A method for correcting the complex micro-topography effect on typhoon parameterized wind field

Abstract

The present invention discloses a method for correcting the complex micro-topography effect on a typhoon parameterized wind field, comprising: S1, generating a data set, at least including: wind field data considering the micro-topography effect obtained by numerical model simulation; wind field data not considering the micro-topography effect obtained by parameterized wind field model simulation; geographical environment data characterizing the micro-topography; S2, constructing a micro-topography correction model: processing the training data as the micro-topography correction model; constructing a micro-topography correction model based on a deep learning network and training it, and capturing the spatial characteristics of the micro-topography and the wind field and the coupling relationship between the micro-topography and the wind field through the micro-topography correction model; S3, inputting the wind field data and geographical environment data not considering the micro-topography effect into the trained micro-topography correction model, thereby obtaining a parameterized wind field considering the micro-topography effect. The present invention effectively considers the influence of complex micro-topography and can provide a more reliable typhoon parameterized wind field for engineering applications.

Description

Technical Field

[0001] The present invention relates to the technical field of typhoon wind field simulation, and in particular to a method for correcting the influence of complex micro-topography on a typhoon parameterized wind field. Background Art

[0002] Typhoon disasters are a type of meteorological disaster that often occurs in coastal and some inland areas, generally causing serious casualties and property losses. In order to scientifically assess the risks caused by typhoons to infrastructure systems (such as power systems, building communities, etc.) and to optimize system design and lifecycle operation and maintenance management decisions, it is necessary to establish an analysis framework consisting of disaster sites, vulnerability, losses, and decisions. Among them, the disaster site model provides the most basic and core disaster input (such as typhoon wind field) for the entire analysis framework, and then conducts subsequent analysis. The typhoon parameterized model is the most commonly used disaster site model in the engineering community that serves lifecycle analysis and long-term engineering problems. It can efficiently generate a large number of typhoon wind fields (including wind speed and wind direction). The resulting wind field conforms to historical statistical characteristics, but because it is based on simplified atmospheric dynamics equations and statistical laws, its wind field can only capture the laws of large-scale wind circles, and cannot consider the serious impact of micro-scale terrain on near-ground wind fields. Especially for coastal areas with complex terrain, the impact of micro-topography on near-ground wind fields is very serious (can cause a maximum wind speed difference of 2 times). Therefore, it is necessary to consider the impact of microtopography on the near-surface wind field in the typhoon parameterized wind field and obtain a more accurate near-surface wind field as the disaster site input, so as to ensure the reliability of subsequent typhoon-induced losses and decision-making analysis.

[0003] As a vector field, the typhoon wind field includes two parts: wind speed and wind direction. As for wind direction, there is no method in the existing typhoon parameterized wind field research that can simulate the change of typhoon wind direction under the influence of complex micro-topography, which directly leads to the inability of the wind field results obtained by the existing model to provide reliable disaster field input for structures whose wind disaster vulnerability is related to wind direction (such as transmission towers, super high-rise buildings, etc.). As for wind speed, the existing research mainly uses two paths to consider the influence of micro-topography in the wind speed of the parameterized wind field, but there is currently no simulation method that can efficiently consider the influence of actual micro-topography. The first path is to construct a more complex parameterized wind field model, that is, to establish a three-dimensional coordinate system that changes with the height of the terrain, so as to directly consider the influence of micro-topography when solving the dynamic equation. Although the result can better reflect the overall influence of the micro-topography, the calculation is time-consuming and the applicable area is small, which makes this method difficult to apply in long-term engineering problems; the second path is the reduction coefficient method, which is also the most commonly used method at present. By constructing the relationship between the wind speed difference with and without considering the influence of micro-topography and the terrain parameters (such as mountain height and slope, etc.), the wind field results are directly post-processed. Although the correlation is taken into account, the premise assumption of this method does not conform to the actual situation: this method assumes that the mountain is an independent peak (that is, a mountain protruding in the middle of a large plain with no mountains next to it). Or there is only one lower mountain), under this assumption, the movement of airflow around and on the mountain is greatly simplified, making it easy to obtain the desired correlation. However, the actual situation does not meet this assumption, that is, the mountain body is continuously undulating, and there are complex micro-topography states such as mountains, hills, passes and plains. At this time, the change in wind speed is difficult to quantify by simple statistical methods. At the same time, the wind speed reduction coefficient in the existing methods is often related to the wind direction, that is, the reduction coefficient value is different for different wind directions. Conversely, reliable wind direction is a prerequisite for obtaining a reliable reduction coefficient. However, as mentioned above, there is no model in the existing technology that can obtain the wind direction field under the influence of complex micro-topography. Therefore, the above two deficiencies directly lead to a reduction in the reliability of the results obtained by the reduction coefficient method. Summary of the invention

[0004] In order to overcome the deficiencies of the above technologies, the present invention provides a method for correcting the complex micro-topography effects on a typhoon parameterized wind field.

[0005] Terminology explanation:

[0006] RWM:Radius to the Maximum Wind speed, maximum wind speed radius.

[0007] WRF: Weather Research and Forecasting model, mesoscale numerical weather forecast model.

[0008] Dynamic downscaling method: It is to embed a low-resolution global climate model into a high-resolution regional climate model, use the global model to provide initial and boundary conditions for the regional climate model, and obtain high-resolution prediction information describing regional climate characteristics.

[0009] The technical solution adopted by the present invention to overcome the technical problems is:

[0010] A method for correcting the complex micro-topography effect on a typhoon parameterized wind field comprises the following steps:

[0011] S1. Generate a data set, the data set at least including:

[0012] Wind field data considering micro-topography effects obtained from numerical model simulation;

[0013] Wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography;

[0014] Geographical environmental data representing micro-topography;

[0015] S2. Construct micro-topography correction model:

[0016] Processing the training data as a micro-topography correction model;

[0017] A micro-topography correction model is constructed and trained based on a deep learning network. The spatial characteristics of the micro-topography and wind field as well as the coupling relationship between the micro-topography and wind field are captured through the micro-topography correction model.

[0018] S3. Input the wind field data and geographic environment data without considering the influence of micro-topography into the trained micro-topography correction model, so as to obtain a parameterized wind field considering the influence of micro-topography.

[0019] Furthermore, in step S1, the wind field data considering the micro-topography effect obtained by numerical model simulation is obtained as follows:

[0020] After at least the mesoscale typhoon wind field data output by the global climate model and the geographical environment data and ocean and atmospheric condition data of the target area are input into the regional scale numerical model as boundary conditions, the high-resolution typhoon wind field data are simulated by the dynamic downscaling method, that is, the wind field data considering the influence of micro-topography, wherein the resolution of the mesoscale typhoon wind field is 20km~100km, and the resolution of the high-resolution typhoon wind field is 100m~1km. The high-resolution typhoon wind field data includes the wind speed of the near-ground wind field and near-surface wind direction , used as the target value for micro-topography correction model training.

[0021] Furthermore, in step S1, the wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography is obtained as follows:

[0022] The parameterized wind field model is based on simplified atmospheric dynamics equations and statistical equations obtained from the analysis of historical typhoon data. By inputting several typhoon characteristic parameters, the corresponding parameterized wind field results are obtained, that is, wind field data without considering the influence of micro-topography. The parameterized wind field results include parameterized wind field wind speed and parameterized wind field direction , used as the predicted value of the micro-topography correction model.

[0023] Furthermore, the wind speed of the parameterized wind field and parameterized wind field direction The calculation is completed by the following formula:

[0024] (1)

[0025] (2)

[0026] (3)

[0027] (4)

[0028] In formulas (1)~(4): Indicates the wind speed at gradient height; Indicates the gradient height wind direction; Indicates the wind speed near the ground; Indicates the wind direction near the ground; footer Indicates the location point where wind speed or wind direction needs to be calculated; Indicates the moving speed of the typhoon center; Indicates the angle between the location point and the moving direction of the typhoon center; represents the Coriolis force; Indicates the distance between the location point and the typhoon center; represents the Holland coefficient; Indicates the relative pressure difference at the center of a typhoon; Indicates the air density; Indicates the maximum wind speed radius; Indicates the moving direction of the typhoon center; represents the rough velocity; represents the von Karman constant; represents the gradient height; Indicates the surface roughness; and These are all parameters of the parameterized wind field model;

[0029] in, , , , The expressions are as follows:

[0030] (5)

[0031] (6)

[0032] (7)

[0033] (8)

[0034] (9)

[0035] In formulas (5)~(9): represents the Earth's rotation speed; Indicates the latitude of the typhoon center; Indicates the central pressure of the typhoon; represents the atmospheric boundary layer height;

[0036] Typhoon center moving speed , the direction of movement of the typhoon center , Typhoon center pressure and the typhoon's center latitude These four parameters are extracted from the wind field results simulated by the numerical model. Apply calculation formulas (1) to (9) to obtain the parameterized wind speed in the corresponding target area: and parameterized wind field direction .

[0037] Furthermore, in step S1, the geographical environment data characterizing the micro-topography specifically includes:

[0038] Digital elevation model of the target area ,slope , Slope Surface roughness , used as explanatory variables in the microtopography correction model.

[0039] Furthermore, in step S2, the training data used as the micro-topography correction model includes the predicted values , , , , , , and the target value and ;

[0040] Processing the training data as a micro-topography correction model includes:

[0041] First, the geographic coordinates of the target value are reprojected to ensure that the target value and the predicted value belong to the same calculation grid in the same projection coordinate system;

[0042] Secondly, the predicted value data is standardized, that is, by calculating the maximum and minimum values ​​of each variable, the numerical range of each variable is limited to [0,1] to eliminate the impact of the difference in the numerical range of the data itself on the training of the micro-topography correction model.

[0043] Furthermore, a set of weights is assigned to each predicted value to form the input Used to build a deep learning based network, where and Represent the width and height of the image respectively, and 6 is the number of image channels, that is, 6 predictor variables;

[0044] (10)

[0045] In formula (10), , , , , , They are , , , , , The six weight parameters are all two-dimensional vectors, which represent the different weights of the six predictor variables when predicting wind speed and wind direction.

[0046] Furthermore, in step S2, the main body of the micro-terrain correction model constructed based on the deep learning network is a fully convolutional neural network, including two sets of encoders and decoders, namely a wind speed encoder and a wind speed decoder for wind speed and a wind direction encoder and a wind direction decoder for wind direction;

[0047] The wind speed encoder consists of five convolutional layers and two pooling layers. For the input image data of wind speed, the convolution kernel of the convolution layer extracts key features from each channel dimension of the input, and then the pooling operation of the pooling layer is used to stabilize and compress the feature map. The hidden tensor is obtained through multi-layer convolution and pooling operations. ;

[0048] The wind speed decoder consists of two deconvolution layers and six convolution layers to perform a hidden tensor Use deconvolution operation to further extract channel features and gradually enlarge the image to match the target value;

[0049] The wind direction encoder consists of six convolutional layers and two pooling layers. For the input image data of wind direction, the convolution kernel of the convolutional layer extracts key features from each channel dimension of the input, and then the pooling operation of the pooling layer is used to stabilize and compress the feature map. The hidden tensor is obtained through multi-layer convolution and pooling operations. ;

[0050] The wind direction decoder consists of two deconvolution layers and seven convolution layers to perform a hidden tensor Use deconvolution operation to further extract channel features and gradually enlarge the image to match the target value;

[0051] Through the fully convolutional neural network, we finally achieved the capture of two layers of features: one is the spatial characteristics of the micro-topography and the wind field itself; the other is the coupling relationship between the wind field and the micro-topography, that is, the changing law of the wind field in complex micro-topography, and this law is presented in the form of two-dimensional space.

[0052] Furthermore, the wind speed encoder specifically includes: for the input image data of the wind speed, first convolution is performed through the first convolution layer and the second convolution layer; secondly, average pooling is performed through the first pooling layer; then convolution is performed through the third convolution layer and the fourth convolution layer; then average pooling is performed through the second pooling layer; finally, convolution is performed through the fifth convolution layer to obtain a hidden tensor ;

[0053] The wind speed decoder specifically includes: First, the channel features are extracted and the image is enlarged through the first deconvolution layer; secondly, the channel features are extracted through the first convolution layer; then the high-dimensional feature map obtained by the wind speed encoder in the encoding process is introduced into the wind speed decoder using the residual connection method; then convolution is performed through the second and third convolution layers; then the image is enlarged through the second deconvolution layer; then convolution is performed through the fourth convolution layer and residual connection is performed in the channel dimension; finally, the image features are further extracted through the fifth and sixth convolution layers, and finally the output tensor, i.e., the wind speed prediction result, is obtained;

[0054] The wind direction encoder specifically includes: for the input image data of wind direction, first convolution is performed through the first convolution layer and the second convolution layer; secondly, maximum pooling is performed through the first pooling layer; then convolution is performed through the third convolution layer and the fourth convolution layer; then maximum pooling is performed through the second pooling layer; finally, convolution is performed through the fifth convolution layer and the sixth convolution layer to obtain the hidden tensor. ;

[0055] The wind direction decoder specifically includes: First, the channel features are extracted and the image is enlarged through the first deconvolution layer; secondly, the channel features are extracted through the first convolution layer and the second convolution layer; then the high-dimensional feature map obtained by the wind direction encoder in the encoding process is introduced into the wind direction decoder using the residual connection method; then convolution is performed through the third convolution layer and the fourth convolution layer; then the image is enlarged through the second deconvolution layer; finally, the image features are further extracted through the fifth convolution layer, the sixth convolution layer and the seventh convolution layer, and finally the output tensor, i.e., the wind direction prediction result, is obtained;

[0056] Among them, the relationship between the change of image size before and after pooling and the pooling layer parameters conforms to formula (11):

[0057] (11)

[0058] In formula (11), and Respectively represent the input image size and output image size. When calculating the image width, and Respectively represent the width of the input image and the width of the output image. When calculating the height of the image, and Represent the height of the input image and the height of the output image respectively; Indicates floor operation; Indicates the stride size; Indicates the pooling window size;

[0059] The relationship between the change in image size before and after deconvolution and the deconvolution layer parameters conforms to formula (12):

[0060] (12)

[0061] In formula (12), and Respectively represent the input image size and output image size; when calculating the image width, and Respectively represent the width of the input image and the width of the output image. When calculating the height of the image, and Represent the height of the input image and the height of the output image respectively; Indicates the stride size; Indicates the filling quantity; Represents the convolution kernel size.

[0062] Furthermore, the output of the micro-topography correction model is the predicted value ,in, and Represent the width and height of the image respectively, 2 is the number of output image channels, i.e. wind speed and wind direction ;

[0063] The following loss function is designed when training the micro-terrain correction model. By applying loss weights to the areas where the wind speed in the wind speed prediction results of the target area is in the top 20% of the wind speed distribution from high to low, the micro-terrain correction model can focus on and predict the area.

[0064] (13)

[0065] (14)

[0066] In formulas (13) and (14), and Respectively represent the loss functions used in calculating wind speed prediction and wind direction prediction; For the image pixels; is the number of pixels in the image; and They are The predicted wind speed value and target wind speed value at the point; This is the maximum wind speed in this image; and They are Predicted wind direction and target wind direction value at the point; and They are the optimization parameters for wind speed prediction and wind direction prediction respectively;

[0067] Finally, the micro-topography correction model is trained based on the constructed deep learning network to obtain the micro-topography correction model parameters with the minimum error. , to achieve the best result prediction;

[0068] (15)

[0069] In formula (15), the parameters Refers to the parameters of the network finally obtained through training in the deep learning network; Refers to the loss function expressed by formulas (13) and (14); Refers to the target value used in micro-terrain correction model training, that is, and ; Refers to the output of the micro-topography correction model, that is, and argmin refers to the optimization process, that is, the process of continuous training of the micro-terrain correction model. Its optimization goal is to find a set of parameters for the micro-terrain correction model. , so that through The loss obtained is the lowest, and finally a trained micro-terrain correction model is obtained.

[0070] The beneficial effects of the present invention are:

[0071] Through the complex micro-topography correction method for typhoon parameterized wind field described in the present invention, the complex micro-topography effect is efficiently considered in the typhoon parameterized wind field, and the spatial characteristics of micro-topography and wind field and the coupling relationship between micro-topography and wind field are captured by deep learning methods, so as to provide a more reliable typhoon parameterized wind field for engineering applications. With respect to wind speed, it makes up for the shortcomings of low calculation efficiency and over-simplification of complex micro-topography conditions in reality as independent peaks in the existing technology, and can efficiently consider the changes in wind speed in complex micro-topography; with respect to wind direction, it fills the gap that there is no method to realize micro-topography correction of wind direction at present. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a schematic diagram of the principle structure of a method for correcting the complex micro-topography impact on a typhoon parameterized wind field according to an embodiment of the present invention.

[0073] Figure 2 This is a structural principle diagram of the main fully convolutional neural network of the micro-topography correction model described in an embodiment of the present invention.

[0074] Figure 3 This is a comparison chart of the output results of the micro-terrain correction model constructed in an embodiment of the present invention and the target value on the validation set.

[0075] Figure 4 This is a comparison diagram of the wind field predicted by the micro-topography correction model and the target wind field in a certain area shown in an embodiment of the present invention.

[0076] Figure 5 It is a schematic diagram comparing the wind fields obtained without micro-topography correction and after complex micro-topography correction using the method described in this embodiment. DETAILED DESCRIPTION

[0077] In order to facilitate those skilled in the art to better understand the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following is only exemplary and does not limit the protection scope of the present invention.

[0078] The present invention discloses a method for correcting the influence of complex micro-topography on a typhoon parameterized wind field, comprising the following steps:

[0079] S1. Generate a data set, the data set at least including:

[0080] Wind field data considering micro-topography effects obtained from numerical model simulation;

[0081] Wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography;

[0082] Geographical environmental data representing micro-topography;

[0083] S2. Construct micro-topography correction model:

[0084] Processing the training data as a micro-topography correction model;

[0085] A micro-topography correction model is constructed and trained based on a deep learning network. The spatial characteristics of the micro-topography and wind field as well as the coupling relationship between the micro-topography and wind field are captured through the micro-topography correction model.

[0086] S3. Input the wind field data and geographic environment data without considering the influence of micro-topography into the trained micro-topography correction model, so as to obtain a parameterized wind field considering the influence of micro-topography.

[0087] In order to better understand the above technical solution, the exemplary embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. These are only exemplary embodiments of the present invention. However, it should be understood that the present invention can also be implemented in various forms and is not limited to the embodiments described herein. These embodiments are intended to enable those skilled in the art to understand the present invention more clearly and thoroughly.

[0088] The present embodiment provides a method for correcting the complex micro-topography effect on the typhoon parameterized wind field. Figure 1 As shown, the following steps are included:

[0089] S1. Generate a data set, which includes at least three parts:

[0090] 1) Wind field data considering micro-topography effects obtained by numerical model simulation;

[0091] 2) Wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography;

[0092] 3) Geographical environment data characterizing microtopography.

[0093] In this embodiment, the wind field data considering the micro-topography effect obtained by numerical model simulation is obtained as follows:

[0094] After at least the mesoscale typhoon wind field data output by the global climate model and the geographical environment data and ocean and atmospheric condition data of the target area are input into the regional scale numerical model as boundary conditions, the high-resolution typhoon wind field data are simulated by the dynamic downscaling method, that is, the wind field data considering the influence of micro-topography, wherein the resolution of the mesoscale typhoon wind field is 20km~100km, and the resolution of the high-resolution typhoon wind field is 100m~1km. The high-resolution typhoon wind field data includes the wind speed of the near-ground wind field and near-surface wind direction , used as the target value for micro-topography correction model training.

[0095] Existing numerical models in the meteorological field can be used, such as the mesoscale numerical meteorological forecast model WRF, and the mesoscale typhoon wind field data published by the meteorological center and the geographical environment data and ocean and atmospheric condition data of the target area can be input. By setting a three-layer nested calculation grid dynamic downscaling method, numerical simulation wind field results with horizontal resolutions of 4km, 1.33km and 443m can be gradually simulated. Among them, the result with a resolution of 443m belongs to the high-resolution typhoon wind field data described in this embodiment, and its near-ground wind field wind speed is and near-surface wind direction The influence of complex micro-topography on the typhoon near-surface wind field can be well considered. and Used as the target value for micro-terrain correction model training.

[0096] In this embodiment, the wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography is obtained as follows:

[0097] The parameterized wind field model is based on simplified atmospheric dynamics equations and statistical equations obtained from the analysis of historical typhoon data. By inputting several typhoon characteristic parameters, the corresponding parameterized wind field results are obtained, that is, wind field data without considering the influence of micro-topography. The parameterized wind field results include parameterized wind field wind speed and parameterized wind field direction , used as the predicted value of the micro-topography correction model.

[0098] Furthermore, the wind speed of the parameterized wind field and parameterized wind field direction The calculation is completed by the following formula: (1)

[0099] (2)

[0100] (3)

[0101] (4)

[0102] In formulas (1)~(4): Indicates the wind speed at gradient height; Indicates the gradient height wind direction; Indicates the wind speed near the ground; Indicates wind direction near the ground; footer Indicates the location point where wind speed or wind direction needs to be calculated; Indicates the moving speed of the typhoon center; Indicates the angle between the location point and the moving direction of the typhoon center; represents the Coriolis force; Indicates the distance between the location point and the typhoon center; represents the Holland coefficient; Indicates the relative pressure difference at the center of a typhoon; Indicates the air density; Indicates the maximum wind speed radius; Indicates the moving direction of the typhoon center; represents the rough velocity; represents the von Karman constant; represents the gradient height; Indicates the surface roughness; and These are all parameters of the parameterized wind field model;

[0103] in, , , , The expressions are as follows:

[0104] (5)

[0105] (6)

[0106] (7)

[0107] (8)

[0108] (9)

[0109] In formulas (5)~(9): represents the Earth's rotation speed; Indicates the latitude of the typhoon center; Indicates the central pressure of the typhoon; represents the atmospheric boundary layer height;

[0110] Typhoon center moving speed , the direction of movement of the typhoon center , Typhoon center pressure and the typhoon's center latitude These four parameters are extracted from the wind field results simulated by the numerical model. Apply calculation formulas (1) to (9) to obtain the parameterized wind speed in the corresponding target area: and parameterized wind field direction .

[0111] In this embodiment, the geographical environment data representing the micro-topography specifically includes:

[0112] Digital elevation model of the target area ,slope , Slope Surface roughness , which is used as the explanatory variable of the micro-topography correction model. Inputting geographic environment data in the construction of the micro-topography correction model can explicitly express geographic features and accelerate the learning efficiency of the model. Specifically, using ArcGIS software, input the digital elevation model of the target area (i.e. DEM, Digital Elevation Model) directly calculates the slope of the target area and slope aspect , obtain the surface cover type of the target area through public data and obtain the surface roughness according to the public corresponding rules .

[0113] S2. Construct a micro-topography correction model, including:

[0114] S2.1. Processing the training data as a micro-topography correction model.

[0115] In this embodiment, the training data of the micro-topography correction model includes the predicted values , , , , , , and the target value and .

[0116] Processing the training data as a micro-topography correction model includes:

[0117] First, the geographic coordinates of the target value are reprojected to ensure that the target value and the predicted value belong to the same calculation grid in the same projection coordinate system;

[0118] Secondly, the predicted value data is standardized, that is, by calculating the maximum and minimum values ​​of each variable, the numerical range of each variable is limited to [0,1] to eliminate the impact of the difference in the numerical range of the data itself on the training of the micro-topography correction model.

[0119] S2.2. Construct and train a micro-topography correction model based on a deep learning network to capture the spatial characteristics of the micro-topography and wind field as well as the coupling relationship between the micro-topography and wind field.

[0120] One of the key points in implementing the method for correcting the complex micro-topography effect on the typhoon parameterized wind field described in this embodiment is how to deal with the relative importance between the predicted values ​​to ensure that the micro-topography correction model pays attention to the most important predicted variables for obtaining the optimal results during the training process. In order to ensure that the most important predicted variables are given the greatest attention, a set of weights is assigned to each predicted value to form an input Used to build a deep learning based network, where and They represent the width and height of the image respectively, and 6 is the number of image channels, that is, 6 predictor variables.

[0121] (10)

[0122] In formula (10), , , , , , They are , , , , , The six weight parameters are all two-dimensional vectors, which represent the different weights of the six predictor variables when predicting wind speed and wind direction.

[0123] Specifically, the main body of the micro-terrain correction model constructed based on the deep learning network is a fully convolutional neural network. In this embodiment, considering the differences in wind speed and wind direction, two sets of encoders and decoders are included, such as Figure 2 As shown, there are a wind speed encoder and a wind speed decoder for wind speed, and a wind direction encoder and a wind direction decoder for wind direction, respectively.

[0124] The wind speed encoder consists of five convolutional layers and two pooling layers. For the input image data of wind speed, the convolution kernel of the convolution layer extracts key features from each channel dimension of the input, and then the pooling operation of the pooling layer is used to stabilize and compress the feature map. The hidden tensor is obtained through multi-layer convolution and pooling operations. .

[0125] The wind speed decoder consists of two deconvolution layers and six convolution layers to perform a hidden tensor Deconvolution operations are used to further extract channel features and gradually enlarge the image to match the target value.

[0126] The wind direction encoder consists of six convolutional layers and two pooling layers. For the input image data of wind direction, the convolution kernel of the convolutional layer extracts key features from each channel dimension of the input, and then the pooling operation of the pooling layer is used to stabilize and compress the feature map. The hidden tensor is obtained through multi-layer convolution and pooling operations. .

[0127] The wind direction decoder consists of two deconvolution layers and seven convolution layers to perform a hidden tensor Deconvolution operations are used to further extract channel features and gradually enlarge the image to match the target value.

[0128] Among them, the relationship between the change of image size before and after pooling and the pooling layer parameters conforms to formula (11):

[0129] (11)

[0130] In formula (11), and Respectively represent the input image size and output image size. When calculating the image width, and Respectively represent the width of the input image and the width of the output image. When calculating the height of the image, and Represent the height of the input image and the height of the output image respectively; Indicates floor operation; Indicates the stride size; Indicates the pooling window size;

[0131] The relationship between the change in image size before and after deconvolution and the deconvolution layer parameters conforms to formula (12):

[0132] (12)

[0133] In formula (12), and Respectively represent the input image size and output image size; when calculating the image width, and Respectively represent the width of the input image and the width of the output image. When calculating the height of the image, and Represent the height of the input image and the height of the output image respectively; Indicates the stride size; Indicates the filling quantity; Represents the convolution kernel size.

[0134] In this embodiment, Figure 2As shown in the figure, the wind speed encoder specifically includes: for the wind speed input image data of size 50×50×6, first convolution is performed through the first convolution layer and the second convolution layer to obtain a tensor of size 50×50×32; secondly, through the first pooling layer, the average pooling method is used to obtain a tensor of size 25×25×32; then through the third convolution layer and the fourth convolution layer, the tensor of size 25×25×128 is obtained by convolution; then through the second pooling layer, the average pooling method is used to obtain a tensor of size 12×12×128; finally, the fifth convolution layer is used to perform a convolution operation to obtain a hidden tensor of size 12×12×256 .

[0135] The wind speed decoder specifically includes: for an input tensor of size 12×12×256 (i.e., a hidden tensor ), firstly, the channel features are extracted through the first deconvolution layer and the image is enlarged to obtain a tensor of size 25×25×256; secondly, the channel features are extracted through the first convolution layer to obtain a tensor of size 25×25×128; then, the high-dimensional feature map obtained by the wind speed encoder in the encoding process is introduced into the wind speed decoder using the residual connection method, that is, the tensor of size 25×25×128 in the wind speed encoder is connected with the tensor of size 25×25×128 obtained in the previous step at the channel scale to obtain a tensor of size 25×25×256; then, the second convolution layer is used to extract the channel features and the image is enlarged to obtain a tensor of size 25×25×256. The image is enlarged by the second deconvolution layer to obtain a tensor of 50×50×64. The image is then enlarged by the fourth convolution layer to obtain a tensor of 50×50×32. On the basis of this tensor, the tensor of 50×50×32 in the wind speed encoder is connected in the channel dimension through the residual connection method to obtain a tensor of 50×50×64. Finally, the image features are further extracted through the fifth and sixth convolution layers to obtain an output tensor of 50×50×1, which is the wind speed prediction result.

[0136] The wind direction encoder specifically includes: for the wind direction input image data of size 50×50×6, first convolution is performed through the first convolution layer and the second convolution layer to obtain a tensor of size 50×50×32; secondly, through the first pooling layer, the maximum pooling method is used to obtain a tensor of size 25×25×32; then through the third convolution layer and the fourth convolution layer, convolution is performed to obtain a tensor of size 25×25×128; then through the second pooling layer, the maximum pooling method is used to obtain a tensor of size 12×12×128; finally, through the fifth convolution layer and the sixth convolution layer, convolution operation is performed to obtain a hidden tensor of size 12×12×512 .

[0137] The wind direction decoder specifically includes: for the input tensor of size 12×12×512 (i.e., the hidden tensor ), firstly, the channel features are extracted through the first deconvolution layer and the image is enlarged to obtain a tensor of size 25×25×512; secondly, the channel features are extracted through the first convolution layer and the second convolution layer to obtain a tensor of size 25×25×128; then, the high-dimensional feature map obtained by the wind direction encoder in the encoding process is introduced into the wind direction decoder using the residual connection method, that is, the tensor of size 25×25×128 in the wind direction encoder is combined with the tensor of size 25×25 ×128 tensors are connected at the channel scale to obtain a tensor of size 25×25×256; it is then passed through the third and fourth convolutional layers to obtain a tensor of size 25×25×64; the image is then enlarged through the second deconvolution layer to obtain a tensor of size 50×50×64; finally, the image features are further extracted through the fifth, sixth, and seventh convolutional layers, and finally an output tensor of size 50×50×1 is obtained, which is the wind direction prediction result.

[0138] Through the fully convolutional neural network described in this embodiment, it is finally possible to capture two layers of features:

[0139] One is the spatial characteristics of the micro-topography and wind field itself. This is because the convolutional neural network itself captures spatial (that is, two-dimensional) features through convolution kernels and extracts the most important features through pooling layers.

[0140] The second is the coupling relationship between the wind field and micro-topography, that is, the changing pattern of the wind field in complex micro-topography, and this pattern is presented in the form of two-dimensional space. The reasons are: (1) The numerical model is used to simulate the target value, and the numerical model itself has the function of capturing this layer of relationship, but because the calculation speed of the numerical model is too slow, it is limited in practical application; (2) In addition to wind speed and wind direction, the input parameters also include terrain parameters, mainly digital elevation model ,slope , Slope Surface roughness These four parameters directly characterize the micro-topography features and help the neural network capture the coupling relationship.

[0141] Furthermore, the output of the micro-topography correction model is the predicted value ,in, and Represent the width and height of the image respectively, 2 is the number of output image channels, i.e. wind speed and wind direction .

[0142] When training the micro-terrain correction model, the choice of loss function is very critical. Therefore, this embodiment designs the following loss function, which applies loss weights to the areas where the wind speed in the wind speed prediction results of the target area is in the top 20% of the wind speed distribution from high to low, so as to achieve the micro-terrain correction model's attention and prediction for the area.

[0143] (13)

[0144] (14)

[0145] In formulas (13) and (14), and Respectively represent the loss functions used in calculating wind speed prediction and wind direction prediction; For the image pixels; is the number of pixels in the image; and They are The predicted wind speed value and target wind speed value at the point; This is the maximum wind speed in this image; and They are Predicted wind direction and target wind direction value at the point; and They are the optimization parameters for wind speed prediction and wind direction prediction respectively.

[0146] Finally, the micro-terrain correction model is trained based on the constructed deep learning network. In this embodiment, the training process of the micro-terrain correction model is carried out for a total of 200 training cycles (epochs), and the batch size (batchsize) is set to 50. After each training, the loss value is calculated by back propagation, and the micro-terrain correction model parameters are updated; specifically, the PyTorch framework in python is used for model construction and training; the optimizer uses the Adam algorithm, the learning rate is set to 0.0001, and the weight decay coefficient is 0.00001. In order to ensure the optimal performance of the micro-terrain correction model and prevent overfitting, the learning rate is dynamically adjusted when the loss value reaches a plateau. In addition, the "early stopping" strategy is adopted. When the loss value on the validation set no longer decreases for 20 consecutive training cycles, the training is stopped to reduce the risk of overfitting.

[0147] After training the micro-topography correction model, the micro-topography correction model parameters with the minimum error are obtained. , to achieve the best result prediction.

[0148] (15)

[0149] In formula (15), the parameters Refers to the parameters of the network finally obtained through training in the deep learning network; Refers to the loss function expressed by formulas (13) and (14); Refers to the target value used in micro-terrain correction model training, that is, and ; Refers to the output of the micro-topography correction model, that is, and ; Due to the target value and Output results of micro-terrain correction model training and There is an error, so it is calculated through the loss function and The error and and The error is then optimized; argmin refers to the optimization process, that is, the process of continuous training of the micro-terrain correction model. Its optimization goal is to find a set of parameters for the micro-terrain correction model. , so that through The loss obtained is the lowest, and finally a trained micro-terrain correction model is obtained.

[0150] S3. Input the wind field data and geographic environment data without considering the influence of micro-topography into the trained micro-topography correction model, so as to obtain a parameterized wind field considering the influence of micro-topography.

[0151] Step S3 is to apply the micro-topography correction model. Specifically, the wind field data (parameterized wind field wind speed) obtained by simulating the parameterized wind field model without considering the micro-topography effect is converted into the wind field data (parameterized wind field wind speed) without considering the micro-topography effect. and parameterized wind field direction ) and geographic environment data (digital elevation model) used to characterize microtopography ,slope , Slope Surface roughness ), which is input into the micro-topography correction model trained in step S2, and the parameterized wind field considering the influence of micro-topography can be directly obtained, that is, the output and .

[0152] The following is an experiment to illustrate the effect of the method for correcting the complex micro-topography influence on a typhoon parameterized wind field described in this embodiment.

[0153] like Figure 3As shown, it is a comparison chart of the output results of the micro-terrain correction model constructed in this embodiment and the target value on the validation set. The x-axis is the prediction result of the micro-terrain correction model, and the y-axis is the target value. Since the output result of the micro-terrain correction model is a two-dimensional image of size 50×50 (width and height of the image), in order to reflect the correlation between the output result of the micro-terrain correction model and the target value, the mean and standard deviation of the output result of the micro-terrain correction model of each two-dimensional image in the validation set are calculated (that is, for each 50×50 two-dimensional image, the mean and standard deviation of the prediction results of its 2500 grid points are calculated; finally, each image in the validation set has 4 values ​​corresponding to the output results of the micro-terrain correction model, namely the wind speed mean, wind speed standard deviation, wind direction mean, and wind direction standard deviation). The mean is used to characterize the average level of the two-dimensional image prediction results, and the standard deviation is used to characterize whether the fluctuations of wind speed or wind direction can be well grasped in the two-dimensional image prediction results (because the essence of the micro-terrain correction model described in this embodiment is to care about fluctuations or changes caused by terrain). Therefore, for Figure 3 Figure (a) in the figure describes the relationship between the overall situation of wind speed forecast and the target value in a 50×50 image; Figure 3 Figure (c) in the figure describes the relationship between the overall situation of wind direction prediction and the target value in a 50×50 image; Figure 3 Figures (b) and (d) in the figure describe whether the fluctuations or changes caused by terrain are well captured in the 50×50 image. In addition, wind speed and wind direction can be simply regarded as two variables, and the evaluation paths for the two of them are similar. Specifically, Figure 3 Figure (a) in the figure describes the prediction of the mean wind speed in each sample by the micro-topography correction model, which represents the overall average level of wind speed in the target area; Figure 3 Figure (b) in the figure describes the prediction of the standard deviation of wind speed in each sample by the micro-topography correction model, which represents the simulation ability of wind speed fluctuations in the target area; Figure 3 Figure (c) in the figure describes the mean wind direction prediction of the micro-topography correction model in each sample, representing the overall level of wind direction in the target area; Figure 3 Figure (d) in the figure describes the prediction of the standard deviation of wind direction in each sample by the micro-topography correction model, which represents the simulation ability of wind direction fluctuations in the target area. Figure 3 It can be seen from the four sub-figures (a), (b), (c) and (d) that the prediction effect of the micro-topography correction model described in this embodiment is good, and the predicted value and the target value are closely distributed around y=x.

[0154] Table 1 Prediction error of micro-terrain correction model on the validation set

[0155]

[0156] Table 1 shows the prediction error data of the micro-terrain correction model on the validation set. Since the analysis of the disaster site is also divided into different spatial scales (such as point-structure monomer; local-building community in a community; regional-provincial power transmission system) according to the different spatial scales of specific problems in engineering applications, in order to verify that the micro-terrain correction model obtained by the method described in this embodiment can achieve good results at different spatial scales, the absolute error between the predicted value and the target value of the micro-terrain correction model at different scales is shown in Table 1. As shown in Table 1, the micro-terrain correction model described in this embodiment can achieve good prediction results at all spatial scales.

[0157] Figure 4 The comparison between the wind field predicted by the micro-topography correction model and the target wind field in a certain area is shown. The white part in the figure is the sea surface, so no prediction is made. Figure 4 Figure (a) and Figure 4 By comparing with Figure (b) in , it can be found that the constructed micro-topography correction model can better capture the wind field under the influence of micro-topography. In order to further verify the effect and superiority of the method described in this embodiment, the micro-topography correction model described in this embodiment is applied to a typhoon that occurred in a certain area in 2020 to display the results.

[0158] Figure 5 A schematic diagram comparing the wind fields obtained without micro-topography correction and after complex micro-topography correction using the method described in this embodiment is shown. Figure 5 Figure (a) is a schematic diagram of the parameterized wind field without micro-topography correction, which is the result obtained directly through the parameterized wind field model; Figure 4 Figure (b) is a schematic diagram of the wind field obtained after the complex micro-topography correction is performed by the method described in this embodiment. By comparison, it can be found that the traditional parameterized wind field model cannot consider the impact of micro-topography on the near-ground wind field, resulting in a very smooth wind field result that cannot reflect the specific surface conditions; by applying the micro-topography correction model established in the method described in this embodiment, the impact of complex surface micro-topography can be captured, and a wind field result that is more in line with the actual situation can be obtained.

[0159] Through the method for correcting the complex micro-topography effect on the typhoon parameterized wind field described in this embodiment, it is possible to efficiently consider the complex micro-topography effect in the typhoon parameterized wind field, and to capture the spatial characteristics of the micro-topography and the wind field and the coupling relationship between the micro-topography and the wind field through a deep learning method, so as to provide a more reliable typhoon parameterized wind field for engineering applications. With respect to wind speed, it makes up for the shortcomings of the existing technology of low calculation efficiency and over-simplification of the complex micro-topography conditions in reality as independent peaks, and can efficiently consider the changes in wind speed in complex micro-topography; with respect to wind direction, it fills the gap that there is currently no method to achieve micro-topography correction of wind direction.

[0160] The above only describes the basic principles and preferred embodiments of the present invention. Those skilled in the art may make many changes and improvements based on the above description, and these changes and improvements should fall within the protection scope of the present invention.

Claims

1. A method for correcting the complex micro-topography effect on typhoon parameterized wind field, characterized in that: The steps include: S1. Generate a data set, the data set at least including: Wind field data considering micro-topography effects obtained from numerical model simulation; Wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography; Geographical environmental data representing micro-topography; The wind field data considering the influence of micro-topography obtained by numerical model simulation is obtained as follows: After at least the mesoscale typhoon wind field data output by the global climate model and the geographical environment data and ocean and atmospheric condition data of the target area are input into the regional scale numerical model as boundary conditions, the high-resolution typhoon wind field data are simulated by the dynamic downscaling method, that is, the wind field data considering the influence of micro-topography, wherein the resolution of the mesoscale typhoon wind field is 20km~100km, and the resolution of the high-resolution typhoon wind field is 100m~1km. The high-resolution typhoon wind field data includes the wind speed of the near-ground wind field and near-surface wind direction , used as the target value for micro-topography correction model training; The wind field data obtained by simulating the parameterized wind field model without considering the influence of micro-topography is obtained as follows: The parameterized wind field model is based on simplified atmospheric dynamics equations and statistical equations obtained from the analysis of historical typhoon data. By inputting several typhoon characteristic parameters, the corresponding parameterized wind field results are obtained, that is, wind field data without considering the influence of micro-topography. The parameterized wind field results include parameterized wind field wind speed and parameterized wind field direction , used as the predicted value of the micro-topography correction model; S2. Construct micro-topography correction model: Processing the training data as a micro-topography correction model; A micro-topography correction model is constructed and trained based on a deep learning network. The spatial characteristics of the micro-topography and wind field as well as the coupling relationship between the micro-topography and wind field are captured through the micro-topography correction model. Among them, the main body of the micro-terrain correction model built based on the deep learning network is a fully convolutional neural network, which includes two sets of encoder-decoders, namely, a wind speed encoder and a wind speed decoder for wind speed and a wind direction encoder and a wind direction decoder for wind direction; The wind speed encoder consists of five convolutional layers and two pooling layers. For the input image data of wind speed, the convolution kernel of the convolution layer extracts key features from each channel dimension of the input, and then the pooling operation of the pooling layer is used to stabilize and compress the feature map. The hidden tensor is obtained through multi-layer convolution and pooling operations. ; The wind speed decoder consists of two deconvolution layers and six convolution layers to perform a hidden tensor Use deconvolution operation to further extract channel features and gradually enlarge the image to match the target value; The wind direction encoder consists of six convolutional layers and two pooling layers. For the input image data of wind direction, the convolution kernel of the convolutional layer extracts key features from each channel dimension of the input, and then the pooling operation of the pooling layer is used to stabilize and compress the feature map. The hidden tensor is obtained through multi-layer convolution and pooling operations. ; The wind direction decoder consists of two deconvolution layers and seven convolution layers to perform a hidden tensor Use deconvolution operation to further extract channel features and gradually enlarge the image to match the target value; Through the fully convolutional neural network, we finally achieved the capture of two layers of features: one is the spatial characteristics of the micro-topography and the wind field itself; the other is the coupling relationship between the wind field and the micro-topography, that is, the changing law of the wind field in the complex micro-topography, and this law is presented in the form of two-dimensional space; S3. Input the wind field data and geographic environment data without considering the influence of micro-topography into the trained micro-topography correction model, so as to obtain a parameterized wind field considering the influence of micro-topography.

2. The method for correcting the complex micro-topography effect on the typhoon parameterized wind field according to claim 1 is characterized in that: Parameterized wind speed and parameterized wind field direction The calculation is completed by the following formula: (1) （2） （3） （4） In formulas (1)~(4): Indicates the wind speed at gradient height; Indicates the gradient height wind direction; Indicates the wind speed near the ground; Indicates the wind direction near the ground; footer Indicates the location point where wind speed or wind direction needs to be calculated; Indicates the moving speed of the typhoon center; Indicates the angle between the location point and the moving direction of the typhoon center; represents the Coriolis force; Indicates the distance between the location point and the typhoon center; represents the Holland coefficient; Indicates the relative pressure difference at the center of a typhoon; Indicates the air density; Indicates the maximum wind speed radius; Indicates the moving direction of the typhoon center; represents the rough velocity; represents the von Karman constant; represents the gradient height; Indicates the surface roughness; and These are all parameters of the parameterized wind field model; in, , , , The expressions are as follows: （5） （6） （7） （8） （9） In formulas (5)~(9): represents the Earth's rotation speed; Indicates the latitude of the typhoon center; Indicates the central pressure of the typhoon; represents the height of the atmospheric boundary layer; Typhoon center moving speed , the direction of movement of the typhoon center , Typhoon center pressure and the typhoon's center latitude These four parameters are extracted from the wind field results simulated by the numerical model. Apply calculation formulas (1) to (9) to obtain the parameterized wind speed in the corresponding target area: and parameterized wind field direction .

3. The method for correcting the complex micro-topography effect on the typhoon parameterized wind field according to claim 1 is characterized in that: In step S1, the geographical environment data representing the micro-topography specifically includes: Digital elevation model of the target area ,slope , Slope Surface roughness , used as explanatory variables in the microtopography correction model.

4. The method for correcting the complex micro-topography effect on the typhoon parameterized wind field according to claim 3 is characterized in that: In step S2, the training data of the micro-topography correction model includes the predicted value , , , , , , and the target value and ; Processing the training data as a micro-topography correction model includes: First, the geographic coordinates of the target value are reprojected to ensure that the target value and the predicted value belong to the same calculation grid in the same projection coordinate system; Secondly, the predicted value data is standardized, that is, by calculating the maximum and minimum values ​​of each variable, the numerical range of each variable is limited to [0,1] to eliminate the impact of the difference in the numerical range of the data itself on the training of the micro-topography correction model.

5. The method for correcting the complex micro-topography effect on the typhoon parameterized wind field according to claim 4 is characterized in that: Assign a set of weights to each predicted value to form the input Used to build a deep learning based network, where and Represent the width and height of the image respectively, and 6 is the number of image channels, that is, 6 predictor variables; （10） In formula (10), , , , , , They are , , , , , The six weight parameters are all two-dimensional vectors, which represent the different weights of the six predictor variables when predicting wind speed and wind direction.

6. The method for correcting the complex micro-topography effect on the typhoon parameterized wind field according to any one of claims 1 to 5, characterized in that: The wind speed encoder specifically includes: for the input image data of wind speed, first convolution is performed through the first convolution layer and the second convolution layer; secondly, average pooling is performed through the first pooling layer; then convolution is performed through the third convolution layer and the fourth convolution layer; then average pooling is performed through the second pooling layer; finally, convolution is performed through the fifth convolution layer to obtain the hidden tensor ; The wind speed decoder specifically includes: First, the channel features are extracted and the image is enlarged through the first deconvolution layer; secondly, the channel features are extracted through the first convolution layer; then the high-dimensional feature map obtained by the wind speed encoder in the encoding process is introduced into the wind speed decoder using the residual connection method; then convolution is performed through the second and third convolution layers; then the image is enlarged through the second deconvolution layer; then convolution is performed through the fourth convolution layer and residual connection is performed in the channel dimension; finally, the image features are further extracted through the fifth and sixth convolution layers, and finally the output tensor, i.e., the wind speed prediction result, is obtained; The wind direction encoder specifically includes: for the input image data of wind direction, first convolution is performed through the first convolution layer and the second convolution layer; secondly, maximum pooling is performed through the first pooling layer; then convolution is performed through the third convolution layer and the fourth convolution layer; then maximum pooling is performed through the second pooling layer; finally, convolution is performed through the fifth convolution layer and the sixth convolution layer to obtain the hidden tensor. ; The wind direction decoder specifically includes: First, the channel features are extracted and the image is enlarged through the first deconvolution layer; secondly, the channel features are extracted through the first convolution layer and the second convolution layer; then the high-dimensional feature map obtained by the wind direction encoder in the encoding process is introduced into the wind direction decoder using the residual connection method; then convolution is performed through the third convolution layer and the fourth convolution layer; then the image is enlarged through the second deconvolution layer; finally, the image features are further extracted through the fifth convolution layer, the sixth convolution layer and the seventh convolution layer, and finally the output tensor, i.e., the wind direction prediction result, is obtained; Among them, the relationship between the change of image size before and after pooling and the pooling layer parameters conforms to formula (11): （11） In formula (11), and Respectively represent the input image size and output image size. When calculating the image width, and Respectively represent the width of the input image and the width of the output image. When calculating the height of the image, and Represent the height of the input image and the height of the output image respectively; Indicates floor operation; Indicates the stride size; Indicates the pooling window size; The relationship between the change in image size before and after deconvolution and the deconvolution layer parameters conforms to formula (12): （12） In formula (12), and Respectively represent the input image size and output image size; when calculating the image width, and Respectively represent the width of the input image and the width of the output image. When calculating the height of the image, and Represent the height of the input image and the height of the output image respectively; Indicates the stride size; Indicates the filling quantity; Represents the convolution kernel size.

7. The method for correcting the complex micro-topography effect on the typhoon parameterized wind field according to claim 6, characterized in that: The output of the micro-topography correction model is the predicted value ,in, and Represent the width and height of the image respectively, 2 is the number of output image channels, i.e. wind speed and wind direction ; The following loss function is designed when training the micro-terrain correction model. By applying loss weights to the areas where the wind speed in the wind speed prediction results of the target area is in the top 20% of the wind speed distribution from high to low, the micro-terrain correction model can focus on and predict the area. （13） （14） In formulas (13) and (14), and Respectively represent the loss functions used in calculating wind speed prediction and wind direction prediction; For the image pixels; is the number of pixels in the image; and They are The predicted wind speed value and target wind speed value at the point; This is the maximum wind speed in this image; and They are Predicted wind direction and target wind direction value at the point; and They are the optimization parameters for wind speed prediction and wind direction prediction respectively; Finally, the micro-topography correction model is trained based on the constructed deep learning network to obtain the micro-topography correction model parameters with the minimum error. , to achieve the best result prediction; （15） In formula (15), the parameters Refers to the parameters of the network finally obtained through training in the deep learning network; Refers to the loss function expressed by formulas (13) and (14); Refers to the target value used in micro-terrain correction model training, that is, and ; Refers to the output of the micro-topography correction model, that is, and argmin refers to the optimization process, that is, the process of continuous training of the micro-terrain correction model. Its optimization goal is to find a set of parameters for the micro-terrain correction model. , so that through The loss obtained is the lowest, and finally a trained micro-terrain correction model is obtained.

Images