A deep learning-based complex terrain area typhoon wind field downscaling method and system

Abstract

The application relates to a deep learning-based typhoon wind field downscaling method and system for a complex terrain area, a data acquisition and standardization preprocessing module acquires wind field and terrain data and performs standardization preprocessing; a first typhoon wind field grid prediction module inputs the standardized wind field and terrain data into a multiscale wind-terrain space correlation extraction network to generate a first typhoon wind field grid prediction value; a second typhoon wind field grid prediction module inputs a station space position map and a station feature correlation map established according to meteorological observation station data and standardized data into a meteorological observation station space feature extraction meteorological observation station space linkage network to generate a second typhoon wind field grid prediction value; and a fusion prediction final typhoon wind field grid module performs weighted fusion on the generated first and second typhoon wind field grid prediction values to obtain a final typhoon wind field grid prediction value; the typhoon wind field downscaling method improves the accuracy and efficiency of typhoon wind field downscaling in a complex terrain area.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning and meteorological calculation, and relates to a complex terrain area typhoon wind field downscaling method and system based on deep learning. BACKGROUND

[0002] Typhoon wind field is crucial for typhoon mechanism research and typhoon disaster prevention. Traditional global climate models, atmospheric reanalysis products and long-term weather forecast models usually use wind fields with a spatial resolution of 0.25° to 1°, and existing weather scale prediction models provide wind fields with a spatial resolution of kilometers, which is insufficient to simulate or capture the core dynamics and physical properties of typhoon. In this case, it is necessary to downscale the typhoon wind field to quickly generate high-precision and high-spatial-resolution typhoon wind field to support scientific research or business service applications related to typhoon. In addition, high-resolution typhoon wind field can reflect fine local wind conditions during typhoon activity and provide more targeted typhoon disaster prediction and warning services, which can effectively reduce the loss caused by typhoon disasters.

[0003] However, the current traditional downscaling method is limited by the speed and effect of wind field downscaling, especially the accuracy of typhoon wind field downscaling under highly complex terrain conditions, resulting in low quality and availability of typhoon wind field, and unable to obtain accurate downscaling results in time.

[0004] For example, patent CN116595366A introduces a sea surface wind field spatial downscaling method and system based on multi-task learning, which uses a multi-task training dataset to train and optimize a generative adversarial network to achieve sea surface wind field spatial downscaling, and uses buoy data to verify the downscaling result. However, this patent method is only applicable to the downscaling task of sea wind field and cannot be applied to the downscaling task under complex land terrain conditions.

[0005] Patent CN118366046A discloses a wind field downscaling method based on deep learning combined with terrain, which can achieve downscaling prediction from 0.25°x0.25° to 0.1°x0.1° in latitude and longitude resolution. However, the downscaling method of this patent takes ERA5-Land data of the European Center (ECMWF) as observation data for model training, and does not consider the error between ERA5-Land data and real meteorological station observations, resulting in a large error between the downscaling result and the real meteorological station observations.

[0006] Patent CN118761026A discloses a near-surface wind field downscaling method based on deep learning, which considers the complex nonlinear relationship between wind and terrain and trains the network with meteorological station observations as true values, and can realize wind field downscaling in complex terrain conditions. However, the downscaling method of this patent does not consider the wind speed difference under typhoon weather conditions and cannot be applied to typhoon wind field downscaling.

[0007] Patent CN110298115A discloses a power transmission line risk prediction method based on statistical downscaling of typhoon wind field, which can obtain microscale typhoon forecast wind speed field by downscaling mesoscale typhoon forecast wind field. This patent uses a linear regression model as a statistical downscaling method, which cannot consider the complex nonlinear relationship between typhoon wind field and terrain in complex terrain areas.

[0008] In summary, the existing typhoon wind field downscaling methods have the following problems: (1) Some methods are only applicable to specific scenarios, such as the sea surface wind field downscaling method of patent CN116595366A, which cannot be applied to complex landforms; (2) Some methods have defects in data processing, such as patent CN118366046A, which does not consider the error between ERA5-Land data and real observations, resulting in a large deviation between the downscaling results and actual observations; (3) Some methods perform poorly under complex weather conditions, such as patent CN118761026A, which does not consider the wind speed difference under typhoon weather and is not suitable for typhoon wind field downscaling; (4) The linear regression model used in patent CN110298115A cannot handle the nonlinear relationship between complex terrain and typhoon wind field. These problems limit the universality and accuracy of existing methods in typhoon wind field downscaling. SUMMARY

[0009] Therefore, to solve the problem of slow speed and low accuracy of traditional downscaling methods in complex terrain conditions, leading to poor quality and availability of wind field, and difficulty in providing accurate typhoon wind field downscaling results in a timely manner, the present application provides a complex terrain area typhoon wind field downscaling method and system based on deep learning, which adaptively weighs the wind-terrain correlation under different typhoon conditions, different geographical areas, and different terrain conditions through a deep learning network, and eliminates the uncertainty in the downscaling process using meteorological station observation information, improving the accuracy and efficiency of typhoon wind field downscaling in complex terrain areas.

[0010] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0011] A complex terrain area typhoon wind field downscaling method based on deep learning, comprising the following steps:

[0012] S1, data acquisition and standardization preprocessing; obtaining meteorological observation site data, low-resolution typhoon wind field data and high-resolution terrain data , the obtained engineering typhoon model wind field data and terrain data are standardized and preprocessed to generate standardized data;

[0013] S2, first typhoon wind field grid prediction; the standardized wind field and terrain data are input into the multi-scale wind-terrain spatial feature extraction typhoon wind field and terrain network, through the feature high-dimensional projection module , multi-scale high-dimensional feature fusion module FM, first full connection prediction module to generate wind speed and wind direction prediction based on wind-terrain correlation consideration; wherein the standardized wind field and terrain data are input into the feature high-dimensional projection module , the feature high-dimensional projection module outputs the total high-dimensional features of the wind field and terrain ; the total high-dimensional features of the typhoon wind field and terrain are input into the multi-scale feature fusion module FM, and the multi-scale wind-terrain high-dimensional fusion features are output ; the first full connection prediction module uses the multi-scale wind-terrain high-dimensional fusion features for prediction, and outputs the first typhoon wind field grid prediction value;

[0014] S3, second typhoon wind field grid prediction; according to the meteorological observation site data, a site spatial position map and a site feature correlation map are established, and the two kinds of site data maps and the standardized data obtained in step S1 are input into the meteorological observation site spatial feature extraction meteorological observation site spatial linkage network, through the spatial position correlation pipeline flow, the feature correlation pipeline flow and the second full connection prediction module, the wind speed and wind direction prediction considering the spatial linkage between the meteorological observation sites are generated; wherein the spatial position correlation pipeline flow, the feature correlation pipeline flow output the site spatial information features , the second full connection prediction module uses the grid spatial information features for prediction, and outputs the second typhoon wind field grid prediction value;

[0015] S4, fusion prediction final typhoon wind field grid; obtaining typhoon information , grid spatial information , site density information input into the adaptive weighting network, generate the first weight corresponding to the first typhoon wind field grid prediction value and the second weight corresponding to the second typhoon wind field grid prediction value, based on the first weight and the second weight, the first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value are weighted and fused to obtain the final typhoon wind field grid prediction value.

[0016] Further, the data standardization process in step S1 is:

[0017]

[0018]

[0019] wherein, represents a standardization operation, represents input low-resolution typhoon wind field data, represents input high-resolution terrain data, represents standardized typhoon wind field data, represents standardized terrain data.

[0020] Further, step S2 is specifically:

[0021] S21, a feature high-dimensional projection module with the standardized typhoon wind field data and terrain data as input, the typhoon wind field data and terrain data are processed into high-dimensional features respectively, and then the high-dimensional features are spliced and output as the input of the multi-scale high-dimensional feature fusion module FM. The feature high-dimensional projection module The calculation formula is as follows:

[0022]

[0023]

[0024]

[0025] wherein, represents a feature high-dimensional projection module, represents high-dimensional typhoon wind field features processed by the feature high-dimensional projection module, represents high-dimensional terrain features processed by the feature high-dimensional projection module, represents a splicing operation, represents total high-dimensional features of typhoon wind field and terrain;

[0026] Each feature high-dimensional projection module is composed of three convolution modules, two Maxpooling pooling layers and an attention layer, each convolution module contains a convolution layer and a Silu activation layer, and the calculation process of each feature high-dimensional projection module is as follows:

[0027]

[0028]

[0029]

[0030]

[0031] wherein, represents the input of the feature extraction module, is a convolution operation, is a SiLU activation operation, is a Batch Normalization operation, is a max-pooling operation, represents the attention module processing operation, is the output of the feature high-dimensional projection module. S22, the multi-scale feature fusion module FM takes the high-dimensional feature as input, and fuses the high-dimensional wind field feature and the high-dimensional terrain feature through three multi-scale fusion (MSF) layers and three attention modules, captures the correlation between the wind field and the terrain, and outputs a multi-scale high-dimensional fusion feature

[0032] S22, the multi-scale feature fusion module FM takes the high-dimensional feature as input, and fuses the high-dimensional wind field feature and the high-dimensional terrain feature through three multi-scale fusion (MSF) layers and three attention modules, captures the correlation between the wind field and the terrain, and outputs a multi-scale high-dimensional fusion feature

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] wherein is the input high-dimensional feature, is the feature fusion operation of the module, is the feature fused by the first MSF layer, is the feature adjusted by the first attention module weight, is the feature fused by the second MSF layer, is the feature adjusted by the second attention module weight, is the feature fused by the third MSF layer is the feature adjusted by the third attention module weight, and the feature fusion is completed through the three MSF layers and the three attention modules to obtain the fusion feature ​​​;

[0040] S23, the first fully connected prediction module takes the multi-scale wind-terrain high-dimensional features as input, and projects the fusion features to a high-dimensional space through four fully connected-activation layers and a fully connected layer to generate multi-scale spatial feature extraction network predictions. The network formula is as follows:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] wherein is a linear layer conversion, is a SiLU activation operation, is a flattening operation, is the feature extracted by the first fully connected-activation layer, is the feature extracted by the second fully connected-activation layer, is the feature extracted by the third fully connected-activation layer, is the feature extracted by the fourth fully connected-activation layer, is the final prediction of the multi-scale spatial feature extraction network.

[0047] Further, the attention module in step S21 includes a wind residual spatial attention mechanism WR-SA and a channel attention mechanism CA; the wind residual spatial attention mechanism WR-SA aims to amplify the weight of key spatial positions, and the formula is as follows:

[0048]

[0049]

[0050]

[0051]

[0052] wherein represents input grid data, represents input standardized typhoon wind field grid data, represents taking the grid center value and copying it as the corresponding dimension, is a convolution operation, is​ activation operation, representing point multiplication operation, representing the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, representing the weight of each grid position, representing the grid position weight processed by two kinds of pooling operations, used for matching dimension and dimension, representing the output processed by the spatial attention mechanism.

[0053] The wind residual spatial attention mechanism uses the residual of the wind field grid and its center grid point as input, dynamically allocates the weight of each spatial position, and effectively improves the performance of the network.

[0054] The channel attention mechanism CA aims to amplify the weight of the key channel, and the formula is as follows:

[0055]

[0056]

[0057]

[0058]

[0059] wherein representing the input grid data, is an average pooling operation, which is used to average the variables of each channel into a channel response value, is a Linear linear layer conversion, is a SiLU activation operation, is activation operation, representing point multiplication operation, representing the channel response value corresponding to each channel, representing the high-dimensional channel response feature processed by the linear layer and activation operation, representing the weight corresponding to each channel, representing the output processed by the channel attention mechanism.

[0060] This mechanism can dynamically allocate channel weights according to the importance of the channel, so that the model pays more attention to important channels and improves the network performance.

[0061] The feature high-dimensional projection module receives typhoon wind field data and terrain data as input, and outputs high-dimensional features corresponding thereto This network design is simple and efficient, which can process typhoon wind field data and terrain data into corresponding high-dimensional features, and use attention mechanism to amplify the weight of key channels and key spatial positions, which helps the network capture the interaction between typhoon wind field and terrain.

[0062] Further, each feature fusion module MSF in step S22 is composed of three cavity convolution layers with different expansion rates, three SiLU activation layers, one convolution fusion module and one attention module. Each convolution module contains a convolution layer. The network formula is as follows:

[0063]

[0064]

[0065]

[0066]

[0067] wherein represents a cavity convolution with an expansion rate of 1, represents a cavity convolution with an expansion rate of 2, represents a cavity convolution with an expansion rate of 3, is the local scale fusion feature processed by the cavity convolution module with an expansion rate of 1, is the medium scale fusion feature processed by the cavity convolution module with an expansion rate of 2, is the large scale fusion feature processed by the cavity convolution module with an expansion rate of 3, is the multi-scale high-dimensional fusion feature output by the MSF layer.

[0068] The multi-scale high-dimensional feature fusion module FM receives the high-dimensional feature as input and outputs the multi-scale wind-terrain high-dimensional fusion feature . This module uses a cavity convolution module to capture the wind-terrain correlation at different scales and fuse it, so that the network can identify changes in wind behavior and subtle changes in terrain, improving the network's prediction ability.

[0069] Further, step S3 is specifically:

[0070] S31, establish a station spatial position map; based on meteorological observation station data, each station is regarded as a node in the graph structure, and the real observation value of the station, the corresponding typhoon wind field value of the station location, and the corresponding topographic feature value of the station location are regarded as the station features. The graph modeling process takes the spatial position of the station as the basis for constructing the edges between nodes. If the Euclidean distance between the longitude and latitude of any two nodes is less than a set threshold R, it is considered that there is an edge between the two nodes, and the edge plays an important role in the subsequent message passing mechanism;

[0071] S32, establish a station feature correlation map; based on meteorological observation station data, each station is regarded as a node in the graph structure, and the real observation value of the station, the corresponding typhoon wind field value of the station location, and the corresponding topographic feature value of the station location are regarded as the station features. The graph modeling process takes the station feature correlation as the basis for constructing the edges between nodes. If the cosine similarity of the features between any two nodes is greater than a set threshold T, it is considered that there is an edge between the two nodes;

[0072] S33, the meteorological observation station spatial linkage network includes two different pipeline flows for processing two modeling graphs. The network takes the modeling graph and node features as input and outputs the prediction based on the station spatial linkage consideration. The spatial position pipeline flow formula for processing the spatial position modeling graph is as follows:

[0073]

[0074]

[0075] wherein, represents the network input, represents a message aggregation layer, represents a dense fully connected layer. Each message aggregation layer is composed of an average aggregator, a maximum pooling aggregator, and an adaptive weighted fusion layer, and the calculation method is as follows:

[0076]

[0077]

[0078]

[0079]

[0080] wherein, is the average aggregated feature of the i-th node, is the number of neighbor nodes of the i-th node, is the feature of the j-th neighbor node, is the max-pooling aggregated feature for the i-th node, W and b represent trainable weights and bias, respectively, is the weight for the average aggregator, is the weight for the max-pooling aggregator.

[0081] The message aggregation layer aggregates the average and max features in the neighbor nodes, and can dynamically allocate the weights of the two features according to different nodes and different graph characteristics, effectively exerting the spatial linkage between sites and improving the prediction performance of the network.

[0082] The correlation pipeline for processing the feature correlation modeling graph includes one message aggregation layer and two graph attention message aggregation layers (GAT), and the formula is as follows:

[0083]

[0084]

[0085]

[0086] wherein, represents the network input, represents the message aggregation layer, represents the graph attention message aggregation layer. Each GAT layer includes a multi-head attention mechanism, which can dynamically weight the messages transmitted by the neighbor nodes according to the importance of the neighbor nodes. The parameters of each attention head are independent, and the outputs of all attention heads are connected The calculation method is as follows:

[0087]

[0088] wherein, K is the number of heads of the multi-head attention, represents the i-th node of the l+1 layer embedding hidden variable, is the activation operation, is the number of neighbor nodes of the i-th node, is the normalized attention matrix, is the linear transformation weight matrix of the k-th attention head.

[0089] Through the multi-head attention mechanism, the network can mine the potential association between the nodes and their neighbor nodes, thereby fully capturing the spatial linkage between the target grid points and the sites.

[0090] S34, the second fully connected prediction module takes the outputs of the two different pipeline streams and For input, the features are projected to a high-dimensional space through four fully-connected-activation layers and a fully-connected layer, and the output is generated based on the consideration of the spatial linkage of the stations. The network formula is as follows:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] wherein is the feature extracted by the first fully-connected-activation layer of the second fully-connected prediction module, is the feature extracted by the second fully-connected-activation layer of the second fully-connected prediction module, is the feature extracted by the third fully-connected-activation layer of the second fully-connected prediction module, is the feature extracted by the fourth fully-connected-activation layer of the second fully-connected prediction module.

[0097] Further, step S4 is specifically: an adaptive weighting network is used to generate the weights of the above two prediction values, and the final prediction is generated by weighted addition according to the weights. The adaptive weighting network has the following characteristics: the adaptive weighting network uses the typhoon information , the grid point space information , and the station density information as input data. The typhoon information includes typhoon intensity, typhoon center longitude and latitude, minimum pressure near the typhoon center, two-minute average maximum sustained wind near the tropical cyclone center, two-minute average sustained wind speed, and typhoon landing time. The grid point space information includes the longitude and latitude of the target grid point, the distance from the typhoon center, the azimuth angle relative to the typhoon center, the timestamp, the typhoon wind field data and the terrain data, the typhoon wind field data and the terrain data are the typhoon wind field grid and the terrain grid centered on the target grid point, and the typhoon wind field grid prediction value is the wind speed and wind direction prediction value of the target grid point. The station density information includes the number of edges of the node represented by the target grid point in the station spatial position graph and the station feature correlation graph. The adaptive weighting network is composed of five fully-connected-activation layers, one linear layer, and one SoftPlus activation layer. The network formula is as follows:

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] wherein, is the feature extracted by the first fully-connected-activation layer of the adaptive weighting network, is the feature extracted by the second fully-connected-activation layer of the adaptive weighting network, is the feature extracted by the third fully-connected-activation layer of the adaptive weighting network, is the feature extracted by the fourth fully-connected-activation layer of the adaptive weighting network, is the feature extracted by the fifth fully-connected-activation layer of the adaptive weighting network, is the first weight of the first typhoon wind field grid prediction value output by the multi-scale wind-terrain spatial correlation extraction network, is the second weight of the second typhoon wind field grid prediction value output by the meteorological observation site spatial linkage network, is the first typhoon wind field grid prediction value, is the second typhoon wind field grid prediction value, is the final prediction.

[0106] The deep learning-based typhoon wind field downscaling system for complex terrain areas comprises:

[0107] a data acquisition and standardization preprocessing module, which is responsible for acquiring meteorological observation site data, low-resolution typhoon wind field data and high-resolution terrain data from data sources, and performing standardization preprocessing on the wind field and terrain data to generate standardized data;

[0108] a first typhoon wind field grid prediction module, which inputs the standardized wind field and terrain data into the multi-scale wind-terrain spatial correlation extraction network, generates the first typhoon wind field grid prediction value based on the consideration of wind-terrain correlation through the feature high-dimensional projection module PM, the multi-scale high-dimensional feature fusion module FM and the first fully-connected prediction module;

[0109] The second typhoon wind field grid prediction module will generate a second typhoon wind field grid prediction value based on the spatial linkage between meteorological observation stations according to the spatial position map and the station feature correlation map established based on meteorological observation station data, the standardized data obtained by the data acquisition and standardization preprocessing module, the meteorological observation station spatial feature extraction meteorological observation station spatial linkage network, the spatial position correlation pipeline flow, the feature correlation pipeline flow and the second full connection prediction module;

[0110] The fusion prediction final typhoon wind field grid module inputs the typhoon information at the target grid point , the grid point spatial information , the station density information into the adaptive weighting network to generate the weight of the first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value, and to perform weighted fusion on the two prediction values to obtain the final typhoon wind field grid prediction value.

[0111] An electronic device comprises a memory for storing a computer program and a processor for implementing the steps of the above-mentioned complex terrain area typhoon wind field downscaling method based on deep learning when executing the computer program.

[0112] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above-mentioned complex terrain area typhoon wind field downscaling method based on deep learning.

[0113] The beneficial effects of the present application are:

[0114] 1. The complex terrain area typhoon wind field downscaling method based on deep learning disclosed in the application aims to solve the problems of insufficient accuracy and poor generalization ability of the prior art under complex terrain conditions. First, low-resolution engineering typhoon wind field data, high-resolution terrain data and site observation data are obtained, and after standardization processing, the model input standardized data is formed. Then, a multi-scale wind-terrain spatial correlation extraction network based on convolutional neural network is used to capture the spatial correlation between the typhoon wind field grid and the terrain grid. The network inputs the typhoon wind field grid and the terrain grid centered on the target grid point, and outputs the wind speed and wind direction prediction of the target grid point. In particular, the wind residual spatial attention mechanism and the channel attention mechanism are introduced in the multi-scale wind-terrain spatial correlation extraction network, which helps the model focus on important grid positions and important variables. Subsequently, a meteorological observation site spatial linkage network based on graph neural network is used to extract the spatial correlation between the target grid point and the meteorological observation site. The network inputs the meteorological observation site data, and outputs the wind speed and wind direction prediction of the target grid point; finally, according to the typhoon feature information, the spatial position information of the target grid point and the site distribution density of the region where the target grid point is located, an adaptive weighting network is used to weight and add the predictions based on two different considerations to generate the final wind field grid. In the training and optimization process of the multi-scale wind-terrain spatial correlation extraction network model, the back propagation method is used to feedback and correct the model parameters to improve the prediction accuracy and reduce the error. The grid comprehensively considers the correlation between the typhoon wind field and the terrain and the spatial linkage between the meteorological observation sites, and automatically adapts to different typhoon conditions, different geographical positions and different site distribution densities.

[0115] 2. The complex terrain area typhoon wind field downscaling method based on deep learning disclosed in the application captures the wind-terrain spatial correlation and the spatial linkage between the meteorological observation sites through two different network structures, respectively, and then uses an adaptive weighting network to dynamically weight the predictions of the two networks to obtain the final wind field downscaling result. The application comprehensively considers the correlation between the near-surface wind field and the terrain and the spatial linkage between the sites, and automatically adapts to different typhoon conditions, geographical positions and meteorological observation site distribution densities, which not only improves the accuracy of typhoon wind field downscaling, but also enhances the robustness of the downscaling effect, has important application value for typhoon mechanism research and disaster prevention, and can meet the typhoon wind field downscaling tasks under various scenarios.

[0116] 3. The deep learning-based complex terrain area typhoon wind field downscaling method disclosed in the application is a typhoon wind field downscaling method applied in typhoon research and typhoon disaster prediction and prevention. The deep learning network is used to process typhoon wind field data in different geographical regions, site distribution densities and typhoon conditions, so as to quickly and accurately generate a high-resolution typhoon wind field map, improve the accuracy and robustness of the typhoon wind field grid, and has important significance for the fields of typhoon research, typhoon prediction, disaster prevention and the like. The near-surface wind field downscaling method has the advantages of fast processing speed, good downscaling effect and strong robustness, and can meet the near-surface typhoon wind field downscaling tasks in various scenes. In addition, the implementation of the method does not require complex hardware support and can be easily deployed on existing computing platforms, providing an efficient typhoon wind field downscaling processing tool for the meteorological and disaster fields.

[0117] Other advantages, objects, and features of the present application will be understood by those skilled in the art from the following specification in conjunction with the accompanying drawings. The present application's objectives and other advantages will be realized and attained by the embodiments particularly pointed out in the specification as follows. BRIEF DESCRIPTION OF DRAWINGS

[0118] In order to make the objectives, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be given below in conjunction with the accompanying drawings, in which:

[0119] Figure 1 A flowchart of the deep learning-based complex terrain area typhoon wind field downscaling method of the present application;

[0120] Figure 2 A flowchart of the step S2 multi-scale wind-terrain spatial correlation network construction of the present application;

[0121] Figure 3 A flowchart of the step S3 meteorological observation site spatial linkage network construction of the present application;

[0122] Figure 4 A flowchart of the step S4 adaptive weighting network construction of the present application. DETAILED DESCRIPTION

[0123] The embodiments of the present application are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure in the specification. The present application can also be implemented or applied in other different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application.

[0124] As Figure 1A deep learning-based complex terrain area typhoon wind field downscaling method is shown, comprising the following steps:

[0125] S1, data acquisition and standardization preprocessing; acquiring meteorological observation site data, low resolution typhoon wind field data and high resolution terrain data , standardizing the acquired engineering typhoon model wind field data and terrain data to generate standardized data; the standardized wind field data and standardized terrain data are represented as and respectively;

[0126] The data standardization process is as follows:

[0127]

[0128]

[0129] wherein, represents a standardization operation, represents the input low resolution typhoon wind field data, represents the input high resolution terrain data, represents the standardized typhoon wind field data, and represents the standardized terrain data.

[0130] S2, first typhoon wind field grid prediction; as Figure 2 shown, the standardized wind field and terrain data are input into a multi-scale wind-terrain spatial correlation extraction network, and the multi-scale wind-terrain spatial correlation extraction network is used to generate wind speed and wind direction prediction based on wind-terrain correlation consideration;

[0131] The multi-scale wind-terrain spatial correlation extraction network outputs the wind speed and wind direction prediction of the target grid point by inputting the typhoon wind field grid and terrain grid centered on the target grid point. The multi-scale wind-terrain spatial correlation extraction network is used to generate prediction based on wind-terrain correlation consideration. The multi-scale wind-terrain spatial correlation extraction network is composed of a feature high-dimensional projection module , a multi-scale high-dimensional feature fusion module FM, and a first full connection prediction module. The standardized wind field and terrain data are input into the feature high-dimensional projection module , and the feature high-dimensional projection module outputs the total high-dimensional features of the wind field and terrain . The total high-dimensional features of the typhoon wind field and terrain are input into the multi-scale feature fusion module FM, and the multi-scale wind-terrain high-dimensional fusion features are output. The first full connection prediction module uses the multi-scale wind-terrain high-dimensional fusion features The first typhoon wind field grid prediction value is outputted by performing prediction, and the specific process is as follows:

[0132] S21, feature high-dimensional projection module The standardized typhoon wind field data and the terrain data are taken as inputs, and the typhoon wind field data and the terrain data are processed into high-dimensional features respectively. After splicing, the output is taken as the input of the multi-scale high-dimensional feature fusion module FM. The feature high-dimensional projection module The calculation formula is as follows:

[0133]

[0134]

[0135]

[0136] Among them, represents the feature high-dimensional projection module, represents the high-dimensional typhoon wind field feature processed by the feature high-dimensional projection module, represents the high-dimensional terrain feature processed by the feature high-dimensional projection module, represents the splicing operation, represents the total high-dimensional feature of the typhoon wind field and the terrain.

[0137] Each feature high-dimensional projection module is composed of three convolution modules, two Maxpooling pooling layers and an attention layer, each convolution module includes a convolution layer and a Silu activation layer, and the calculation process of the feature high-dimensional projection module is as follows:

[0138]

[0139]

[0140]

[0141]

[0142] Among them, represents the input of the feature extraction module , is a convolution operation, is a SiLU activation operation, is a Batch Normalization batch normalization operation, is a maximum pooling operation, represents an attention module processing operation, is a feature high-dimensional projection module The output.

[0143] The attention module includes the Wind Residual-Spatial Attention (WR-SA) mechanism and the Channel Attention (CA) mechanism. The WR-SA mechanism aims to amplify the weights of key spatial locations, as shown in the following formula:

[0144]

[0145]

[0146]

[0147]

[0148] in Represents input grid data, The input represents standardized typhoon wind field grid data. This represents taking the center value of the grid and copying it. The corresponding dimensions For convolution operations, for Activation operation, Represents the dot product operation. This represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid. The weight representing each grid position, This represents the grid position weights after processing through two pooling operations, used for matching. Dimensions and Dimension This represents the output after processing by the spatial attention mechanism.

[0149] This wind residual spatial attention mechanism uses the residual between the wind field grid and its central grid point as input, dynamically assigning weights to each spatial location, effectively improving the network performance.

[0150] The Channel Attention (CA) mechanism aims to amplify the weights of key channels, as shown in the following formula:

[0151]

[0152]

[0153]

[0154]

[0155] in Represents input grid data, This is an average pooling operation used to average the variables of each channel into a single channel response value. For linear layer transformation, For SiLU activation operation, for Activation operation, Represents the dot product operation. This represents the channel response value corresponding to each channel. Represents passing through linear layers and High-dimensional channel response features after activation operation processing This represents the weight corresponding to each channel. This represents the output after processing by the channel attention mechanism.

[0156] This mechanism allows the model to dynamically allocate channel weights based on channel importance, making the model pay more attention to important channels and improving network performance.

[0157] The high-dimensional feature projection module receives typhoon wind field data and terrain data as input and outputs the corresponding high-dimensional features. This network design is simple and efficient, capable of processing typhoon wind field data and terrain data into corresponding high-dimensional features, and using an attention mechanism to amplify the weights of key channels and key spatial locations, which helps the network capture the interaction between typhoon wind fields and terrain.

[0158] S22, Multi-scale Feature Fusion Module FM uses high-dimensional features As input, the high-dimensional wind field features and high-dimensional terrain features are fused through three multi-scale fusion (MSF) layers and three attention modules to capture the correlation between wind field and terrain, and output multi-scale high-dimensional fused features. This serves as the input to the first fully connected prediction module. The calculation formula for the multi-scale feature fusion module is as follows:

[0159]

[0160]

[0161]

[0162]

[0163]

[0164]

[0165] in For input high-dimensional features, For Feature fusion operation of the module, For the first MSF layer fused features, For the first attention module weight adjusted features, For the second MSF layer fused features, For the second attention module weight adjusted features, For the third MSF layer fused features For the third attention module weight adjusted features, after three MSF layers and three attention modules, the feature fusion work is completed, and the fused features .

[0166] Each feature fusion module MSF is composed of three different dilated rate cavity convolution layers, three SiLU activation layers, a convolution fusion module and an attention module. Each convolution module contains a convolution layer. The network formula is as follows:

[0167]

[0168]

[0169]

[0170]

[0171] Wherein represents a dilated rate of 1 cavity convolution, represents a dilated rate of 2 cavity convolution, represents a dilated rate of 3 cavity convolution, is the local scale fusion feature processed by the dilated rate of 1 cavity convolution module, is the medium scale fusion feature processed by the dilated rate of 2 cavity convolution module, is the large scale fusion feature processed by the dilated rate of 3 cavity convolution module, is the multi-scale high-dimensional fusion feature output by the MSF layer.

[0172] The multi-scale high-dimensional feature fusion module FM receives high-dimensional features as input, and outputs multi-scale wind-terrain high-dimensional fusion features . This module uses a cavity convolution module to capture the wind field-terrain correlation at different scales and fuse it, so that the network can identify the changes of wind behavior and the subtle changes of terrain, and improve the prediction ability of the network.

[0173] S23, the first full connection prediction module takes multi-scale wind-terrain high-dimensional features For input, the fused features are passed through four fully-connected-activation layers and a fully-connected layer projected into a high-dimensional space and a multi-scale spatial feature extraction network is generated. The network formula is as follows:

[0174]

[0175]

[0176]

[0177]

[0178]

[0179] wherein is a linear layer conversion, is a SiLU activation operation, is a flattening operation, is the feature extracted by the first fully-connected-activation layer, is the feature extracted by the second fully-connected-activation layer, is the feature extracted by the third fully-connected-activation layer, is the feature extracted by the fourth fully-connected-activation layer, is the final prediction of the multi-scale spatial feature extraction network.

[0180] The multi-scale spatial feature extraction network receives typhoon wind field data and terrain data as input and outputs a prediction based on wind-terrain correlation considerations. This network design is simple and effective, capable of capturing the interaction between wind field and terrain at different scales, and using attention mechanism to amplify the weight of key channels and key spatial positions, generating accurate near-surface wind field grids.

[0181] S3, second typhoon wind field grid prediction; as Figure 3 shown, a site spatial position map and a site feature correlation map are established according to the meteorological observation site data. The two kinds of site data maps and the standardized data are input into the site spatial feature extraction meteorological observation site spatial linkage network, and a prediction based on site spatial linkage considerations is output, which uses real site observation to guide the generation of prediction in site-free areas, and obtains high-quality near-surface wind field grids;

[0182] The meteorological observation site spatial linkage network includes a spatial position correlation pipeline, a feature correlation pipeline and a second fully-connected prediction module. The spatial position correlation pipeline and the feature correlation pipeline output site spatial information features , and the second fully-connected prediction module uses grid spatial information features The second typhoon wind field grid prediction value is outputted according to the prediction, and the specific process is as follows:

[0183] S31, a station space position graph is established; each station is regarded as a node in the graph structure based on the meteorological observation station data, and the real observation value of the station, the typhoon wind field value corresponding to the position of the station and the terrain feature value corresponding to the position of the station are regarded as the station features. The graph modeling process takes the station space position as the basis for constructing the edges between nodes. If the Euclidean distance between the longitude and latitude of any two nodes is less than a set threshold R, it is considered that there is an edge between the two nodes, and the edge plays an important role in the subsequent message passing mechanism;

[0184] S32, a station feature correlation graph is established; each station is regarded as a node in the graph structure based on the meteorological observation station data, and the real observation value of the station, the typhoon wind field value corresponding to the position of the station and the terrain feature value corresponding to the position of the station are regarded as the station features. The graph modeling process takes the station feature correlation as the basis for constructing the edges between nodes. If the feature cosine similarity between any two nodes is greater than a set threshold T, it is considered that there is an edge between the two nodes;

[0185] S33, the meteorological observation station space linkage network includes two different pipeline flows for processing two modeling graphs. The network takes the modeling graph and the node feature as the input, and outputs the prediction based on the station space linkage consideration. The space position pipeline flow formula for processing the space position modeling graph is as follows:

[0186]

[0187]

[0188] wherein, represents the network input, represents a message aggregation layer, represents a dense fully connected layer. Each message aggregation layer is composed of an average aggregator, a maximum pooling aggregator and an adaptive weighted fusion layer, and the calculation method is as follows:

[0189]

[0190]

[0191]

[0192]

[0193] wherein, is the average aggregated feature of the i th node, is the number of neighbor nodes of the i th node, The feature of the jth neighbor node, The maximum pooling aggregated feature of the ith node, W and b represent the trainable weight and bias, respectively, The weight occupied by the average aggregator, The weight occupied by the maximum pooling aggregator.

[0194] The message aggregation layer aggregates the average feature and the maximum feature in the neighbor nodes, and can dynamically allocate the weights of the two features according to different nodes and different graph characteristics, effectively exerting the spatial linkage between sites and improving the prediction performance of the network.

[0195] The correlation pipeline for processing the feature correlation modeling graph includes one message aggregation layer and two graph attention message aggregation layers (GAT), and the formula is as follows:

[0196]

[0197]

[0198]

[0199] wherein, represents the network input, represents the message aggregation layer, represents the graph attention message aggregation layer. Each GAT layer includes a multi-head attention mechanism, which can dynamically weight the messages transmitted by the neighbor nodes according to the importance of the neighbor nodes. The parameters of each attention head are independent, and the outputs of all attention heads are connected The calculation method is as follows:

[0200]

[0201] wherein, K is the number of heads of the multi-head attention, represents the (l+1)th layer embedding hidden variable of the ith node, is an activation operation, is the number of neighbor nodes of the ith node, is the normalized attention matrix, is the linear transformation weight matrix of the kth attention head.

[0202] Through the multi-head attention mechanism, the network can mine the potential association between the nodes and their neighbor nodes, thereby fully capturing the spatial linkage between the target grid points and the sites.

[0203] The second fully connected prediction module takes the outputs of the two different pipeline streams and For input, the features are projected to a high-dimensional space through four fully connected-activation layers and a fully connected layer, and the output is generated based on the site space linkage consideration. The network formula is as follows:

[0204]

[0205]

[0206]

[0207]

[0208]

[0209] wherein is the feature extracted by the first fully connected-activation layer of the second fully connected prediction module, is the feature extracted by the second fully connected-activation layer of the second fully connected prediction module, is the feature extracted by the third fully connected-activation layer of the second fully connected prediction module, is the feature extracted by the fourth fully connected-activation layer of the second fully connected prediction module.

[0210] The network receives site data, wind field data and terrain data as input, and outputs a prediction based on site space linkage consideration. This network can capture the implicit correlation between the target grid point and the site, so as to use real site observations to guide the generation of predictions in site-free areas, thereby obtaining a high-quality near-surface wind field grid.

[0211] S4, fuse the final typhoon wind field grid; as Figure 4 shown to obtain typhoon information at the target grid point , grid space information , site density information input the adaptive weighting network to generate a first weight corresponding to the first typhoon wind field grid prediction value and a second weight corresponding to the second typhoon wind field grid prediction value, and to perform weighted fusion on the first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value based on the first weight and the second weight to obtain a final typhoon wind field grid prediction value. Specifically:

[0212] The adaptive weighting network is used to generate the weights of the above two prediction values, and to perform weighted addition according to the weights, thereby generating the final prediction. Wherein, the adaptive weighting network has the following characteristics: the adaptive weighting network uses typhoon information at the target grid point , grid space information The grid point space information includes typhoon intensity, typhoon center longitude and latitude, minimum pressure near the typhoon center, two-minute average maximum sustained wind near the tropical cyclone center, two-minute average sustained wind speed, and typhoon landing time. The grid point space information includes target grid point longitude and latitude, distance from the typhoon center, azimuth angle relative to the typhoon center, timestamp, typhoon wind field data, and terrain data, the typhoon wind field data and terrain data are typhoon wind field grid and terrain grid centered on the target grid point, and the typhoon wind field grid prediction value is the wind speed and wind direction prediction value of the target grid point. The site density information The site density information includes the number of edges of the node represented by the target grid point in the site space position graph and the site feature correlation graph. The adaptive weighted network is composed of five fully connected-activation layers, one linear layer, and one SoftPlus activation layer. The network formula is as follows:

[0213]

[0214]

[0215]

[0216]

[0217]

[0218]

[0219]

[0220] wherein, is the feature extracted by the first fully connected-activation layer of the adaptive weighted network, is the feature extracted by the second fully connected-activation layer of the adaptive weighted network, is the feature extracted by the third fully connected-activation layer of the adaptive weighted network, is the feature extracted by the fourth fully connected-activation layer of the adaptive weighted network, is the feature extracted by the fifth fully connected-activation layer of the adaptive weighted network, is the first weight of the first typhoon wind field grid prediction value output by the multi-scale wind-terrain spatial correlation extraction network, is the second weight of the second typhoon wind field grid prediction value output by the meteorological observation site space linkage network, is the first typhoon wind field grid prediction value, is the second typhoon wind field grid prediction value, is the final prediction.

[0221] The network can dynamically adjust the multi-scale spatial feature extraction network prediction value weight and the station spatial linkage network prediction value weight according to the geographical region where the target grid is located, the site distribution density and the typhoon condition, which can not only improve the prediction performance of the network, but also improve the robustness of the network, ensure the stability of the prediction result, and generate high-quality near-surface typhoon wind field.

[0222] The training process of the multi-scale wind-terrain spatial correlation extraction network based on the convolutional neural network includes model training, model verification and model testing. The model training stage includes forward propagation and backward propagation, the network parameters are optimized through the loss function, and the learning process of the model is realized. In the model verification stage, the generalization of the model is evaluated by using the five-fold cross-validation method, and the model hyperparameters are adjusted according to the evaluation results. In the model testing stage, independent test samples are used to test the performance of the trained model, to ensure the robustness and accuracy of the model on unknown data.

[0223] The specific embodiments of the present application provide a new typhoon near-surface wind field downscaling solution for related field technicians, which helps to promote the development of meteorological data downscaling technology. The protection scope should be subject to the claims. The implementation of this method will greatly improve the quality and availability of typhoon near-surface wind field under complex terrain conditions, and thus plays an important role in typhoon prediction and disaster prevention. By providing higher quality typhoon wind field grid, more accurate scientific research and decision making can be carried out, which is crucial for improving the efficiency and safety of these fields. In addition, the downscaling technology of the present application can also be applied to other fields such as scientific research and risk assessment, where high-quality wind field grid is also crucial. In summary, the present application provides a powerful and flexible tool for improving the quality of near-surface typhoon wind field, opening up new possibilities for various applications.

[0224] The near-surface wind field downscaling method of the present application can efficiently and high-quality complete the downscaling task, providing strong technical support for related field research and application. Through the present application, users can obtain clearer and more accurate wind field maps, and thus carry out more effective scientific research and decision making.

[0225] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the purpose and scope of the present application, which should be covered by the claims of the present application.

Claims

1. A deep learning-based method for downscaling typhoon wind fields in complex terrain regions, characterized in that, Includes the following steps: S1. Data Acquisition and Standardization Preprocessing; This involves processing the acquired meteorological observation station data and low-resolution typhoon wind field data. and high-resolution terrain data Standardized preprocessing generates standardized data; S2. First Typhoon Wind Field Grid Prediction: Standardized wind field and terrain data are input into multi-scale wind-terrain spatial feature extraction to extract the typhoon wind field and terrain network, which is then processed through a feature high-dimensional projection module. The multi-scale high-dimensional feature fusion module (FM) and the first fully connected prediction module generate wind speed and direction predictions based on wind-topography correlation; among them, the standardized wind field and topographic data input feature high-dimensional projection module... Feature high-dimensional projection module Output wind field and topographic total high-dimensional features ; Overall high-dimensional characteristics of typhoon wind field and topography Input the multi-scale feature fusion module FM, output multi-scale wind-topography high-dimensional fused features. The first fully connected prediction module utilizes multi-scale wind-topography high-dimensional fusion features. Make predictions; S3, second typhoon wind field grid prediction; Based on meteorological observation station data, a spatial location map and a station feature correlation map are established. The two types of station data maps and the standardized data obtained in step S1 are input into the meteorological observation station spatial feature extraction meteorological observation station spatial linkage network. Through the spatial location correlation pipeline flow, the feature correlation pipeline flow and the second fully connected prediction module, wind speed and wind direction predictions are generated considering the spatial linkage between meteorological observation stations. Among them, spatial location-related pipeline flow and feature-related pipeline flow output station spatial information features The second fully connected prediction module utilizes the spatial information features of grid points. Make predictions; S4. Fusion prediction of the final typhoon wind field grid; typhoon information at the target grid points. Grid spatial information Site density information The adaptive weighted network is input to generate a first weight corresponding to the first typhoon wind field grid prediction value and a second weight corresponding to the second typhoon wind field grid prediction value. The first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value are weighted and fused based on the first weight and the second weight to obtain the final typhoon wind field grid prediction value.

2. The method for downscaling typhoon wind fields in complex terrain areas as described in claim 1, characterized in that, The data standardization process in step S1 is as follows: in, Represents standardized operation. This represents the input low-resolution typhoon wind field data. Represents the input high-resolution terrain data, This represents standardized typhoon wind field data. This represents standardized terrain data.

3. The method for downscaling typhoon wind fields in complex terrain areas as described in claim 2, characterized in that, Step S2 is as follows: S21, Feature High-Dimensional Projection Module Using standardized typhoon wind field data and topographic data as input, the typhoon wind field data and topographic data are processed into high-dimensional features respectively. Then, high-dimensional features The output after stitching is used as input to the multi-scale high-dimensional feature fusion module FM; the feature high-dimensional projection module The calculation formula is as follows: in, High-dimensional projection module representing characteristics This represents the high-dimensional typhoon wind field characteristics after processing by the feature high-dimensional projection module. This represents the high-dimensional terrain features after processing by the feature high-dimensional projection module. This represents a splicing operation. Represents the overall high-dimensional characteristics of typhoon wind field and topography; High-dimensional projection module for each feature It consists of three convolutional modules, two maxpooling layers, and an attention layer. Each convolutional module contains a convolutional layer and a Silu activation layer, and each feature has a high-dimensional projection module. The calculation process is as follows: in, Representative feature extraction module Input, For convolution operations, For SiLU activation operation, For Batch Normalization, For max pooling operation, This represents the attention module processing operation. For feature high-dimensional projection module The output; S22, Multi-scale Feature Fusion Module FM uses high-dimensional features As input, the high-dimensional wind field features and high-dimensional terrain features are fused through three multi-scale fusion layers (MSF) and three attention modules to capture the correlation between wind field and terrain, and output multi-scale high-dimensional fused features. The multi-scale feature fusion module FM calculation formula is as follows: in To input high-dimensional features, for Feature fusion operation of the module Features of the first MSF layer fusion The features are after weight adjustment for the first attention module. Features of the second MSF layer fusion The features are after weight adjustment for the second attention module. Features fused for the third MSF layer The features, after weight adjustment by the third attention module, undergo feature fusion through three MSF layers and three attention modules to obtain the fused features. ; S23, The first fully connected prediction module uses multi-scale wind-topography high-dimensional features As input, features are fused through four fully connected activation layers and one fully connected layer. Projecting onto a high-dimensional space and generating a multi-scale spatial feature extraction network for prediction; the network formula is as follows: in For linear layer transformation, For SiLU activation operation, For flattening operation, Features extracted from the first fully connected-activation layer. Features extracted for the second fully connected-activation layer Features extracted for the third fully connected-activation layer Features extracted from the fourth fully connected-activation layer. This is the final prediction for the multi-scale spatial feature extraction network.

4. The method for downscaling typhoon wind fields in complex terrain areas as described in claim 3, characterized in that, The attention module in step S21 includes the wind residual spatial attention mechanism WR-SA and the channel attention mechanism CA; the wind residual spatial attention mechanism WR-SA aims to amplify the weights of key spatial locations, as shown in the following formula: in Represents input grid data, The input represents standardized typhoon wind field grid data. This represents taking the center value of the grid and copying it. The corresponding dimensions For convolution operations, for Activation operation, Represents the dot product operation. This represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid. The weight representing each grid position, This represents the grid position weights after processing through two pooling operations, used for matching. Dimensions and Dimension This represents the output after processing by the spatial attention mechanism; The Channel Attention (CA) mechanism aims to amplify the weights of key channels, as shown in the following formula: in Represents input grid data, This is an average pooling operation used to average the variables of each channel into a single channel response value. For linear layer transformation, For SiLU activation operation, for Activation operation, Represents the dot product operation. This represents the channel response value corresponding to each channel. Represents passing through linear layers and High-dimensional channel response features after activation operation processing This represents the weight corresponding to each channel. This represents the output after processing by the channel attention mechanism.

5. The method for downscaling typhoon wind fields in complex terrain areas as described in claim 4, characterized in that, In step S22, each feature fusion module (MSF) consists of three dilated convolutional layers with different dilation rates, three SiLU activation layers, one convolutional fusion module, and one attention module; each convolutional module contains convolutional layers, and the network formula is as follows: in A dilated convolution with an inflation rate of 1 A dilated convolution with an inflation rate of 2, A dilated convolution with an inflation rate of 3 is represented. This represents the local scale fusion features after processing by a dilated convolutional module with an inflation rate of 1. This is a medium-scale fused feature after being processed by a dilated convolutional module with an inflation rate of 2. This represents the large-scale fused features after processing by a dilated convolutional module with an inflation rate of 3. This represents the multi-scale, high-dimensional fusion features output by the MSF layer.

6. The method for downscaling typhoon wind fields in complex terrain areas as described in claim 5, characterized in that, Step S3 is as follows: S31. Establish a station spatial location map based on the spatial location of meteorological observation stations as the basis for constructing edges between nodes. If the Euclidean distance between any two nodes is less than a set threshold R, then it is considered that there is an edge between the two nodes. S32. Establish a station feature correlation graph based on the feature correlation of meteorological observation stations as the basis for constructing edges between nodes. If the feature cosine similarity between any two nodes is greater than the set threshold T, then it is considered that there is an edge between the two nodes. S33. The spatial linkage network of meteorological observation stations takes the modeling map and node features as input and outputs predictions based on station spatial linkage considerations. The spatial location pipeline flow formula used to process the spatial location modeling map is as follows: in, Represents network input. represent Message aggregation layer Represents a dense fully connected layer; each The message aggregation layer consists of an average aggregator, a max-pooling aggregator, and an adaptive weighted fusion layer, calculated as follows: in, The average aggregation feature of the i-th node, Let be the number of neighboring nodes of the i-th node. The features of the j-th neighbor node are... Let W be the max-pooling aggregated feature of the i-th node, and let W and b represent the trainable weights and biases, respectively. The weight of the average aggregator. The weight assigned to the maximum pooling aggregator; The correlation pipeline used to process feature correlation modeling graphs contains a The message aggregation layer and two GAT graph attention message aggregation layers are defined by the following formula: in, Represents network input. represent Message aggregation layer The GAT (Graph Attention Message Aggregation) layer represents the graph attention message aggregation layer. Each GAT layer contains a multi-head attention mechanism that dynamically weights the messages passed by neighboring nodes based on their importance. Each attention head has independent parameters, and the outputs of all attention heads are concatenated. The calculation method is as follows: Where K is the number of heads in multi-head attention. This represents the (l+1)th level embedded latent variable of the i-th node. For activation, Let be the number of neighboring nodes of the i-th node. The standardized attention matrix, Let be the linear transformation weight matrix for the k-th attention head; S34, the second fully connected prediction module outputs two different pipeline flows. and The input is processed through four fully connected-activation layers and one fully connected layer to project the features into a high-dimensional space, generating an output based on site spatial linkage considerations; the network formula is as follows: in Features extracted from the first fully connected-activation layer of the second fully connected prediction module. The features extracted from the second fully connected-activation layer of the second fully connected prediction module. The features extracted from the third fully connected-activation layer of the second fully connected prediction module. The features are extracted from the fourth fully connected-activation layer of the second fully connected prediction module.

7. The method for downscaling typhoon wind fields in complex terrain areas as described in claim 6, characterized in that, Step S4 specifically involves: the adaptive weighted network using typhoon information at the target grid point. Grid spatial information Site density information The typhoon information serves as input data. This includes typhoon intensity, typhoon center latitude and longitude, minimum air pressure near the typhoon center, two-minute average maximum sustained winds near the tropical cyclone center, two-minute average sustained wind speeds, and typhoon landfall duration; the gridded spatial information. This includes the latitude and longitude of the target grid point, its distance from the typhoon center, its azimuth relative to the typhoon center, timestamp, typhoon wind field data, and terrain data. The typhoon wind field data and terrain data are typhoon wind field grids and terrain grids centered on the target grid point. The predicted values ​​of the typhoon wind field grid are the predicted wind speed and wind direction of the target grid point; the station density information... This includes the number of edges in the site spatial location map and the site feature correlation map represented by the target grid point; the adaptive weighted network consists of five fully connected activation layers, one linear layer, and one SoftPlus activation layer; the network formula is as follows: in, Features extracted from the first fully connected-activation layer of the adaptive weighted network. Features extracted from the second fully connected-activation layer of the adaptive weighted network. Features extracted from the third fully connected-activation layer of the adaptive weighted network. Features extracted from the fourth fully connected-activation layer of the adaptive weighted network. Features extracted from the fifth fully connected-activation layer of the adaptive weighted network. The first weight is used to extract the first typhoon wind field grid prediction value from the multi-scale wind-topography spatial correlation extraction network output. The second weight of the second typhoon wind field grid prediction value output by the spatial linkage network of meteorological observation stations. This is the grid prediction value for the wind field of the first typhoon. This is the grid prediction value for the wind field of the second typhoon. This is the final prediction.

8. A deep learning-based typhoon wind field downscaling system for complex terrain regions, characterized in that, include: The data acquisition and standardization preprocessing module is responsible for acquiring meteorological observation station data and low-resolution typhoon wind field data from data sources. and high-resolution terrain data Furthermore, standardized preprocessing was performed on wind field and topographic data to generate standardized data; The first typhoon wind field grid prediction module inputs standardized wind field and terrain data into a multi-scale wind-terrain spatial correlation extraction network. Through the feature high-dimensional projection module PM, the multi-scale high-dimensional feature fusion module FM, and the first fully connected prediction module, it generates the first typhoon wind field grid prediction value based on wind-terrain correlation consideration. The second typhoon wind field grid prediction module will input the station spatial location map and station feature correlation map established based on meteorological observation station data, as well as the standardized data obtained by the data acquisition and standardization preprocessing module, into the meteorological observation station spatial feature extraction meteorological observation station spatial linkage network. Through the spatial location correlation pipeline flow, feature correlation pipeline flow and the second fully connected prediction module, the second typhoon wind field grid prediction value based on the spatial linkage between meteorological observation stations will be generated. The fusion prediction final typhoon wind field grid module integrates typhoon information at the target grid points. Grid spatial information Site density information The adaptive weighted network is input to generate weights for the first and second typhoon wind field grid predictions. The two predictions are then weighted and fused to obtain the final typhoon wind field grid prediction.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute a computer program to implement the steps of the deep learning-based typhoon wind field downscaling method for complex terrain regions as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the steps of the deep learning-based typhoon wind field downscaling method for complex terrain areas as described in any one of claims 1 to 7.

Images