maps.
In the convolution layer 27c, convolution filter processing is performed on the feature map generated in the convolution layer 27b. The convolution layer 27c is equipped with a predetermined number of filters, and performs convolution filter processing using these filters, and further performs batch normalization processing and pooling processing as necessary. Generate a predetermined number of feature maps.
The weights of the filters in each convolutional layer 27a, 27b, 27c are adjusted by machine learning.
The feature map generated in the convolution layer 27c becomes an input to the transposed convolution layer 28c of the transposed convolution processing unit 28.

The transposed convolution processing unit 28 has a structure symmetrical to that of the convolution processing unit 27. That is, although the learning input data 14 was compressed to a low dimension by the convolution processing section 27, the transposition and convolution processing section 28 operates to expand and restore the data compressed to a low dimension. More specifically, a transposed convolution layer 28c executes a transposed convolution process corresponding to the convolution layer 27c, a transposed convolution layer 28b executes a transposed convolution process corresponding to the convolution layer 27b, and a transposed convolution process corresponding to the convolution layer 27a. Output data is generated by sequentially passing through the transposed convolution layer 28a to be executed.
In this embodiment, the output data of the transposed convolution processing unit 28, that is, the machine learning device 24, is expressed in the building information data 3 corresponding to the building information data 3 and wind direction data 4 in the input learning input data 14. This is the learning estimation result 15 of the wind speed and direction in the area where the wind speed and direction were determined. Similar to the wind speed distribution data 5, the learning estimation result 15 is decomposed into components corresponding to each of the X direction, Y direction, and Z direction at each point within the area expressed in the building information data 3. These are three 1-channel images in which wind speed and wind direction information is stored, and these are output as one 3-channel image.

In the machine learning device 24, machine learning is performed so that the estimation result in the learning estimation result 15 becomes a value close to the wind speed distribution data 5 corresponding to the input learning input data 14, that is, the teacher data. For this purpose, the learning unit 20 compares the wind speed distribution data 5 corresponding to the learning input data 14 and the learning estimation result 15, which is the output when the learning input data 14 is input, pixel by pixel. For example, the squared error of the difference in pixel values between each pixel is calculated as a cost function.
Then, the machine learning device 24 performs machine learning by adjusting the weight values of each filter using an error back propagation method or the like so as to reduce this cost function.
As a result, when the learning input data 14 is input, the machine learning device 24 outputs a learning estimation result 15 that is close to the wind speed distribution data 5 corresponding to the building information data 3 and wind direction data 4 included therein. It is learned as follows.

When the learning is completed, the learning unit 20 stores parameters such as weight values of each filter whose adjustment has been completed in the learning model parameter storage unit 22 as learning model parameters. The learning model parameters stored in the learning model parameter storage section 22 are acquired by the wind speed distribution estimating section 21, which will be described later, and a learning model 25 for estimating the wind speed distribution is constructed.
That is, the learning unit 20 generates a trained learning model 25 in which appropriate learning parameters have been learned, which is used as a program module that is part of artificial intelligence software.

Next, the behavior of each component when estimating the wind speed distribution will be explained.
As already explained, when estimating the wind speed distribution, the wind speed distribution estimation device 1 divides a wide area (the entire area) into a plurality of individual areas, estimates the wind speed distribution in each individual area, and uses these estimation results. Join. Therefore, the estimation input data 30, which is input data to the wind speed distribution estimating device 1 at the time of estimating the wind speed distribution, is information corresponding to a wide entire area, for example, 250 m or more on a side. The input data 30 at the time of estimation is building information data 31 and wind direction data 32 in which the area and wind direction where the wind speed distribution is actually estimated are stored as information. The building information data 31 includes a building shape image 33 and a building height image 34, each of which was used as learning data 2 during learning, except that the image is larger because it corresponds to the entire area. It is configured similarly to the building shape image 6 and the building height image 7. The wind direction data 32 includes a wind direction X component image 35 and a wind direction Y component image 36, each of which was used as learning data 2 during learning, except that the image is larger because it corresponds to the entire area. It is configured similarly to the wind direction X component image 8 and the wind direction Y component image 9. The building shape image 33, the building height image 34, the wind direction This is image data that has been combined as a single 4-channel image.
In this embodiment, when each image of the estimation input data 30 is input to the wind speed distribution estimating unit 21, the value of each pixel is set to the maximum value of the pixel value, for example, 255. It is scaled to a range of values from 0 to 1 by dividing by .

The whole region dividing unit 50 divides the whole region, which is the target of wind speed distribution estimation, into a plurality of individual regions so as to have overlapping parts that overlap with each other. More precisely, the whole region dividing unit 50 divides the estimation input data 30, which is an image corresponding to the whole region, into individual input data, which is an image corresponding to individual regions, such that the individual regions overlap each other. do.
FIG. 4 is an explanatory diagram of the entire area dividing section 50. To simplify the explanation, the estimation input data 30 corresponding to the entire area GR has a resolution of 20 pixels in both the horizontal direction X and the vertical direction Y, and from this, the resolution in the horizontal direction A case will be described in which individual input data corresponding to the individual region LR of 8 pixels in both is cut out and divided. In reality, both the estimation input data 30 and the individual input data have a resolution greater than that described above.
In the following description, when the entire region GR is written, the estimation input data 30 corresponding thereto is also indicated, and when the individual region LR is written, the corresponding individual input data is also indicated. In addition, when the entire region GR or individual region LR is viewed as an image, the position of the pixel in these images is determined by the coordinate value when the upper left pixel of the image is set as the origin (0, 0), that is, (horizontal direction pixel coordinate value in the vertical direction Y).

First, as shown in FIG. 4(a), the whole region dividing unit 50 matches the pixel located at the origin (0, 0) of the upper left of the individual region LR with the origin (0, 0) of the whole region GR. The individual region LR1 obtained when the calculation is performed is cut out and divided from the entire region GR, that is, the input data 30 at the time of estimation, as individual input data.
Next, as shown in FIG. 4B, the entire area dividing unit 50 divides the individual area LR by a certain stride (amount of movement, 4 pixels in this embodiment) in the horizontal direction from the individual area LR1. is moved in the horizontal direction X. As a result, the origin of the individual region LR is located at (4,0) of the entire region GR, and the individual region LR2 in this state is cut out and divided from the entire region GR.
The stride size in the lateral direction X is set smaller than the resolution in the lateral direction X of the individual region LR. More specifically, it is desirable that the stride size in the lateral direction X be less than half the resolution in the lateral direction X of the individual region LR. Thereby, the individual region LR1 and the individual region LR2 are divided from the entire region GR so as to have an overlapping portion OR (OR1) that overlaps with each other.
Further, as shown in FIG. 4(c), the overall area dividing unit 50 moves the individual area LR in the lateral direction X by a stride in the lateral direction X from the individual area LR2, and divides the individual area LR3 in this state into the entire Cut out and divide from region GR. The individual region LR2 and the individual region LR3 are divided from the entire region GR so as to have an overlapping portion OR2 that overlaps with each other.
Similarly, as shown in FIG. 4(d), the entire area dividing unit 50 moves the individual area LR in the horizontal direction X by a stride in the horizontal direction X from the individual area LR3, and divides the individual area LR4 in this state. , the entire region GR is cut out and divided.

As shown in FIG. 4D, when the movement and division of the individual region LR at the top of the entire region GR is completed, the entire region dividing unit 50 moves the individual region LR in the vertical direction Y. That is, as shown in FIG. 4(e), the entire area dividing unit 50 divides the individual area LR1 shown in FIG. 4(a) by a certain stride in the vertical direction Y (4 pixels in this embodiment). , moves the individual region LR in the vertical direction Y. Thereby, the origin of the individual region LR is located at (0, 4) of the whole region GR, and the individual region LR5 in this state is cut out and divided from the whole region GR.
The size of the stride in the vertical direction Y is set smaller than the resolution of the individual region LR in the vertical direction Y. More specifically, it is desirable that the stride size in the vertical direction Y be less than half the resolution of the individual region LR in the vertical direction Y. Thereby, the individual region LR1 and the individual region LR5 are divided from the entire region GR so as to have an overlapping portion OR3 that overlaps with each other.
Next, the entire area dividing unit 50 moves the individual area LR in the lateral direction X by a stride in the lateral direction X from the individual area LR5, as shown in FIG. 4(f). As a result, the origin of the individual region LR is located at (4, 4) of the entire region GR, and the individual region LR6 in this state is cut out and divided from the entire region GR.
The individual region LR5 and the individual region LR6 are divided from the entire region GR so as to have an overlapping portion OR4 that overlaps with each other. Further, the vertically adjacent individual regions LR2 and LR6 are divided from the entire region GR so as to have an overlapping portion OR5 that overlaps with each other.
Similarly, as shown in FIG. 4(g), the entire area dividing unit 50 moves the individual area LR in the horizontal direction X by a stride in the horizontal direction X from the individual area LR6, and divides the individual area LR7 in this state. , the entire region GR is cut out and divided.

In this way, by cutting out the individual regions LR while moving them in the horizontal direction Split the input data. In the example of FIG. 4, the entire area GR is divided into 16 individual areas LR. FIG. 4(h) shows the individual region LR16 that is divided last.

Next, the wind speed distribution estimation unit 21 acquires learning model parameters from the learning model parameter storage unit 22 and constructs a learning model 25. FIG. 5 is a schematic explanatory diagram of the learning model 25. The wind speed distribution estimation unit 21 estimates the wind speed distribution by executing the learning model 25 as a program on a CPU, for example.
The wind speed distribution estimation unit 21 sequentially inputs the individual input data corresponding to each of the individual regions LR1 to LR16 generated as described above to the learning model 25, and estimates the wind speed distribution corresponding to each individual region LR.

More specifically, when the wind speed distribution estimation unit 21 inputs individual input data to the learning model 25, the learning model 25 sequentially traces the convolutional layers 27a, 27b, 27c and the transposed convolutional layers 28c, 28b, 28a. Executes convolution processing and transposed convolution processing. Finally, an individual image corresponding to the estimation input data 30 is output from the transposed convolution layer 28a.
Similar to the learning estimation results 15, the individual images include wind speeds decomposed into components corresponding to each of the X direction, Y direction, and Z direction at each point in the area expressed in the building information data 31; These are three 1-channel images in which wind direction information is stored, and these are output as one 3-channel image.
The wind speed distribution estimating unit 21 calculates one-channel images corresponding to each of the X direction, Y direction, and Z direction from the individual images, that is, one channel image corresponding to each direction, for each individual image corresponding to each individual input data. An X component individual image, a Y component individual image, and a Z component individual image, which are component-decomposed wind speed and wind direction information, are extracted.

The wind speed distribution estimating unit 21 synthesizes the information of each of the X component individual image, the Y component individual image, and the Z component individual image to generate an individual estimation result image (an estimation result of the wind speed distribution in the individual region). The individual estimation result image is an image having the same resolution as the X component individual image, the Y component individual image, and the Z component individual image. The wind speed distribution estimating unit 21 calculates a pixel value from each of these images for each corresponding pixel of the X component individual image, Y component individual image, and Z component individual image, that is, the wind direction and wind speed at the point corresponding to the pixel. Obtain the values corresponding to each of the X component value, Y component value, and Z component value of Synthesize on the coordinate system. The wind speed distribution estimation unit 21 calculates the wind speed based on the magnitude of the vector generated by the synthesis, and the wind direction based on the direction. The wind speed distribution estimating unit 21 calculates wind speed and wind direction information for all pixels, and generates one image, that is, an individual estimation result image by combining them.
In this way, the wind speed distribution estimating unit 21 calculates the wind speed at each of a plurality of points (for example, one pixel in each figure in FIG. 4) included in the individual region LR, that is, for each individual input data. Then, the wind speed distribution within the individual region LR is estimated as an individual estimation result image.
In the example using FIG. 4, 16 individual estimation result images corresponding to each of 16 individual regions LR1 to LR16 are output.

For the individual estimation result image, which is the estimation result of the wind speed distribution in the individual region LR, output as described above, the weight set in the weight setting section 51, which will be described next, is applicable.
FIG. 6 is an example in which the weights provided to the individual regions LR are expressed by a weighting function W having a three-dimensional curved surface. The weighting function W is set to have the shape of a probability density function of normal distribution when viewed from the horizontal direction X or the vertical direction Y. The probability density function is set so that its median value is located at the center of each of the horizontal direction X and vertical direction Y of the individual region LR. In addition, the domain of the probability density function is set between -2 times the standard deviation and twice the standard deviation, and this domain is set to match the area in the horizontal direction X and vertical direction Y of the individual region LR. , a probability density function is provided. As a result, approximately 95% of the probability density function realized as a normal distribution is provided in association with the individual region LR.

As a result of the weight being set corresponding to the individual region LR based on the weighting function W as described above, the value of the weight becomes larger for a point located closer to the center of the individual region LR, and for a point located closer to the outer edge of the individual region LR. The smaller the weight value becomes. For example, a point PA located at the center of the individual region LR in both the horizontal direction X and the vertical direction Y is associated with a weight of 1, as indicated by the value VA. Further, a point PB located at the outer edge of the individual region LR in both the horizontal direction X and the vertical direction Y is associated with a weight close to 0, for example, 0.1, as shown as a value VB. Further, at a point PC located between points PA and PB, a value between them, such as 0.6, is associated.

A weighting function W as described above is individually set in each of the individual regions LR1 to LR16. In this embodiment, the same weighting function W is set for each of the individual regions LR1 to LR16. Therefore, even if the same point is located on the entire area GR, if it is a duplicate point included in the overlap area OR, the weight corresponding to the duplicate point is It differs depending on the regions LR1 to LR16.
For example, point P1 shown in FIG. 4 belongs only to individual region LR1. Therefore, point P1 is not an overlapping point. Further, since the point P1 is located at the outer edge of the individual region LR1 in both the horizontal direction X and the vertical direction Y, the weight of the point P1 in the individual region LR1 is a small value, for example, 0.1.

Moreover, point P2 shown in FIG. 4 is an overlapping point belonging to two individual regions LR1 and LR2.
Although this point P2 is located at the outer edge in the vertical direction Y in the individual region LR1, it is located at the center in the horizontal direction X, so it is larger than 0.1 but smaller than 1, for example, 0. It can take a weight value such as .3.
Moreover, since the point P2 is located at the outer edge of the individual region LR2 in both the horizontal direction X and the vertical direction Y, the weight of the point P2 in the individual region LR2 is a small value, for example, 0.1.

Furthermore, point P3 shown in FIG. 4 is an overlapping point belonging to four individual regions LR1, LR2, LR5, and LR6.
Since this point P3 is located at the outer edge of the individual region LR1 in both the horizontal direction X and the vertical direction Y, the weight of the point P3 in the individual region LR1 is a small value, for example, 0.1.
In addition, point P3 is located at the outer edge in the vertical direction Y in the individual region LR2, but is located at the center in the horizontal direction X, so it is larger than 0.1 but smaller than 1, e.g. It can take a weight value such as 0.3. Also in the individual region LR5, the weight of the point P3 is, for example, 0.3 for the same reason.
Furthermore, since the point P3 is located at the center of the individual region LR6 in both the horizontal direction X and the vertical direction Y, the weight of the point P3 in the individual region LR6 is a large value, for example 1.
In this way, different values are set as weights at each overlapping point depending on the individual region LR to which it belongs.

As described above, in the weight setting unit 51, a weight is individually set for each individual region LR for each of the overlapping points that are included in the overlapping portion OR.
Furthermore, in each of the plurality of individual regions LR, the weight setting unit 51 determines that when the median value of the probability density function of the normal distribution is located at the center of the individual region LR, The value of the probability density function at a position is set as the weight at that point.

The estimation result adjusting unit 52 adjusts the estimation result of each point including the overlapping point based on the weight of each overlapping point set in the weight setting unit 51 in each individual region LR.
More specifically, for each point including overlapping points, the estimation result adjustment unit 52 first calculates the sum SL of the weights set for the point in each individual region LR. Next, in each of the individual regions LR, the estimation result adjustment unit 52 calculates the division value obtained by dividing the weight set at the point by the total weight SL and the estimation result of the wind speed at the point, that is, the individual estimation result image. An individual region adjustment value LV, which is a multiplication value obtained by multiplying the pixel value of the corresponding pixel, is calculated. Finally, the estimation result adjustment unit 52 calculates the sum of the individual region adjustment values LV in each of the individual regions LR, and calculates this as the estimation result at the point where the wind speed has been adjusted.

For example, point P1 shown in FIG. 4 is not an overlapping point and belongs only to individual region LR1. If the weight of the point P1 in this individual region LR1 is, for example, 0.1, the estimation result adjustment unit 52 first calculates 0.1 as the total weight SL. Next, the estimation result adjustment unit 52 calculates the division value obtained by dividing the weight 0.1 of the point P1 in the individual region LR1 by the total weight SL, and the estimated result V1 of the wind speed at the point P1, that is, the individual region corresponding to the individual region LR1. The individual region adjustment value LV is calculated by multiplying by the wind speed estimation result V1 at the point P1 in the estimation result image. In this case, the division value is 0.1/0.1=1, so the individual area adjustment value LV is V1. Finally, the estimation result adjustment unit 52 calculates the value of V1 as the sum of the individual region adjustment values LV in each of the individual regions LR, and calculates this as the estimation result for the point P1 with the wind speed adjusted.
In other words, at a point P that is not an overlapping point, the wind speed calculated for the point P in the individual estimation result image corresponding to the individual area LR including the point P is unchanged as the estimated result with the wind speed adjusted. Calculated.

Further, point P2 shown in FIG. 4 is an overlapping point and belongs to two individual regions LR1 and LR2. If the weight of the point P2 in the individual region LR1 is, for example, 0.3, and the weight of the point P2 in the individual region LR2 is, for example, 0.1, the estimation result adjustment unit 52 first sets the total weight SL to 0.3+0.1. Calculate =0.4. Next, the estimation result adjustment unit 52 uses the estimation result V21 of the wind speed at the point P2 in the individual estimation result image corresponding to the individual region LR1 to set the individual region adjustment value LV21 in the individual region LR1 to LV21=V21×( 0.1/0.4). Further, the estimation result adjustment unit 52 uses the estimation result V22 of the wind speed at the point P2 in the individual estimation result image corresponding to the individual region LR2 to calculate the individual region adjustment value LV22 in the individual region LR2, LV22=V22×(0 .3/0.4). Finally, the estimation result adjustment unit 52 calculates the value V2=LV21+LV22 as the sum of the individual region adjustment values LV21 and LV22 in each of the individual regions LR1 and LR2, and uses this as the estimated result of the adjusted wind speed at the point P2. calculate.

Furthermore, point P3 shown in FIG. 4 is an overlapping point and belongs to four individual regions LR1, LR2, LR5, and LR6. If the weights of the point P3 in each of the individual regions LR1, LR2, LR5, and LR6 are, for example, 0.1, 0.3, 0.3, and 1, the estimation result adjustment unit 52 first sets the total weight SL to 0.1, 0.3, 0.3, and 1. Calculate 1+0.3+0.3+1=1.7. Next, the estimation result adjustment unit 52 uses the estimation results V31, V32, V35, and V36 of the wind speed at the point P3 in the individual estimation result images corresponding to each of the individual regions LR1, LR2, LR5, and LR6 to The individual region adjustment values LV31, LV32, LV35, and LV36 in each of LR1, LR2, LR5, and LR6 are determined as follows: LV31=V31×(0.1/1.7), LV32=V32×(0.3/1. 7), LV35=V35×(0.3/1.7), LV36=V36×(1/1.7). Finally, the estimation result adjustment unit 52 calculates the value V3=LV31+LV32+LV35+LV36 as the sum of the individual region adjustment values LV31, LV32, LV35, and LV36 in each of the individual regions LR1, LR2, LR5, and LR6, and calculates the value V3=LV31+LV32+LV35+LV36 at the point P3. Calculated as an estimated result with wind speed adjusted.

As exemplified by points P2 and P3, for overlapping points included in a plurality of individual regions LR, the wind speed estimation result at the point P in the individual estimation result image corresponding to each individual region LR is weighted. averaged. As explained using FIG. 6, the weight used at this time is implemented by a function W such that the weight is large near the center of each individual region LR and small near the outer edge. This reduces the influence near the outer edge of the individual region LR, which causes a reduction in the estimation accuracy of the wind speed distribution.
Further, the number of times of overlap, that is, the number of individual regions LR to which the point P belongs differs for each point P. Therefore, the weight used for averaging is a value standardized by dividing the weight of each individual region LR by the total weight SL.
In this way, the estimation result adjustment unit 52 provides the estimation results of the weight and wind speed of the points P1, P2, P3 in each of the individual regions LR for each of the points P1, P2, P3 including the overlapping points P2, P3. Based on LV1, LV21, LV22, LV31, LV32, LV35, and LV36, adjusted estimation results V1, V2, and V3 of the wind speeds at the points P1, P2, and P3 are calculated.

The estimation result combination unit 53 combines the estimation results of the wind speed distribution in the plurality of individual regions LR based on the estimation results adjusted by the estimation result adjustment unit 52, and generates the estimation result of the wind speed distribution in the entire region GR. do.
More specifically, the estimation result combining unit 53 has the same resolution as the estimation input data 30 (for example, 20 pixels in both the horizontal direction X and the vertical direction Y according to the example shown in FIG. 4), which corresponds to the entire region GR. An estimation result combined image (estimated result of wind speed distribution in the entire region GR) 54 is generated. In the combined estimation result image 54, each pixel stores the adjusted estimation results V1, V2, and V3 of the wind speed calculated by the estimation result adjustment unit 52 regarding the point P corresponding to the pixel.
The estimation result combination unit 53 outputs the estimation result combination image 54 to the outside.

Next, a wind speed distribution estimation method using the above-mentioned wind speed distribution estimation device 1 will be explained using FIGS. 1 to 6 and FIG. 7. FIG. 7(a) is a flowchart when learning the wind speed distribution, and FIG. 7(b) is a flowchart when estimating the wind speed distribution.
First, the operation of each component of the wind speed distribution estimation device 1 during wind speed distribution learning will be described.
The learning unit 20 inputs the building information data 3 and the wind direction data 4 as learning input data 14 to the convolution processing unit 27 of the machine learning device 24, and outputs a learning estimation result 15.
The learning unit 20 acquires the wind speed distribution data 5 corresponding to the input learning input data 14, compares the wind speed distribution data 5, that is, the teacher data, and the learning estimation result 15 pixel by pixel, and calculates, for example, the difference between each pixel. Calculate the squared error of the pixel value difference as a cost function.
Then, the machine learning device 24 performs machine learning by adjusting the weight values of each filter by using an error back propagation method or the like so as to reduce this cost function (step S1).
When the learning is completed, the learning unit 20 stores parameters such as weight values of each filter whose adjustment has been completed in the learning model parameter storage unit 22 as learning model parameters (step S3).

Next, the behavior of each component when estimating the wind speed distribution will be explained.
As a premise, in the wind speed distribution estimation device 1, if the entire region GR to be estimated is larger than the region that can be estimated by the wind speed distribution estimation unit 21, it is divided into a plurality of individual regions LR, and the wind speed distribution is estimated in each individual region. do. Therefore, if the region to be estimated (the entire region GR) is equal to or smaller than the individual region LR, there is no need to divide it into multiple regions.
The entire region dividing unit 50 divides the entire region for which wind speed distribution is estimated into a plurality of individual regions LR so as to have overlapping portions, and generates individual input data (step S11).
The wind speed distribution estimation unit 21 acquires the learning model parameters from the learning model parameter storage unit 22 and constructs the learning model 25 (step S13).
The wind speed distribution estimation unit 21 sequentially inputs individual input data corresponding to each of the individual regions LR to the learning model 25, estimates the wind speed distribution corresponding to each individual region LR, and generates an individual estimation result image ( Step S15).
For each point including overlapping points, the estimation result adjustment unit 52 first calculates the sum SL of the weights set for the point in each individual region LR. Next, in each of the individual regions LR, the estimation result adjustment unit 52 calculates the division value obtained by dividing the weight set at the point by the total weight SL and the estimation result of the wind speed at the point, that is, the individual estimation result image. An individual region adjustment value LV, which is a multiplication value obtained by multiplying the pixel value of the corresponding pixel, is calculated. Finally, the estimation result adjustment unit 52 calculates the sum of the individual region adjustment values LV in each of the individual regions LR, and calculates this as the estimation result at the point where the wind speed has been adjusted.
The estimation result combination unit 53 combines the estimation results of the wind speed distribution in the plurality of individual regions LR based on the estimation results adjusted by the estimation result adjustment unit 52 to obtain the estimation result of the wind speed distribution in the entire region GR. , generates the estimation result combined image 54 (step S17).

Next, the effects of the above-described wind speed distribution estimation device 1 and wind speed distribution estimation method will be explained.

The wind speed distribution estimating device 1 of the present embodiment is a wind speed distribution estimating device 1 that estimates the wind speed distribution around a building, and is configured so that the entire region GR, which is the estimation target of the wind speed distribution, has an overlapping portion OR that overlaps with each other. , the entire region dividing section 50 divides into a plurality of individual regions LR, the wind speed distribution estimating section 21 estimates the wind speed distribution within the plurality of individual regions LR, and the overlapping region OR is performed for each of the individual regions LR. The wind speed at the overlapping point is adjusted based on the weight setting unit 51, which has an individual weight set for each overlapping point that is the point P included in the overlapping point, and the estimation results of the weight and wind speed distribution of the overlapping point. Based on the adjusted estimation results, the estimation result adjusting unit 52 calculates the estimation results based on the adjusted estimation results, and combines the estimation results of the wind speed distribution in the plurality of individual regions LR to obtain the estimation result 54 of the wind speed distribution of the entire region. and an estimation result combination unit 53 that generates the estimation result.
Further, the wind speed distribution estimation method of the present embodiment is a wind speed distribution estimation method for estimating the wind speed distribution around a building, and the entire region GR, which is the estimation target of the wind speed distribution, is Then, it is divided into a plurality of individual regions LR, the wind speed distribution within the plurality of individual regions LR is estimated, and for each individual region LR, each overlapping point, which is a point P included in the overlapping portion OR, is individually calculated. Based on the set weights and the estimation results of the wind speed distribution, calculate the estimation results in which the wind speeds at the overlapped points are adjusted, and estimate the wind speed distributions in multiple individual regions LR based on the adjusted estimation results. The results are combined to generate an estimation result 54 of the wind speed distribution in the entire region GR.
According to the above configuration, the entire region GR, which is the target for estimating the wind speed distribution, is divided into a plurality of individual regions LR, and the wind speed distribution is estimated for each of the plurality of individual regions LR. The results are combined to generate an estimation result 54 of the wind speed distribution in the entire region GR.
That is, since the wide region GR is divided into small individual regions LR and the wind speed distribution is estimated for each of these, the wind speed distribution of the wide region can be estimated relatively easily.
Here, the entire region GR is divided into a plurality of individual regions LR so as to have overlapping portions OR that overlap with each other. That is, when dividing the entire region GR into individual regions LR, the division is not simply done by one boundary line, but is divided so that adjacent individual regions LR have a common overlapping portion OR. In this state, when combining the estimation results of the individual regions LR as the estimation results of the whole region GR, each individual region LR is The estimation results are adjusted. Therefore, discontinuity between individual regions LR is alleviated.
Furthermore, the overlapping portion OR is included in a plurality of mutually adjacent individual regions LR from different directions. Therefore, even if the estimation accuracy of the overlapping part OR located near the outer edge of each individual region LR is reduced due to the lack of information on buildings located outside the outer edge, these buildings will not be able to estimate the overlapping part OR. Since it is located within another individual region LR that is adjacent to it on both sides, the missing information described above is reflected in the estimation of the wind speed distribution in this adjacent individual region LR. Therefore, when combining the estimation results of the individual regions LR as the estimation results of the entire region GR, information regarding all the buildings located around the overlapping region OR is reflected and adjusted in the estimation results of the overlapping region OR. This makes it possible to suppress a decrease in estimation accuracy near the outer edge of the individual region LR, compared to a case where the evaluation target region is simply divided by a boundary line without providing an overlapping portion OR.
Due to the above reasons, the accuracy of estimating the wind speed distribution can be improved over almost the entire wide area.

Furthermore, the estimation result adjustment unit calculates the sum of weights set for each of the overlapping points, and divides the weight by the sum of the weights for each of the weights set for the overlapping points. After that, a multiplication value is calculated by multiplying the estimation result at the overlapped point, and the sum of the multiplication values is calculated, thereby calculating an adjusted estimation result for the overlapped point.
According to the above configuration, the above wind speed distribution estimation device 1 can be appropriately realized.

In addition, in the weight setting unit 51, when the median value of the probability density function of the normal distribution is located at the center of each individual region LR, the value of the probability density function at the position corresponding to the point P in the individual region is It is set as the weight at the point P.
As already explained, for overlapping points included in a plurality of individual regions LR, the wind speed estimation results at the point P in the individual estimation result images corresponding to each individual region LR are weighted and averaged. . If the weights used at this time are implemented using a function that is large near the center of each individual region LR and small near the outer edge, the influence near the outer edge of the individual region LR, which reduces the estimation accuracy of the wind speed distribution, will be small. Become.
Therefore, according to the above configuration, the above wind speed distribution estimation device 1 can be appropriately realized.

FIG. 8 is an example in which three-dimensional building information and the combined estimation result image 54 output by the wind speed distribution estimation device 1 in the above embodiment are superimposed. As shown in FIG. 8, according to the configuration of the embodiment described above, it is possible to output a wind speed distribution with continuous values over almost the entire wide area.

Note that the wind speed distribution estimation device and wind speed distribution estimation method of the present invention are not limited to the above-described embodiments described with reference to the drawings, and various other modifications can be considered within the technical scope thereof.

For example, in the above embodiment, an FCN is used as the machine learning device 24 and the learning model 25, but a neural network with another structure may be used as long as the estimation accuracy is not impaired.
Furthermore, the structure of the FCN is not limited to the above structure described using FIGS. 2 and 5, and may have other structures. For example, in the above embodiment, the convolution processing section 27 and the transposed convolution processing section 28 are both explained using a typical example of a three-layer structure, but the number of layers of each may be other than three, The number of layers in the convolution processing section 27 and the number of layers in the transposed and convolution processing section 28 may be different.
Furthermore, in the embodiment described above, the wind speed distribution estimation unit 21 estimated the wind speed distribution in the individual region LR using the learning model 25 based on a full-layer convolutional network. However, the wind speed distribution estimation unit 21 may estimate the wind speed distribution in the individual region LR by wind tunnel experiment or computational fluid analysis.

Further, in the embodiment described above, as the learning data 2, an image having a resolution corresponding to a small area, which is equivalent to the individual input data, is used. The learning data 2 itself may also be generated by cutting out and dividing data of some wider area in the same manner as the entire area dividing section 50.
Further, in the above embodiment, it has been explained that the stride in the horizontal direction X and the vertical direction Y is desirably half or less of the resolution in the horizontal direction X and vertical direction Y of the individual region LR. In reality, the smaller the stride, the more the number of overlaps at each point, that is, the number of individual regions LR to which point P belongs, increases, and the number of objects to be weighted and averaged increases. The discontinuities between the two are more relaxed. However, since reducing the stride also increases the amount of processing performed by the wind speed distribution estimating device 1, it is actually desirable to determine the size of the stride in consideration of both processing speed and accuracy.

In addition to this, it is possible to select the configurations mentioned in the above embodiments or to change them to other configurations as appropriate, without departing from the gist of the present invention.

1 Wind speed distribution estimation device 51 Weight setting unit 2 Learning data 52 Estimation result adjustment unit 5 Wind speed distribution data 53 Estimation result combination unit 21 Wind speed distribution estimation unit 54 Estimation result combined image (estimation result of wind speed distribution in the entire area)
24 Machine learning device GR Whole region 25 Learning model OR Overlapping part 30 Input data during estimation LR Individual region 50 Whole region dividing part W Weight function (weight)

Claims (2)

A wind speed distribution estimation device that estimates wind speed distribution around a building,
an entire area dividing unit that divides the entire area to be estimated for the wind speed distribution into a plurality of individual areas so as to have overlapping parts that overlap with each other;
a wind speed distribution estimation unit that estimates the wind speed distribution within the plurality of individual regions;
a weight setting unit in which a weight is individually set for each overlapping point that is a point included in the overlapping portion for each of the individual regions;
an estimation result adjustment unit that calculates an estimation result in which the wind speed of the overlapped point is adjusted based on the estimation result of the weight and the wind speed distribution of the overlapped point;
an estimation result combining unit that combines the estimation results of the wind speed distribution in the plurality of individual regions based on the adjusted estimation results to generate an estimation result of the wind speed distribution of the entire region;
Equipped with
The estimation result adjustment unit calculates, for each of the overlapping points, the sum of the weights set to the overlapping points, and adjusts the weight to the weight for each of the weights set to the overlapping points. A wind speed distribution estimating device that calculates the adjusted estimation result of the overlap point by calculating a multiplication value obtained by dividing by the sum and then multiplying the estimation result at the overlap point, and calculating the sum of the multiplication values. . A wind speed distribution estimation method for estimating wind speed distribution around a building, the method comprising:
dividing the entire area to be estimated for the wind speed distribution into a plurality of individual areas so as to have overlapping parts that overlap with each other;
estimating the wind speed distribution within the plurality of individual regions;
For each of the individual regions, based on the individually set weight for each overlapping point that is a point included in the overlapping part and the estimation result of the wind speed distribution, for each of the overlapping points, A multiplication value obtained by calculating the sum of the weights set at the point, dividing each of the weights set at the overlap point by the sum of the weights, and then multiplying the result by the estimation result at the overlap point. By calculating the sum of the multiplied values, an estimated result with the wind speed at the overlapped point adjusted is calculated.
A wind speed distribution estimation method, wherein the estimation results of the wind speed distribution in a plurality of individual regions are combined based on the adjusted estimation results to generate an estimation result of the wind speed distribution in the entire region.

Images