Super-resolution wind resource map generation method and device and readable storage medium

Abstract

According to the super-resolution wind resource atlas generation method and device and the readable storage medium provided by the embodiment of the invention, an electronic device obtains the first wind resource atlas of the target area and the topographic data of the target area, and the first wind resource atlas and the topographic data serve as input and are input into the pre-trained deep learning model. Afterwards, the electronic device processes the first wind resource map and the topographic data by using an attention layer, a deformable convolutional layer and a fusion layer of a deep learning model to obtain a second wind resource map of the target area. By adopting the scheme, the electronic equipment generates the high-resolution wind resource atlas in a deep learning mode, the purpose of improving the precision of the high-resolution wind resource atlas is achieved, and the time cost and the cloud computing cost are saved to a great extent. Moreover, topographic data is introduced as priori knowledge, so that the second wind resource map can adapt to topographic relief, and the purpose of improving the precision of the high-resolution wind resource map is further achieved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of wind farms, in particular to a super-resolution wind resource map generation method, device and readable storage medium. BACKGROUND

[0002] As a clean and renewable energy, wind energy is increasingly valued by countries around the world. Wind resource map is one of the key foundations for large-scale development of wind power.

[0003] Wind resource map is a visualization tool used to describe the distribution of wind energy resources in a specific area, usually presented in the form of a map, with wind speed, wind direction, wind power, etc. marked at different locations. According to the resolution, wind resource map can be divided into low-resolution wind resource map and high-resolution wind resource map. Among them, resolution refers to the actual map range represented by a single grid, low resolution means that each grid cell covers a larger geographic range, and high resolution means that each grid covers a smaller geographic range. Obviously, high-resolution wind resource map is crucial for wind farm site selection and wind turbine arrangement. In the traditional way, bilinear interpolation method is used to interpolate low-resolution wind resource map to obtain high-resolution wind resource map.

[0004] However, the high-resolution wind resource map obtained based on the bilinear interpolation method has low accuracy. SUMMARY

[0005] Embodiments of the present application provide a super-resolution wind resource map generation method, device and readable storage medium, which generates high-resolution wind resource map using deep learning method, and achieves the purpose of improving the accuracy of high-resolution wind resource map.

[0006] In a first aspect, embodiments of the present application provide a super-resolution wind resource map generation method, comprising:

[0007] obtaining a first wind resource map of a target area and topographic data indicating the topography of the target area;

[0008] inputting the first wind resource map and the topographic data into a pre-trained deep learning model, the deep learning model comprising an attention layer, a deformable convolution layer and a fusion layer in sequence;

[0009] processing the first wind resource map and the topographic data using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain a second wind resource map of the target area, the first resolution of the first wind resource map being lower than the second resolution of the second wind resource map.

[0010] In a second aspect, an embodiment of the present application provides an ultra-resolution wind resource map generation device, comprising:

[0011] an acquisition module configured to acquire a first wind resource map of a target region and topographic data indicating a topography of the target region;

[0012] an input module configured to input the first wind resource map and the topographic data into a pre-trained deep learning model, the deep learning model comprising, in sequence, an attention layer, a deformable convolution layer, and a fusion layer;

[0013] a processing module configured to process the first wind resource map and the topographic data by using the attention layer, the deformable convolution layer, and the fusion layer of the deep learning model, to obtain a second wind resource map of the target region, the first resolution of the first wind resource map being lower than the second resolution of the second wind resource map.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to enable the electronic device to implement the method according to the first aspect or any possible implementation manner of the first aspect.

[0015] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to implement the method according to the first aspect or any possible implementation manner of the first aspect when executed by a processor.

[0016] In a fifth aspect, an embodiment of the present application provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the method according to the first aspect or any possible implementation manner of the first aspect.

[0017] The method, device and readable storage medium for generating a super-resolution wind resource map provided by the embodiments of the present application, an electronic device obtains a first wind resource map of a target region and topographic data of the target region, takes the first wind resource map and the topographic data as inputs, and inputs them into a pre-trained deep learning model. The deep learning model sequentially includes an attention layer, a deformable convolution layer and a fusion layer. Then, the electronic device processes the first wind resource map and the topographic data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain a second wind resource map of the target region. The first resolution of the first wind resource map is lower than the second resolution of the second wind resource map. By using this scheme, the electronic device generates a high-resolution wind resource map by using a deep learning method, thereby achieving the purpose of improving the accuracy of the high-resolution wind resource map and greatly saving the time cost and cloud computing cost. Moreover, the topographic data is introduced as prior knowledge, so that the second wind resource map can be adapted to the terrain undulation, thereby further achieving the purpose of improving the accuracy of the high-resolution wind resource map. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 is a flowchart of the method for generating a super-resolution wind resource map provided by the embodiments of the present application;

[0020] Figure 2 is a structural schematic diagram of the deep learning model in the method for generating a super-resolution wind resource map provided by the embodiments of the present application;

[0021] Figure 3 is a structural schematic diagram of the attention sublayer b in the method for generating a super-resolution wind resource map provided by the embodiments of the present application; Figure 2

[0022] Figure 4 is a structural schematic diagram of the deformable convolution in the method for generating a super-resolution wind resource map provided by the embodiments of the present application; Figure 2

[0023] Figure 5 is a flowchart of splicing the second wind resource map of super-resolution in the method for generating a super-resolution wind resource map provided by the embodiments of the present application;

[0024] Figure 6 is a schematic diagram of the second wind resource map of the target range in the method for generating a super-resolution wind resource map provided by the embodiments of the present application;

[0025] Figure 7 ​​A schematic diagram of an ultra-resolution wind resource map generation device provided by an embodiment of the present application;

[0026] Figure 8 A structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0027] A wind resource map is used to indicate key parameters such as wind speed and wind power density at different locations, and is widely used in the fields of wind farm site selection, wind energy potential assessment, and energy planning. According to the geographical range corresponding to each grid, the wind resource map can be divided into low-resolution wind resource maps, high-resolution wind resource maps, etc. The high-resolution wind resource map is also called an ultra-resolution wind resource map. For example, each grid of a low-resolution wind resource map represents 5 kilometers (km) x 5 km, 9 km x 9 km, or a coarser granularity; each grid of an ultra-resolution wind resource map represents 1 km x 1 km, 100 meters x 100 meters, etc. Obviously, the high-resolution wind resource map is more refined. Therefore, how to obtain a high-resolution wind resource map is crucial.

[0028] A conventional scheme for obtaining an ultra-resolution wind resource map includes an ultra-resolution algorithm and a dynamic downscaling method. The ultra-resolution algorithm includes, but is not limited to, a bilinear interpolation method, etc. This scheme obtains a high-resolution wind resource map by interpolating a low-resolution wind resource map. However, this scheme does not consider terrain features, etc., resulting in insufficient accuracy of the ultra-resolution wind resource map in complex terrain areas.

[0029] In the dynamic downscaling method, a numerical model such as a Weather Research and Forecasting Model (WRF) is used to process atmospheric boundary layer parameters, etc., to obtain a high-resolution wind resource map. However, taking the creation of a high-resolution wind resource map for a country 1 as an example, assuming that the area of the country 1 is about 9 million square kilometers to 10 million square kilometers, the single calculation of this scheme takes about 8-12 hours, which is low in efficiency and high in calculation cost.

[0030] Moreover, the above-mentioned ultra-resolution algorithm and dynamic downscaling method have boundary effects when processing global data, affecting data consistency.

[0031] Therefore, an embodiment of the present application provides an ultra-resolution wind resource map generation method, device, and readable storage medium, which generates an ultra-resolution wind resource map by using a deep learning method, and achieves the purpose of improving the accuracy of the ultra-resolution wind resource map.

[0032] The execution subject of the embodiment of the present application is an electronic device, and a pre-trained deep learning model is deployed on the electronic device. The deep learning model sequentially includes an attention layer, a deformable convolution layer, and a fusion layer. The electronic device can obtain a first wind resource map and corresponding topographic data, and obtain a second wind resource map based on the first wind resource map, the topographic data, and the pre-trained deep learning model.

[0033] In the embodiment of the present application, the electronic device has great computing power, storage capacity, and the like. The electronic device can be hardware or software. When the electronic device is hardware, the electronic device is a single server or a distributed server cluster composed of multiple servers, a mobile phone, a notebook computer, and the like. When the electronic device is software, the electronic device can be multiple software modules or a single software module, and the like, and the embodiment of the present application is not limited.

[0034] Figure 1 FIG. 1 is a flowchart of a method for generating a super-resolution wind resource map provided by the embodiment of the present application. The execution subject of the embodiment is the electronic device described above, and the embodiment includes the following steps.

[0035] 101. Obtain a first wind resource map of a target region and topographic data for indicating the topography of the target region.

[0036] In the embodiment of the present application, the target region is a region on the earth, such as a country, a continent, a region of a country, and the like. The super-resolution wind resource map of the target region is unknown, but the low-resolution wind resource map of the target region is known. For example, the global low-resolution wind resource map is known, and the first wind resource map is extracted from the global low-resolution wind resource map according to the latitude and longitude of the target region.

[0037] In the embodiment of the present application, the resolution of the first wind resource map is a first resolution, and the resolution of the second wind resource map is a second resolution. The first resolution is less than the second resolution, that is, the first wind resource map is a low-resolution wind resource map, and the second wind resource map is a super-resolution wind resource map. For example, the first resolution is 9km x 9km, and the second resolution is 3km x 3km.

[0038] The embodiment of the present application does not limit the manner in which the electronic device obtains the first wind resource map and the corresponding topographic data. For example, the electronic device obtains the first wind resource map of the target region and the related topographic data from the cloud; for another example, the electronic device obtains a global first-resolution wind resource map from the cloud, and extracts the first wind resource map from the global first-resolution wind resource map according to the latitude and longitude of the target region. At the same time, the electronic device obtains global topographic data from the cloud, and extracts the topographic data of the target region therefrom.

[0039] Generally, the terrain data has a high resolution, such as 3km*3km. The terrain data includes, but is not limited to, elevation, terrain type, etc. By introducing the elevation and terrain type as prior knowledge, the accuracy of the super-resolution wind resource map of the complex terrain can be enhanced.

[0040] 102. inputting the first wind resource map and the terrain data into a pre-trained deep learning model, the deep learning model sequentially including an attention layer, a deformable convolution layer and a fusion layer.

[0041] In the embodiment of the application, the deep learning model is pre-trained, and the deep learning model sequentially includes an attention layer, a deformable convolution layer and a fusion layer from input to output. Through the attention layer, the positions with high wind speed and low wind speed in the target region, and the positions with high altitude and low altitude in the target region can be distinguished. The ordinary convolution is a fixed rectangle, while the first wind resource map and the second wind resource map have a point-to-point correspondence. However, based on this correspondence, a sub-region of the first wind resource map and a corresponding sub-region on the second wind resource map are not regular shapes. Therefore, by using the deformable convolution, the irregular correspondence between the first wind resource map and the second wind resource map can be learned, which conforms to the real world situation. Moreover, by introducing the elevation and terrain type as prior knowledge, the modeling capability of the complex terrain region is enhanced by dynamically adjusting the receptive field through the deformable convolution (DeformableConv2d).

[0042] 103. processing the first wind resource map and the terrain data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain the second wind resource map of the target region, the first resolution of the first wind resource map being lower than the second resolution of the second wind resource map.

[0043] In the embodiment of the application, the electronic device inputs the first wind resource map and the corresponding terrain data as input into the deep learning model, and the deep learning model processes the first wind resource map by using the attention layer and the deformable convolution layer in sequence, and processes the terrain data by using the attention layer. Then, the processing results of the first wind resource map and the processing results of the terrain data are fused by the fusion layer, so as to obtain the second wind resource map.

[0044] It is found through verification that the super-resolution wind resource map generation method provided in the embodiments of the present application can be used to create a super-resolution wind resource map for a country 1, for example. The area of the country 1 is about 9 million square kilometers to 10 million square kilometers. From inputting a first wind resource map of a first resolution and corresponding terrain data to outputting a second wind resource map of a second resolution, the single calculation is only in seconds, and only two teslav100s are needed for training and reasoning, which greatly saves the time cost. Moreover, the cloud computing cost is greatly saved. In addition, on multiple test sets, the mean squared error (MES) of the annual average wind speed of the super-resolution wind resource map generated by using the deep learning model is within 0.5 meters per second (m / s), and the robustness of the deep learning model is strong.

[0045] The super-resolution wind resource map generation method provided in the embodiments of the present application is that an electronic device obtains a first wind resource map of a target region and terrain data of the target region, takes the first wind resource map and the terrain data as inputs, and inputs them into a pre-trained deep learning model. The deep learning model sequentially includes an attention layer, a deformable convolution layer, and a fusion layer. Then, the electronic device processes the first wind resource map and the terrain data by using the attention layer, the deformable convolution layer, and the fusion layer of the deep learning model to obtain a second wind resource map of the target region. The first resolution of the first wind resource map is lower than the second resolution of the second wind resource map. By using this scheme, the electronic device generates a high-resolution wind resource map by using a deep learning method, and the purpose of improving the accuracy of the high-resolution wind resource map is achieved. The time cost and the cloud computing cost are greatly saved. Moreover, the terrain data is introduced as prior knowledge, so that the second wind resource map can be adapted to the terrain undulation, and the purpose of improving the accuracy of the high-resolution wind resource map is further achieved.

[0046] Optionally, after the electronic device processes the first wind resource map and the terrain data by using the attention layer, the deformable convolution layer, and the fusion layer of the deep learning model to obtain the second wind resource map of the target region, the position of a wind farm in the target region and the arrangement mode of wind turbine generators in the wind farm are planned according to the second wind resource map.

[0047] For example, when a user expects to plan a wind farm in a target region, a super-resolution wind resource map of the target region is first generated by using the scheme described in the embodiments of the present application. Then, the electronic device selects a position suitable for establishing a wind farm in the target region according to the super-resolution wind resource map of the target region. The position is a region with rich wind resources in the target region.

[0048] In addition, when the arrangement of the wind turbine units in the target region is unreasonable or the target region needs to be planned with wind turbine units, the electronic device takes the super-resolution wind resource map of the target region as a basis and arranges the models. For example, which models to choose and how to arrange.

[0049] By adopting the scheme, the electronic device selects the site of the wind farm or arranges the models according to the wind resource map of the target region, so as to accurately select the position of the wind farm and reasonably arrange the wind turbine units, and further to improve the utilization rate of wind energy.

[0050] Optionally, in the implementation, in the process that the electronic device processes the first wind resource map and the terrain data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain the second wind resource map of the target region, first, the electronic device processes the first wind resource map by using the attention layer of the deep learning model to obtain first attention features, and processes the terrain data by using the attention layer of the deep learning model to obtain second attention features. The first attention features are used to distinguish the wind resources at different positions in the target region, and the second attention features are used to distinguish the heights of different positions in the target region. Second, the electronic device processes the first attention features by using the deformable convolution layer of the deep learning model to obtain multi-scale wind speed features, and the multi-scale wind speed features are used to indicate the features of the wind resources in the target region at different scales. Then, the electronic device fuses the multi-scale wind speed features and the second attention features to obtain the second wind resource map of the target region.

[0051] Figure 2 is a structural diagram of a deep learning model in a super-resolution wind resource map generation method provided by an embodiment of the present application. Please refer to Figure 2 The deep learning model includes, from input to output, an attention layer, a deformable convolution layer and a fusion layer. The attention layer includes an attention sub-layer a and an attention sub-layer b. The attention sub-layer a is used to process the first wind resource map to obtain first attention features, and the attention sub-layer b is used to process the terrain data to obtain second attention features.

[0052] In the embodiments of the present application, the role of the attention layer is to highlight the positions that need more attention. For example, for the first wind resource atlas, the positions with higher wind speed in the target region need to be highlighted; for the terrain data, the positions with higher elevation in the template region need to be highlighted. Taking the terrain data as an example, in order to highlight the positions with higher elevation, the height of each position is normalized, and the height of each position after normalization is between 0 and 1. In order to highlight the difference of the terrain height feature, the height of the high-elevation position needs to be set higher, and the height of the low-elevation position needs to be set lower. For example, the height of a high-elevation position is normalized to 0.6, and it is set to 0.8 or even higher through the attention sub-layer b; for example, the height of a low-elevation position is normalized to 0.1, and it is set to 0.01 through the attention sub-layer b.

[0053] Please refer to Figure 2 In the deformable convolution, there are four gray filled rectangles, and each gray filled rectangle represents one 1.45 times progressive adaptive up-sampling. After the original low-resolution first attention feature passes through the first gray filled rectangle, the resolution is multiplied by 1.45; after passing through the second gray filled rectangle, the resolution is multiplied by 1.45 again on the basis of the above multiplication; after passing through the third gray filled rectangle, the resolution is multiplied by 1.45 again on the basis of the second gray filled rectangle, and then one deformation is performed, so that the high-resolution multi-scale wind speed feature is obtained. The multi-scale wind speed feature includes the average wind speed, the turbulence intensity, the wind direction, etc. of the target region. This deep learning model combining deformable convolution is also called a model using a multi-level residual network structure (Residual-in-Residual Dense Block), which captures multi-scale wind speed features through 16 dense residual blocks, combines nearest neighbor interpolation with deformable convolution, realizes 1.45 times progressive adaptive up-sampling, and avoids the blur effect introduced by traditional bicubic interpolation.

[0054] Figure 3 is Figure 2 The structure diagram of the attention sub-layer b. Please refer to Figure 3 The attention sub-layer b includes a normalization unit, three two-dimensional convolutions (conv 2d) and an enhanced terrain unit. Among them, the three two-dimensional convolutions are conv 2d 6 layers, conv 2d 4 layers and conv 2d 2 layers, the conv 2d 6 layers include 6 consecutive convolution layers, the conv 2d 4 layers include 4 convolution layers, and the conv 2d 2 layers include 2 convolution layers. The attention sub-layer b is constructed by stacking multiple two-dimensional convolution layers to build an attention mechanism. These hierarchical structures extract and fuse multi-scale features step by step, and enhance the focus of the deep learning model on the positions with high elevation in the target region.

[0055] Please refer to Figure 3, the normalization unit of the attention sub-layer b normalizes the terrain data to obtain normalized terrain X, and takes the normalized terrain X as the input of the conv 2d 6 layer, the conv 2d 4 layer and the conv 2d 2 layer, so that the two-dimensional convolutions output features A, B and C respectively, the feature A is a multi-scale terrain feature, including but not limited to mountain trend, watershed boundary and the like; the feature B includes but is not limited to a terrain transition zone and the like; and the feature C is a basic terrain, including but not limited to an edge, a local texture and the like. The electronic device obtains an enhanced terrain Y based on the normalized terrain X, the features A, B and C output by the two-dimensional convolutions in the enhanced terrain unit, Y=A×(X+C)+B. The enhanced terrain Y is the second attention feature.

[0056] In the embodiment of the application, the first wind resource atlas is mainly used to indicate the wind speed at each position. Please refer to the processing of the height by the attention sub-layer b for the processing of the wind speed by the attention sub-layer a, which will not be repeated here.

[0057] Figure 4 is Figure 2 The structural diagram of the deformable convolution in the embodiment of the application. Please refer to Figure 4 Generally, the convolution sampling region corresponding to the standard convolution operation is a square of the size of a convolution kernel, such as the larger square in the first attention feature. In the embodiment of the application, the deformable convolution is adopted, and the convolution sampling region corresponding to the deformable convolution is the region shown by the nine small squares in the first attention feature.

[0058] Please refer to Figure 4 The points in a small square region in the multi-scale wind speed feature correspond to the convolution sampling region on the first attention feature, and the convolution sampling region is the region shown by the nine small squares. Each convolution sampling point in the region shown by the nine small squares needs to learn an offset, and there are nine offsets in total, and the sum of all the offsets constitutes an offset field.

[0059] By using this scheme, the electronic device obtains the first attention feature and the second attention feature through the attention layer, thereby highlighting the positions with large wind speed and high altitude in the target region, processes the first attention feature by using the deformable convolution layer to obtain the multi-scale wind speed feature, and fuses the multi-scale wind speed feature and the second attention feature to obtain the wind resource atlas of the second resolution, thereby achieving the purpose of improving the accuracy of the second wind resource atlas.

[0060] Optionally, before the electronic device fuses the multi-scale wind speed features and the second attention features to obtain the second wind resource atlas of the target region, the electronic device further acquires a terrain category of each position in the target region according to the terrain data, and convolves the second attention features and the terrain category, so that the second attention features distinguish the height of each position in the target region according to the terrain category.

[0061] In the embodiments of the present application, the electronic device determines the terrain category of each position in the target region, and different terrain categories are represented by different numbers [0, 1, 2, 3, 4, 5, 6…N]. For example, 0 represents a plain, 1 represents a plateau, 2 represents a canyon, 3 represents a forest, and so on. The terrain category is taken as a new channel of terrain features. Then, after the electronic device processes the terrain data by using the attention layer to obtain the second attention features, the electronic device convolves the second attention features and the terrain category, so that the second attention features distinguish the height of each position in the target region according to the terrain category. For example, the electronic device generates a grid map according to the terrain category, the grid map has the same geographical range as the target region, and each grid is labeled with the terrain category of the corresponding position, such as a plateau, a canyon, and so on. The grid map has the same size and range as the terrain data and the first wind resource atlas. Therefore, the grid map and the feature map corresponding to the second attention features have the same size and range, and the second attention features and the terrain category can be convolved.

[0062] Since different terrain categories pay attention to different heights, for example, a plateau pays attention to an altitude of 500 meters or more, and a plain pays attention to an altitude of 200 meters or less. Therefore, by considering the terrain category of each position in the target region, the second attention features and the terrain category can be adapted, the accuracy of the second attention features is improved, and the accuracy of the second wind resource atlas is improved.

[0063] In the embodiments of the present application, the target region is a small region on the earth, such as a country, a continent, a province of a country, and so on. When a super-resolution wind resource atlas of a larger target range is needed, the larger range needs to be divided into multiple regions, a super-resolution second wind resource atlas is obtained for each region, and the second wind resource atlases of the regions in the target range are spliced to obtain a second wind resource atlas of the target range. The target range is, for example, the whole world, the northern hemisphere, and so on, which is not limited in the embodiments of the present application. Figure 5 FIG. 1 is a flowchart of a method for generating a super-resolution wind resource atlas provided in the embodiments of the present application. The method includes the following steps:

[0064] 501. Divide the target range into multiple regions.

[0065] The electronic device divides the target range into multiple areas based on longitude and latitude. It should be noted that the division of areas in the embodiments of the present application is not to divide the target range into multiple rectangular areas on a plane map, but to divide based on longitude and latitude.

[0066] 502. Projecting the second wind resource atlas of the target area to the target range obtains a first sector, and projecting the second wind resource atlas of the adjacent area of the target area to the target range obtains a second sector.

[0067] In the embodiments of the present application, since the earth is a sphere and the second wind resource atlas output by the deep learning model is a rectangle, the projection of each rectangular second wind resource atlas to the target range is a sector. Therefore, the electronic device projects the second wind resource atlas of each area in the target range to the target range on the surface of the earth, thereby obtaining a sector by sector. Among the sectors, the sector obtained by projecting the second wind resource atlas of the target area is called the first sector, and the sector obtained by projecting the second wind resource atlas of the adjacent area adjacent to the target area is called the second sector. The first sector and the second sector have an overlapping area.

[0068] 503. For each longitude and latitude position in the overlapping area of the first sector and the second sector, a first position closest to the longitude and latitude position is determined from the first sector, and a second position closest to the longitude and latitude position is determined from the second sector.

[0069] In the embodiments of the present application, for each longitude and latitude position in the overlapping area, two positions can be found from the first sector and the second sector by solving the Euclidean distance. The first position is the position in the first sector closest to the longitude and latitude position, and the second position is the position in the second sector closest to the longitude and latitude position.

[0070] 504. According to the wind speed of the first position and the wind speed of the second position, a fused wind speed of the longitude and latitude position is determined to obtain a fused wind speed of each longitude and latitude position in the overlapping area.

[0071] For each longitude and latitude position, the electronic device determines a first position from the first sector and a second position from the second sector, and then performs weighted average on the wind speeds of the two positions to obtain the fused wind speed of the longitude and latitude position. In this way, the electronic device can obtain the fused wind speed of each longitude and latitude position in the overlapping area.

[0072] 505. According to the second wind resource atlas of each area in the target range and the fused wind speed of each longitude and latitude position in the overlapping area between each area and an adjacent area, the second wind resource atlas of each area in the target range is spliced to obtain a second wind resource atlas of the target range.

[0073] The electronic device keeps the wind speed of each position in the non-overlapping area unchanged, and for each latitude and longitude position of the overlapping area, takes the fusion wind speed as the wind speed of the latitude and longitude position, so as to splice the second wind resource map of each area in the target range. Then, the spliced second wind resource map is unfolded, so as to obtain the second wind resource map of the target range. For example, please refer to Figure 6 .

[0074] Figure 6 is a schematic diagram of the second wind resource map of the target range in the super-resolution wind resource map generation method provided by the embodiment of the present application. Please refer to Figure 6 , the second wind resource map is a rectangle, the abscissa represents the longitude, the ordinate represents the latitude, and the gray bar on the right represents the wind speed represented by different gray scales.

[0075] With this scheme, when a second wind resource map of a larger target range needs to be obtained, the target range is divided into multiple areas, a second wind resource map of each area is obtained and spliced, and the purpose of quickly obtaining a second wind resource map of the target range is achieved. At the same time, for the overlapping area, the fusion wind speed is obtained to realize fusion, so as to avoid the existence of boundary effect and achieve the purpose of improving data consistency.

[0076] The above embodiments mainly describe how to generate a second resolution wind resource map of a target area by using a deep learning model. Next, how to train a deep learning model is described in detail.

[0077] When training a deep learning model, first, the electronic device obtains a second resolution supervision wind resource map and a first resolution global wind resource map.

[0078] For example, the first resolution global wind resource map is known, and the resolution of the global wind resource map is, for example, 9km x 9km. The supervision wind resource map is also known, which is a super-resolution wind resource map, and the resolution is, for example, 3km x 3km.

[0079] Secondly, the electronic device generates an initial wind resource map of the second resolution according to the geographical range covered by the supervision wind resource map, and the geographical range covered by the initial wind resource map is the same as the geographical range covered by the supervision wind resource map.

[0080] In the embodiments of the present application, since the low-resolution global wind resource map is a whole, and the geographical range corresponding to the supervised wind resource map is only a region on the earth, such as a country, a province, a continent, etc. Therefore, the electronic device needs to obtain a low-resolution wind resource map covering the same geographical range as the supervised wind resource map. In one way, the electronic device extracts a region from the global wind resource map according to the latitude and longitude coordinates of the geographical range covered by the supervised wind resource map, thereby obtaining a sample wind resource map of a first resolution, and the corresponding geographical range of the sample wind resource map of the first resolution is the same as the geographical range covered by the supervised wind resource map.

[0081] In another way, the electronic device initializes an initial wind resource map of a second resolution, each position in the initial wind resource map has a wind speed of null, and the geographical range covered by the initial wind resource map is the same as the geographical range covered by the supervised wind resource map. The electronic device preprocesses the initial wind resource map according to the global wind resource map to obtain a sample wind resource map of a first resolution. The preprocessing includes but is not limited to establishing a mapping relationship between the corresponding points of the initial wind resource map and the supervised wind resource map by the Euclidean distance, etc.

[0082] After the electronic device obtains the sample wind resource map of a first resolution corresponding to the supervised wind resource map, the electronic device takes the supervised wind resource map as the true value data, and continuously trains the initial model by using the sample wind resource map. In the training process, the parameters of the initial model are adjusted under the guidance of the back propagation function until the deep learning model is trained.

[0083] By using this scheme, a large number of sample wind resource maps are quickly and accurately obtained by constructing the sample wind resource map based on the known global wind resource map, so as to quickly and accurately train the deep learning model.

[0084] Optionally, in the above embodiments, in the process of preprocessing the initial wind resource map according to the global wind resource map to obtain a sample wind resource map of a first resolution, for each grid in the initial wind resource map, the electronic device determines a reference position from the global wind resource map according to the grid position of the grid, and the reference position is the position in the global wind resource map having the minimum Euclidean distance from the grid position. Then, the electronic device determines the wind speed value of the grid position according to the wind speed value of the reference position, to obtain the wind speed value of each grid in the initial wind resource map, and down-samples the initial wind resource map having the wind speed value of each grid to obtain the sample wind resource map of the first resolution.

[0085] For example, the resolution of the supervised wind resource map is 9x9, and there are 81 grids in total. The initial wind resource map is also divided into 81 grids, and the resolution of the sample wind resource map is 3x3, i.e., there are 9 grids in total.

[0086] For each of the 81 grids of the initial wind resource map, the electronic device determines the Euclidean distance between the grid position and each position of the global wind resource map, and sorts the Euclidean distances in ascending order. The position with the smallest Euclidean distance is taken as the reference position, and the wind speed value of the reference position is assigned to the grid. The grid position is, for example, the longitude and latitude of the grid center. In this way, each grid of the initial wind resource map obtains a wind speed value, and there are 81 wind speed values in total. Then, the electronic device down-samples the initial wind resource map with wind speed values for each grid, thereby obtaining a sample wind resource map with only 9 grids, i.e., a sample wind resource map with a first resolution.

[0087] It should be noted that 9x9 and 3x3 in the embodiment represent the number of grids, not 9kmx9km and 3kmx3km. The more the number of grids, the higher the resolution, and the smaller the number of grids, the lower the resolution.

[0088] It should be further noted that the number of grids 9x9 and 3x3 in the above embodiment is only an example, and in practice, the number of grids is much larger than 9x9 and 3x3. For example, in practice, the number of grids is 1000x1000, 5000x5000, 10,000x10,000, etc. The embodiment of the present application is not limited.

[0089] With this scheme, after the electronic device assigns wind speed values to each position of the initial wind resource map, the first resolution sample wind resource map is obtained by down-sampling, achieving the purpose of quickly and accurately obtaining the sample wind resource map.

[0090] Optionally, before the electronic device trains the deep learning model using the sample wind resource map, the supervised wind resource map, and the terrain data of the geographical range covered by the supervised wind resource map, the electronic device further performs mask processing on the region outside the reference region in the supervised wind resource map. The reference region is the region of the known wind resource map in the geographical range covered by the supervised wind resource map.

[0091] For example, the geographical range of the supervised wind resource map covers not only the country where the super-resolution wind resource map is known, but also other countries on the edge. For example, the geographical range is mainly country 1, but there are also parts of country 2, country 3 and country 4 around country 1, and the super-resolution wind resource map of these areas is unknown. In order to avoid the influence of these areas on the training of the deep learning model, the electronic device performs mask processing on the area where country 2, country 3 and country 4 of the supervised wind resource map are located. In this way, the data of this part of the area is invalid or unreliable during the model training process, so it will not affect the training of the deep learning model.

[0092] In addition, the electronic device can also optimize the high gradient area of the wind speed gradient by designing the mask, so as to improve the accuracy of the deep learning model.

[0093] By using this scheme, the unknown area in the supervised wind resource map is masked, so that the unknown area does not affect the training of the deep learning model, thereby focusing on optimizing the concerned area, and achieving the purpose of improving the accuracy of the deep learning model.

[0094] The following is an embodiment of the device of the present application, which can be used to execute the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0095] Figure 7 A schematic diagram of a super-resolution wind resource map generation device provided by an embodiment of the present application. The super-resolution wind resource map generation device 700 includes an acquisition module 71, an input module 72 and a processing module 73.

[0096] The acquisition module 71 is configured to acquire a first wind resource map of a target area and topographic data for indicating the topography of the target area.

[0097] The input module 72 is configured to input the first wind resource map and the topographic data into a pre-trained deep learning model, and the deep learning model sequentially includes an attention layer, a deformable convolution layer and a fusion layer.

[0098] The processing module 73 is configured to process the first wind resource map and the topographic data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model, to obtain a second wind resource map of the target area, and a first resolution of the first wind resource map is lower than a second resolution of the second wind resource map.

[0099] In an implementable manner, when the processing module 73 processes the first wind resource atlas and the terrain data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain the second wind resource atlas of the target region, the first attention feature is obtained by processing the first wind resource atlas by using the attention layer of the deep learning model, the second attention feature is obtained by processing the terrain data by using the attention layer of the deep learning model, the first attention feature is used to distinguish the wind resource level of each position in the target region, and the second attention feature is used to distinguish the height of each position in the target region; the multi-scale wind speed feature is obtained by processing the first attention feature by using the deformable convolution layer of the deep learning model, the multi-scale wind speed feature is used to indicate the characteristics of the wind resource in the target region at different scales, and the second wind resource atlas of the target region is obtained by fusing the multi-scale wind speed feature and the second attention feature.

[0100] In an implementable manner, before the processing module 73 fuses the multi-scale wind speed feature and the second attention feature to obtain the second wind resource atlas of the target region, the processing module 73 is further configured to obtain the terrain category of each position in the target region according to the terrain data; and the second attention feature and the terrain category are convolved so that the second attention feature distinguishes the height of each position in the target region according to the terrain category.

[0101] In an implementable manner, the target region is one of a plurality of regions included in a target range, and after the processing module 73 processes the first wind resource atlas and the terrain data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain the second wind resource atlas of the target region, the processing module 73 is further configured to project the second wind resource atlas of the target region to the target range to obtain a first sector, project the second wind resource atlas of an adjacent region of the target region to the target range to obtain a second sector; for each latitude and longitude position in the overlapping region of the first sector and the second sector, determine a first position closest to the latitude and longitude position from the first sector, and determine a second position closest to the latitude and longitude position from the second sector; determine the fusion wind speed of the latitude and longitude position according to the wind speed of the first position and the wind speed of the second position, to obtain the fusion wind speed of each latitude and longitude position in the overlapping region of each region and the adjacent region; and splice the second wind resource atlas of each region in the target range according to the second wind resource atlas of each region in the target range, the fusion wind speed of each latitude and longitude position in the overlapping region of each region and the adjacent region, to obtain the second wind resource atlas of the target range.

[0102] In a feasible implementation manner, before the input module 72 inputs the first wind resource atlas and the terrain data into the pre-trained deep learning model, the processing module 73 is further configured to acquire a supervised wind resource atlas of the second resolution and a global wind resource atlas of the first resolution, generate an initial wind resource atlas of the second resolution according to a geographical range covered by the supervised wind resource atlas, the geographical range covered by the initial wind resource atlas being the same as the geographical range covered by the supervised wind resource atlas, and preprocess the initial wind resource atlas according to the global wind resource atlas to obtain a sample wind resource atlas of the first resolution, and train the deep learning model by using the sample wind resource atlas, the supervised wind resource atlas and terrain data of the geographical range covered by the supervised wind resource atlas.

[0103] In a feasible implementation manner, when the processing module 73 preprocesses the initial wind resource atlas according to the global wind resource atlas to obtain a sample wind resource atlas of the first resolution, the processing module 73 is configured to determine a reference position from the global wind resource atlas according to a grid position of each grid in the initial wind resource atlas, the reference position being a position in the global wind resource atlas having the minimum Euclidean distance from the grid position, determine a wind speed value of the grid position according to a wind speed value of the reference position to obtain a wind speed value of each grid in the initial wind resource atlas, and down-sample the initial wind resource atlas having wind speed values of each grid to obtain the sample wind resource atlas of the first resolution.

[0104] In a feasible implementation manner, before the processing module 73 trains the deep learning model by using the sample wind resource atlas, the supervised wind resource atlas and terrain data of the geographical range covered by the supervised wind resource atlas, the processing module 73 is further configured to perform mask processing on a region other than a reference region in the supervised wind resource atlas, the reference region being a region of a known wind resource atlas in the geographical range covered by the supervised wind resource atlas.

[0105] In a feasible implementation manner, after the processing module 73 processes the first wind resource atlas and the terrain data by using the attention layer, the deformable convolution layer and the fusion layer of the deep learning model to obtain a second wind resource atlas of the target region, the processing module 73 is further configured to plan a position of a wind farm and an arrangement manner of wind turbine generators in the wind farm from the target region according to the second wind resource atlas.

[0106] The super-resolution wind resource atlas generation apparatus provided by the embodiments of the present application can perform the actions of the electronic device in the above embodiments, and has similar implementation principles and technical effects, which will not be described herein again.

[0107] Figure 8A structural schematic diagram of an electronic device is provided in the embodiments of the present application. The electronic device 800 comprises:

[0108] The processor 81 and the memory 82;

[0109] The memory 82 stores computer instructions;

[0110] The processor 81 executes the computer instructions stored in the memory 82, so that the processor 81 performs the log collection method implemented by the edge cluster or the center cluster as described above.

[0111] The specific implementation process of the processor 81 can refer to the method embodiments described above, which have similar implementation principles and technical effects. Therefore, no further description is given here.

[0112] Optionally, the electronic device 800 further comprises a communication component 83. The processor 81, the memory 82 and the communication component 83 can be connected through a bus 84.

[0113] The embodiments of the present application further provide a computer readable storage medium, wherein the computer readable storage medium stores computer instructions, and the computer instructions are executed by a processor to implement the super-resolution wind resource map generation method as described above.

[0114] The embodiments of the present application further provide a computer program product, which contains a computer program, and the computer program is executed by a processor to implement the super-resolution wind resource map generation method as described above.

[0115] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed here. The present application is intended to cover any variations, uses or adaptive changes of the present application following the general principles of the present application and including those expressly suggested or inherently related to the present technology in the art. The specification and examples are only regarded as exemplary, and the true scope and spirit of the present application are indicated by the following claims.

[0116] It should be understood that the present application is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present application is only limited by the appended claims.

Claims

1. A method for generating super-resolution wind resource maps, characterized in that, include: Obtain a first wind resource map of the target area and topographic data indicating the terrain of the target area; The first wind resource map and the terrain data are input into a pre-trained deep learning model, which includes an attention layer, a deformable convolutional layer and a fusion layer in sequence. The first wind resource map and the terrain data are processed using the attention layer, deformable convolutional layer and fusion layer of the deep learning model to obtain a second wind resource map of the target area. The first resolution of the first wind resource map is lower than the second resolution of the second wind resource map.

2. The method according to claim 1, characterized in that, The step of processing the first wind resource map and the terrain data using the attention layer, deformable convolutional layer, and fusion layer of the deep learning model to obtain the second wind resource map of the target area includes: The first wind resource map is processed using the attention layer of the deep learning model to obtain a first attention feature, and the terrain data is processed using the attention layer of the deep learning model to obtain a second attention feature. The first attention feature is used to distinguish the wind resource level at each location within the target area, and the second attention feature is used to distinguish the height of each location within the target area. The first attention feature is processed using the deformable convolutional layer of the deep learning model to obtain multi-scale wind speed features. These multi-scale wind speed features are used to indicate the characteristics of wind resources within the target area at different scales. The multi-scale wind speed features and the second attention features are fused to obtain a second wind resource map of the target area.

3. The method according to claim 2, characterized in that, Before fusing the multi-scale wind speed features and the second attention features to obtain the second wind resource map of the target area, the method further includes: The terrain category of each location in the target area is obtained based on the terrain data; The second attention feature and the terrain category are convolved so that the second attention feature can distinguish the height of each location within the target area according to the terrain category.

4. The method according to any one of claims 1 to 3, characterized in that, The target region is one of multiple regions encompassed by the target range. After processing the first wind resource map and the terrain data using the attention layer, deformable convolutional layer, and fusion layer of the deep learning model to obtain the second wind resource map of the target region, the method further includes: The second wind resource map of the target area is projected onto the target range to obtain a first sector, and the second wind resource maps of the adjacent areas of the target area are projected onto the target range to obtain a second sector; For each latitude and longitude position in the overlapping area of ​​the first sector and the second sector, a first position closest to the latitude and longitude position is determined from the first sector, and a second position closest to the latitude and longitude position is determined from the second sector; Based on the wind speed at the first location and the wind speed at the second location, the combined wind speed at the latitude and longitude locations is determined to obtain the combined wind speed at each latitude and longitude location within the overlapping area. Based on the second wind resource map of each region within the target range and the fused wind speed at each latitude and longitude position within the overlapping area of ​​each region and its adjacent regions, the second wind resource map of each region within the target range is stitched together to obtain the second wind resource map of the target range.

5. The method according to any one of claims 1 to 3, characterized in that, Before inputting the first wind resource map and the terrain data into the pre-trained deep learning model, the method further includes: Obtain the supervised wind resource map at the second resolution and the global wind resource map at the first resolution; An initial wind resource map of the second resolution is generated based on the geographical area covered by the monitored wind resource map, wherein the geographical area covered by the initial wind resource map is the same as that covered by the monitored wind resource map. The initial wind resource map is preprocessed based on the global wind resource map to obtain a sample wind resource map with a first resolution; The deep learning model is trained using the sample wind resource map, the supervised wind resource map, and the topographic data of the geographical area covered by the supervised wind resource map.

6. The method according to claim 5, characterized in that, The step of preprocessing the initial wind resource map based on the global wind resource map to obtain a sample wind resource map with a first resolution includes: For each grid in the initial wind resource map, a reference position is determined from the global wind resource map based on the grid position. The reference position is the position in the global wind resource map with the smallest Euclidean distance from the grid position. The wind speed value at the grid location is determined based on the wind speed value at the reference location, so as to obtain the wind speed value of each grid in the initial wind resource map; The initial wind resource map with wind speed values ​​for each grid is downsampled to obtain a sample wind resource map with the first resolution.

7. The method according to claim 5, characterized in that, Before training the deep learning model using the sample wind resource map, the supervised wind resource map, and the terrain data of the geographical area covered by the supervised wind resource map, the method further includes: The regions outside the reference region in the monitored wind resource map are masked. The reference region is the area of ​​the known wind resource map within the geographical area covered by the monitored wind resource map.

8. The method according to any one of claims 1 to 3, characterized in that, After processing the first wind resource map and the terrain data using the attention layer, deformable convolutional layer, and fusion layer of the deep learning model to obtain the second wind resource map of the target area, the method further includes: Based on the second wind resource map, the location of the wind farm and the arrangement of the wind turbine generators within the wind farm are planned from the target area.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it causes the terminal device to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images