Complex terrain area typhoon wind field downscaling method and system based on deep learning

Abstract

The invention relates to a complex terrain area typhoon wind field downscaling method and system based on deep learning. 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 topographic data into a multi-scale wind-topographic spatial correlation extraction network to generate a first typhoon wind field grid prediction value; a second typhoon wind field grid prediction module establishes a site space location map and a site feature correlation map according to meteorological observation site data, inputs standardized data into meteorological observation site space features to extract a meteorological observation site space linkage network, and generates a second typhoon wind field grid prediction value; a final typhoon wind field grid fusion prediction module performs weighted fusion on the generated first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value to obtain a final typhoon wind field grid prediction value; according to the typhoon wind field downscaling method, the accuracy and efficiency of typhoon wind field downscaling in a complex terrain area are improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and meteorological computing technology, and relates to a typhoon wind field downscaling method and system in complex terrain areas based on deep learning. Background Art

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

[0003] However, the current traditional downscaling methods are limited by the speed and effect of wind field downscaling, especially the low accuracy of typhoon wind field downscaling under highly complex terrain conditions, resulting in low quality and availability of typhoon wind fields, and the inability to obtain accurate downscaling results in a timely manner.

[0004] For example, patent CN116595366A describes a method and system for spatial downscaling of offshore wind farms based on multi-task learning. This method uses a multi-task training dataset to train and optimize a generative adversarial network to achieve spatial downscaling of offshore wind farms, and verifies the downscaling results using buoy data. However, this patented method is only applicable to downscaling tasks on offshore wind farms and is not suitable for downscaling tasks on complex land terrain.

[0005] Patent CN118366046A discloses a wind field downscaling method based on deep learning and combined with terrain, capable of downscaling predictions with a latitude and longitude resolution of 0.25°×0.25° to 0.1°×0.1°. However, the patent's downscaling method uses the European Center for Meteorological Observations (ECMWF) ERA5-Land data as observation data for model training, without considering the error between the ERA5-Land data and real weather station observations. As a result, there is still a large error between the downscaling results and the real weather station observations.

[0006] Patent CN118761026A discloses a deep learning-based method for downscaling near-surface wind fields. This method considers the complex nonlinear relationship between wind and terrain and uses meteorological station observations as ground truth to train the network. This method is capable of downscaling wind fields in complex terrain. However, this patented downscaling method does not account for wind speed variations in typhoon weather conditions and is therefore not suitable for typhoon wind field downscaling.

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

[0008] In summary, 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 in patent CN116595366A, which cannot be applied to complex land terrain; (2) Some methods have defects in data processing, such as patent CN118366046A 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 do not perform well under complex weather conditions, such as patent CN118761026A 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 of the Invention

[0009] In view of this, in order to solve the problem that traditional downscaling methods have slow downscaling speed and low accuracy in complex terrain conditions, resulting in poor wind field quality and availability, and difficulty in providing accurate typhoon wind field downscaling results in a timely manner, the present invention provides a typhoon wind field downscaling method and system for complex terrain areas based on deep learning. The method uses a deep learning network to adaptively weigh the wind-terrain correlation considerations under different typhoon conditions, different geographical regions, and different terrain conditions, and uses meteorological station observation information to eliminate uncertainty in the downscaling process, thereby improving the accuracy and efficiency of typhoon wind field downscaling in complex terrain areas.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A deep learning-based downscaling method for typhoon wind fields in complex terrain areas includes the following steps:

[0012] S1. Data acquisition and standardization preprocessing; obtain meteorological observation station data and low-resolution typhoon wind field data W LR and high-resolution terrain data T HR , the acquired engineering typhoon model wind field data and terrain data are subjected to standardized preprocessing to generate standardized data;

[0013] S2. First typhoon wind field grid forecast; input the standardized wind field and terrain data into the multi-scale wind-terrain spatial feature to extract the typhoon wind field and terrain network, and generate the wind speed and wind direction forecast based on wind-terrain correlation considerations through the feature high-dimensional projection module PM, the multi-scale high-dimensional feature fusion module FM, and the first fully connected prediction module; the standardized wind field and terrain data are input into the feature high-dimensional projection module PM, and the feature high-dimensional projection module PM outputs the total high-dimensional feature F of the wind field and terrain total ; Typhoon wind field and terrain total high-dimensional characteristics F total Input the multi-scale feature fusion module FM and output the multi-scale wind-terrain high-dimensional fusion feature F fused The first fully connected prediction module uses the multi-scale wind-terrain high-dimensional fusion feature F fused Make predictions and output the first typhoon wind field grid prediction value;

[0014] S3, the second typhoon wind field grid forecast; according to the meteorological observation station data, establish the site spatial location map and the site feature correlation map, input the two site data maps and the standardized data obtained in step S1 into the meteorological observation station spatial feature extraction meteorological observation station spatial linkage network, and generate the wind speed and wind direction forecast considering the spatial linkage between meteorological observation stations through the spatial location correlation pipeline flow, feature correlation pipeline flow and the second fully connected prediction module; wherein the spatial location correlation pipeline flow and feature correlation pipeline flow output site spatial information feature I SI The second fully connected prediction module uses the grid space information feature I SI Make predictions and output the grid prediction values ​​of the second typhoon wind field;

[0015] S4. Fusion forecast of the final typhoon wind field grid; obtaining typhoon information at the target grid point I T , Grid space information I SI , site density information I SD Input the adaptive weighted 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. 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 weightedly fused to obtain the final typhoon wind field grid prediction value.

[0016] Furthermore, the data normalization process in step S1 is:

[0017] WNorm =Norm(W LR )

[0018] T Norm =Norm(T HR )

[0019] Among them, Norm(*) represents the standardization operation, W LR Represents the input low-resolution typhoon wind field data, T HR Represents the input high-resolution terrain data, W Norm represents the standardized typhoon wind field data, T Norm Represents normalized terrain data.

[0020] Further, step S2 is specifically as follows:

[0021] S21, the feature high-dimensional projection module PM takes the standardized typhoon wind field data and terrain data as input, and processes the typhoon wind field data and terrain data into high-dimensional features F total , and then the high-dimensional feature F total The spliced ​​output is used as the input to the multi-scale high-dimensional feature fusion module FM. The calculation formula of the feature high-dimensional projection module PM is as follows:

[0022] F W =PM(W Norm )

[0023] F T =PM(T Norm )

[0024] F total =Concatenate(F W ,F T )

[0025] Among them, PM(*) represents the feature high-dimensional projection module, F W represents the high-dimensional typhoon wind field characteristics after being processed by the feature high-dimensional projection module, F T represents the high-dimensional terrain features processed by the feature high-dimensional projection module, Concatenate(*) represents the splicing operation, and F total Represents the overall high-dimensional characteristics of the typhoon wind field and terrain;

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

[0027] F1=MaxPooling(BN(SiLU(Conv(Input))))

[0028] F2=MaxPooling(BN(SiLU(Conv(F1))))

[0029] F3=BN(SiLU(Conv(F2)))

[0030] F output =Attention(SiLU(Conv(F3)))

[0031] Among them, Input represents the input of the feature extraction module PM, Conv(*) is the convolution operation, SiLU(*) is the SiLU activation operation, BN(*) is the BatchNormalization batch normalization operation, MaxPooling(*) is the maximum pooling operation, Attention(*) represents the attention module processing operation, F output It is the output of the feature high-dimensional projection module PM.

[0032] S22, multi-scale feature fusion module FM with high-dimensional features F total 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 the wind field and the terrain, and output the multi-scale high-dimensional fusion feature F fused , as the input of the first fully connected prediction module. The calculation formula of the multi-scale feature fusion module is as follows:

[0033] T1=MSF(F total )

[0034] T 1_Att =Attention(T1)

[0035] T2=MSF(T 1_Att )

[0036] T 2_Att =Attention(T2)

[0037] T3=MSF(T 2_Att )

[0038] F fused =Attention(T3)

[0039] Among them F total is the input high-dimensional feature, MSF(*) is the feature fusion operation of the MSF module, T1 is the feature fusion of the first MSF layer, T 1_Att is the feature after weight adjustment of the first attention module, T2 is the feature fused by the second MSF layer, and T 2_Attis the feature after weight adjustment of the second attention module, and T3 is the feature T fused by the third MSF layer. 3_Att The feature after weight adjustment of the third attention module is completed through three MSF layers and three attention modules to obtain the fused feature F fused ;

[0040] S23, the first fully connected prediction module uses multi-scale wind-terrain high-dimensional features F fused As input, the fusion feature F is passed through four fully connected activation layers and one fully connected layer. fused Project it into high-dimensional space and generate multi-scale spatial feature extraction network predictions. The network formula is as follows:

[0041] P1=SiLU(FC(Flatten(F fused )))

[0042] P2=SiLU(FC(O1))

[0043] P3=SiLU(FC(O2))

[0044] P4=SiLU(FC(O3))

[0045] P WT =FC(O4)

[0046] Where FC(*) is the Linear layer conversion, SiLU(*) is the SiLU activation operation, Flatten(*) is the flattening operation, P1 is the feature extracted by the first fully connected activation layer, P2 is the feature extracted by the second fully connected activation layer, P3 is the feature extracted by the third fully connected activation layer, P4 is the feature extracted by the fourth fully connected activation layer, and P WT The final prediction of the network is extracted for multi-scale spatial features.

[0047] Furthermore, the attention module in step S21 includes the wind residual space attention mechanism WR-SA and the channel attention mechanism CA; the wind residual space attention mechanism WR-SA is designed to amplify the key spatial position weights, and the formula is as follows:

[0048] A S1 =Input W -Center(W Norm )

[0049] A S2 =Sigmoid(Conv(A S1 ))

[0050] A S3 =MeanPooling(A S2)+Maxpooling(A S2 )

[0051] F out =A S3 ⊙Input

[0052] Where Input represents the input grid data, W Norm Represents the input of standardized typhoon wind field grid data, Center(*) represents taking the grid center value and copying it to the corresponding dimension, Conv(*) is the convolution operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, A S1 Represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, A S2 Represents the weight of each grid position, A S3 Represents the grid position weight after two pooling operations, used to match A S2 Dimension and Input dimension, F out Represents the output after processing by the spatial attention mechanism.

[0053] The wind residual spatial attention mechanism uses the residual of the wind field grid and its central grid as input, dynamically assigns weights to each spatial position, and effectively improves the performance of the network.

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

[0055] A C1 =MeanPooling(Input)

[0056] A C2 =SiLU(FC(A1))

[0057] A C3 =Sigmoid(FC(A2))

[0058] F out =AC3⊙Input

[0059] Where Input represents the input grid data, MeanPooling(*) is the average pooling operation, which is used to average the variables of each channel into a channel response value, FC(*) is the Linear layer conversion, SiLU(*) is the SiLU activation operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, A C1 Represents the channel response value corresponding to each channel, A C2 Represents the high-dimensional channel response features after the linear layer and SiLU activation operation, A C3Represents the weight corresponding to each channel, F out Represents the output after processing by the channel attention mechanism.

[0060] This mechanism enables the model to dynamically assign channel weights based on channel importance, allowing the model to pay more attention to important channels and improve network performance.

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

[0062] Furthermore, each feature fusion module MSF in step S22 consists of three dilated convolutional layers with different dilation rates, three SiLU activation layers, a convolutional fusion module, and an attention module. Each convolutional module contains a convolutional layer. The network formula is as follows:

[0063] S 1.1 =BN(SiLU(D_Conv d＝1 (Input)))

[0064] S 1.2 =BN(SiLU(D_Conv d＝2 (Input)))

[0065] S 1.3 =BN(SiLU(D_Conv d＝3 (Input)))

[0066] S out =Conv(Concatenate(S 1.1 ,S 1.2 ,S 1.3 ))

[0067] Where D_Conv d＝1 Represents a dilated convolution with a dilation rate of 1, D_Conv d＝2 Represents the dilation rate of 2, D_Conv d＝3 represents the dilation rate of 3, S 1.1 is the local scale fusion feature after being processed by the dilation convolution module with a dilation rate of 1, S 1.2 is the medium-scale fusion feature after being processed by the dilation convolution module with a dilation rate of 2, S 1.3 is the large-scale fusion feature after being processed by the dilation rate 3 dilation convolution module, S outIt 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 F total As input, it outputs the multi-scale wind-terrain high-dimensional fusion feature F fused This module uses a dilated convolutional module to capture wind field-terrain correlations at different scales and fuses them, enabling the network to identify changes in wind behavior and subtle changes in terrain, thereby improving the network's prediction capabilities.

[0069] Further, step S3 is specifically as follows:

[0070] S31. Establish a site spatial location graph. Based on meteorological observation site data, treat each site as a node in the graph structure, and use the site's actual observation value, the typhoon wind field value corresponding to the site's location, and the terrain characteristic value corresponding to the site's location as site features. The graph modeling process uses the site's spatial location as the basis for constructing edges between nodes. If the Euclidean distance between any two nodes is less than the set threshold R, an edge is considered to exist between the two nodes. Edges play an important role in the subsequent message transmission mechanism.

[0071] S32. Establish a site feature correlation graph. Based on meteorological observation site data, treat each site as a node in the graph structure, and use the site's actual observation value, the typhoon wind field value corresponding to the site's location, and the terrain feature value corresponding to the site's location as site features. The graph modeling process uses site feature correlation 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 an edge is considered to exist between the two nodes.

[0072] S33. The spatial linkage network for meteorological observation sites includes two different pipeline flows for processing the two modeling graphs. The network takes the modeling graph and node features as input and outputs predictions based on the spatial linkage considerations of the sites. The spatial location pipeline flow formula for processing the spatial location modeling graph is as follows:

[0073] h1=GraphSage(Input)

[0074] h2=Dense(h1)

[0075] Among them, Input represents the network input, GraphSage(*) represents the GraphSage message aggregation layer, and Dense(*) represents the dense fully connected layer. Each GraphSage message aggregation layer consists of an average aggregator, a maximum pooling aggregator, and an adaptive weighted fusion layer. The calculation method is as follows:

[0076]

[0077] Pooli =Maxpooling({SiLU(W·h j +b)|j∈N(i)})

[0078] W m ,W p =Softmax(FC(Meanpooling(Mean i ),Meanpooling(Pool i )))

[0079] H i =W m Mean i +W p Pool i

[0080] Among them, Mean i is the average aggregate feature of the i-th node, N(i) is the number of neighbor nodes of the i-th node, h j is the feature of the jth neighbor node, Pool i is the maximum pooled aggregation feature of the i-th node, W and b represent the trainable weight and bias respectively, W m is the weight of the average aggregator, W p is the weight of the maximum pooling aggregator.

[0081] The message aggregation layer aggregates the average and maximum features of neighboring nodes, and can dynamically assign weights of the two features based on different nodes and different graph characteristics, effectively leveraging the spatial linkage between sites and improving the network's prediction performance.

[0082] The correlation pipeline flow for processing feature correlation modeling graphs consists of a GraphSage message aggregation layer and two graph attention message aggregation layers (GraphATtention, GAT), as follows:

[0083] l1=GraphSage(Input)

[0084] l2=GAT(l1)

[0085] l3=GAT(l3)

[0086] Here, Input represents the network input, GraphSage(*) represents the GraphSage message aggregation layer, and GAT(*) 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. The parameters of each attention head are independent, and the outputs of all attention heads are concatenated. The calculation method is as follows:

[0087]

[0088] Among them, K is the number of heads of multi-head attention, represents the l+1th layer embedded latent variable of the i-th node, σ(*) is the activation operation, N(i) is the number of neighbor nodes of the i-th node, is the normalized attention matrix, W k is the linear transformation weight matrix of the kth attention head.

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

[0090] S34, the second fully connected prediction module takes the outputs h2 and l3 of two different pipeline flows as input, projects the features into a high-dimensional space through four fully connected activation layers and one fully connected layer, and generates output based on site spatial linkage considerations. The network formula is as follows:

[0091] G1=SiLU(FC(Concatenate(h2,l3)))

[0092] G2=SiLU(FC(G1))

[0093] G3=SiLU(FC(G2))

[0094] G4=SiLU(FC(G3))

[0095] P G =FC(G4)

[0096] Among them, G1 is the feature extracted by the first fully connected-activated layer of the second fully connected prediction module, G2 is the feature extracted by the second fully connected-activated layer of the second fully connected prediction module, G3 is the feature extracted by the third fully connected-activated layer of the second fully connected prediction module, and G4 is the feature extracted by the fourth fully connected-activated layer of the second fully connected prediction module.

[0097] Furthermore, step S4 is specifically as follows: the adaptive weighted network is used to generate the weights of the two prediction values, and weighted addition is performed according to the weights to generate the final prediction. The adaptive weighted network has the following characteristics: the adaptive weighted network uses the typhoon information I at the target grid point T , Grid space information I SI , site density information I SD As input data. TIncluding typhoon intensity, typhoon center longitude and latitude, minimum air 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. SI It includes the latitude and longitude of the target grid point, the distance from the typhoon center, the azimuth 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. 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 I SD This includes the number of edges in the site spatial location graph and site feature correlation graph for the node 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:

[0098] D1=Silu(FC(Concatenate(I T ,I SI I SD ))

[0099] D2=Silu(FC(D1))

[0100] D3=Silu(FC(D2))

[0101] D4=Silu(FC(D3))

[0102] D5=Silu(FC(D4))

[0103] D F1 ,D F2 =SoftPlus(FC(D5))

[0104] O Final =(D F1 *P WT +D F2 *P G )

[0105] Among them, D1 is the feature extracted by the first fully connected-activated layer of the adaptive weighted network, D2 is the feature extracted by the second fully connected-activated layer of the adaptive weighted network, D3 is the feature extracted by the third fully connected-activated layer of the adaptive weighted network, D4 is the feature extracted by the fourth fully connected-activated layer of the adaptive weighted network, D5 is the feature extracted by the fifth fully connected-activated layer of the adaptive weighted network, and D F1 The first weight of the first typhoon wind field grid prediction value output by the multi-scale wind-terrain spatial correlation extraction network, D F2 The second weight of the second typhoon wind field grid prediction value output by the spatial linkage network of meteorological observation stations, P WTis the grid prediction value of the first typhoon wind field, P G is the grid prediction value of the second typhoon wind field, P Final For the final prediction.

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

[0107] Data acquisition and standardization preprocessing module, which is responsible for obtaining meteorological observation station data and low-resolution typhoon wind field data W from the data source LR and high-resolution terrain data T HR , and perform standardized preprocessing on wind field and terrain data to generate standardized data;

[0108] The first typhoon wind field grid prediction module inputs the standardized wind field and terrain data into the multi-scale wind-terrain spatial correlation extraction network, and generates the first typhoon wind field grid prediction value based on wind-terrain correlation considerations 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 establish a site spatial location map and a site feature correlation map based on the meteorological observation site data, and input the standardized data obtained by the data acquisition and standardization preprocessing module into the meteorological observation site spatial feature extraction meteorological observation site 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 sites is generated;

[0110] The fusion prediction final typhoon wind field grid module is used to calculate the typhoon information I at the target grid point. T , Grid space information I SI , site density information I SD The adaptive weighted network is input to generate the weights of the first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value, and the two prediction values ​​are weightedly fused 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 deep learning-based typhoon wind field downscaling method for complex terrain areas when executing the computer program.

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

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

[0114] 1. The typhoon wind field downscaling method for complex terrain areas based on deep learning disclosed in the present invention is intended to solve the problems of insufficient accuracy and poor generalization ability of existing technologies under complex terrain conditions. First, low-resolution engineering typhoon wind field data, high-resolution terrain data, and site observation data are obtained and standardized data is formed into model input after standardization processing. Then, a multi-scale wind-terrain spatial correlation extraction network based on a 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 into the multi-scale wind-terrain spatial correlation extraction network to help the model focus on important grid positions and important variables. Subsequently, a spatial linkage network of meteorological observation stations based on a graph neural network was used to extract the spatial correlation between the target grid points and the meteorological observation stations. The network inputs the meteorological observation station data and outputs the wind speed and direction forecasts for the target grid points. Finally, an adaptive weighted network was used to perform a weighted addition of the predictions based on the two different considerations based on the typhoon characteristics, the spatial location information of the target grid points, and the station distribution density in the area where the target grid points are located to generate the final wind field grid. During the training and optimization process of the multi-scale wind-terrain spatial correlation extraction network model, the backpropagation method was used to feedback correct the model parameters to improve prediction accuracy and reduce errors. The grid comprehensively considers the correlation between the typhoon wind field and the terrain and the spatial linkage between meteorological observation stations, automatically adapting to different typhoon conditions, different geographical locations, and different station distribution densities.

[0115] 2. The present invention discloses a method for downscaling typhoon wind fields in complex terrain areas based on deep learning. It uses two different network structures to capture the wind-terrain spatial correlation and the spatial linkage between meteorological observation stations, respectively, and makes predictions based on the two correlations. Then, an adaptive weighted network is used to dynamically weight the predictions of the two networks, thereby obtaining the final wind field downscaling results. The present invention comprehensively considers the correlation between the near-surface wind field and the terrain and the spatial linkage between stations, and automatically adapts to different typhoon conditions, geographical locations, and distribution densities of meteorological observation stations. It not only improves the accuracy of typhoon wind field downscaling, but also enhances the robustness of the downscaling effect. It has important application value for typhoon mechanism research and disaster prevention, and can meet the typhoon wind field downscaling tasks in various scenarios.

[0116] 3. The typhoon wind field downscaling method for complex terrain areas based on deep learning disclosed in the present invention is a typhoon wind field downscaling method used in typhoon research and typhoon disaster forecasting and prevention. By processing typhoon wind field data in different geographical regions, station distribution densities, and typhoon conditions through a deep learning network, high-resolution typhoon wind field maps are generated quickly and accurately, and the accuracy and robustness of the typhoon wind field grid are improved, which is of great significance to the fields of typhoon research, typhoon forecasting, disaster prevention, etc. 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 scenarios. In addition, the implementation of this 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 meteorology and disaster fields.

[0117] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0119] Figure 1 This is a flow chart of the method for downscaling typhoon wind fields in complex terrain areas based on deep learning of the present invention;

[0120] Figure 2 Flowchart of step S2 of the present invention for constructing a multi-scale wind-topography spatial correlation network;

[0121] Figure 3 This is a flow chart of step S3 of the present invention for constructing a spatial linkage network of meteorological observation sites;

[0122] Figure 4 This is a flow chart of step S4 of the present invention for constructing an adaptive weighted network. DETAILED DESCRIPTION

[0123] The following describes the embodiments of the present invention through specific examples. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention.

[0124] like Figure 1A typhoon wind field downscaling method for complex terrain areas based on deep learning is shown, comprising the following steps:

[0125] S1. Data acquisition and standardization preprocessing; obtain meteorological observation station data and low-resolution typhoon wind field data W LR and high-resolution terrain data T HR , the obtained engineering typhoon model wind field data and terrain data are standardized and preprocessed to generate standardized data; the standardized wind field data and standardized terrain data are represented as W Norm and T Norm ;

[0126] The data normalization process is:

[0127] W Norm =Norm(W LR )

[0128] T Norm =Norm(T HR )

[0129] Among them, Norm(*) represents the standardization operation, W LR Represents the input low-resolution typhoon wind field data, T HR Represents the input high-resolution terrain data, W Norm represents the standardized typhoon wind field data, T Norm Represents normalized terrain data.

[0130] S2, the first typhoon wind field grid forecast; Figure 2 As shown, the standardized wind field and terrain data are input into the 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 predictions based on wind-terrain correlation considerations;

[0131] The multi-scale wind-terrain spatial correlation extraction network inputs the typhoon wind field grid and terrain grid centered on the target grid, and outputs the wind speed and wind direction prediction of the target grid. The multi-scale wind-terrain spatial correlation extraction network is used to generate predictions based on wind-terrain correlation considerations. The multi-scale wind-terrain spatial correlation extraction network consists of a feature high-dimensional projection module PM, a multi-scale high-dimensional feature fusion module FM, and a first fully connected prediction module. The standardized wind field and terrain data are input into the feature high-dimensional projection module PM, and the feature high-dimensional projection module PM outputs the total high-dimensional features F of the wind field and terrain. total The typhoon wind field and terrain total high-dimensional features F total Input the multi-scale feature fusion module FM and output the multi-scale wind-terrain high-dimensional fusion feature F fused The first fully connected prediction module uses the multi-scale wind-terrain high-dimensional fusion feature Ffused Make a prediction and output the first typhoon wind field grid prediction value; specifically:

[0132] S21, the feature high-dimensional projection module PM takes the standardized typhoon wind field data and terrain data as input, and processes the typhoon wind field data and terrain data into high-dimensional features F total , and then the high-dimensional feature F total The spliced ​​output is used as the input to the multi-scale high-dimensional feature fusion module FM. The calculation formula of the feature high-dimensional projection module PM is as follows:

[0133] F W =PM(W Norm )

[0134] F T =PM(T Norm )

[0135] F total =Concatenate(F W ,F T )

[0136] Among them, PM(*) represents the feature high-dimensional projection module, F W represents the high-dimensional typhoon wind field characteristics after being processed by the feature high-dimensional projection module, F T represents the high-dimensional terrain features processed by the feature high-dimensional projection module, Concatenate(*) represents the splicing operation, and F total Represents the overall high-dimensional characteristics of typhoon wind field and terrain.

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

[0138] F1=MaxPooling(BN(SiLU(Conv(Input))))

[0139] F2=MaxPooling(BN(SiLU(Conv(F1))))

[0140] F3=BN(SiLU(Conv(F2)))

[0141] F output =Attention(SiLU(Conv(F3)))

[0142] Among them, Input represents the input of the feature extraction module PM, Conv(*) is the convolution operation, SiLU(*) is the SiLU activation operation, BN(*) is the BatchNormalization batch normalization operation, MaxPooling(*) is the maximum pooling operation, Attention(*) represents the attention module processing operation, F output It is the output of the feature high-dimensional projection module PM.

[0143] The attention module includes the wind residual spatial attention mechanism WR-SA and the channel attention mechanism CA. The wind residual spatial attention mechanism WR-SA (Wind Residual-Spatial Attention) aims to amplify the weight of key spatial positions. The formula is as follows:

[0144] A S1 =Input W -Center(W Norm )

[0145] A S2 =Sigmoid(Conv(A S1 ))

[0146] A S3 =MeanPooling(A S2 )+Maxpooling(A S2 )

[0147] F out =A S3 ⊙Input

[0148] Where Input represents the input grid data, W Norm Represents the input of standardized typhoon wind field grid data, Center(*) represents taking the grid center value and copying it to the corresponding dimension, Conv(*) is the convolution operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, A S1 Represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, A S2 Represents the weight of each grid position, A S3 Represents the grid position weight after two pooling operations, used to match A S2 Dimension and Input dimension, F out Represents the output after processing by the spatial attention mechanism.

[0149] The wind residual spatial attention mechanism uses the residual of the wind field grid and its central grid as input, dynamically assigns weights to each spatial position, and effectively improves the performance of the network.

[0150] The channel attention mechanism CA (ChannelAttention) aims to amplify the key channel weights. The formula is as follows:

[0151] A C1 =MeanPooling(Input)

[0152] A C2 =SiLU(FC(A1))

[0153] A C3 =Sigmoid(FC(A2))

[0154] F out =A C3 ⊙Input

[0155] Where Input represents the input grid data, MeanPooling(*) is the average pooling operation, which is used to average the variables of each channel into a channel response value, FC(*) is the Linear layer conversion, SiLU(*) is the SiLU activation operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, A C1 Represents the channel response value corresponding to each channel, A C2 Represents the high-dimensional channel response features after the linear layer and SiLU activation operation, A C3 Represents the weight corresponding to each channel, F out Represents the output after processing by the channel attention mechanism.

[0156] This mechanism enables the model to dynamically assign channel weights based on channel importance, allowing the model to pay more attention to important channels and improve network performance.

[0157] The feature high-dimensional projection module receives typhoon wind field data and terrain data as input and outputs the corresponding high-dimensional features F total This network is concise and efficient in design. It can process typhoon wind field data and terrain data into corresponding high-dimensional features, and uses the attention mechanism to amplify the weights of key channels and key spatial positions, which helps the network capture the interaction between typhoon wind field and terrain.

[0158] S22, multi-scale feature fusion module FM with high-dimensional features F total 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 the wind field and the terrain, and output the multi-scale high-dimensional fusion feature F fused , as the input of the first fully connected prediction module. The calculation formula of the multi-scale feature fusion module is as follows:

[0159] T1=MSF(F total )

[0160] T 1_Att =Attention(T1)

[0161] T2=MSF(T 1_Att )

[0162] T 2_Att =Attention(T2)

[0163] T3=MSF(T 2_Att )

[0164] F fused =Attention(T3)

[0165] Among them F total is the input high-dimensional feature, MSF(*) is the feature fusion operation of the MSF module, T1 is the feature fusion of the first MSF layer, T 1_Att is the feature after weight adjustment of the first attention module, T2 is the feature fused by the second MSF layer, and T 2_Att is the feature after weight adjustment of the second attention module, and T3 is the feature T fused by the third MSF layer. 3_Att The feature after weight adjustment of the third attention module is completed through three MSF layers and three attention modules to obtain the fused feature F fused .

[0166] Each feature fusion module MSF consists of three dilated convolutional layers with different dilation rates, three SiLU activation layers, a convolutional fusion module, and an attention module. Each convolutional module contains convolutional layers. The network formula is as follows:

[0167] S 1.1 =BN(SiLU(D_Conv d＝1 (Input)))

[0168] S 1.2 =BN(SiLU(D_Conv d＝2 (Input)))

[0169] S 1.3 =BN(SiLU(D_Conv d＝3 (Input)))

[0170] S out =Conv(Concatenate(S 1.1 ,S 1.2 ,S 1.3 ))

[0171] Where D_Conv d＝1 Represents a dilated convolution with a dilation rate of 1, D_Conv d＝2 Represents the dilation rate of 2, D_Conv d＝3 represents the dilation rate of 3, S 1.1 is the local scale fusion feature after being processed by the dilation convolution module with a dilation rate of 1, S 1.2 is the medium-scale fusion feature after being processed by the dilation convolution module with a dilation rate of 2, S 1.3 is the large-scale fusion feature after being processed by the dilation rate 3 dilation convolution module, S out It is the multi-scale high-dimensional fusion feature output by the MSF layer.

[0172] The multi-scale high-dimensional feature fusion module FM receives the high-dimensional feature F total As input, it outputs the multi-scale wind-terrain high-dimensional fusion feature F fused This module uses a dilated convolutional module to capture wind field-terrain correlations at different scales and fuses them, enabling the network to identify changes in wind behavior and subtle changes in terrain, thereby improving the network's prediction capabilities.

[0173] S23, the first fully connected prediction module uses multi-scale wind-terrain high-dimensional features F fused As input, the fusion feature F is passed through four fully connected activation layers and one fully connected layer. fused Project it into high-dimensional space and generate multi-scale spatial feature extraction network predictions. The network formula is as follows:

[0174] P1=SiLU(FC(Flatten(F fused )))

[0175] P2=SiLU(FC(O1))

[0176] P3=SiLU(FC(O2))

[0177] P4=SiLU(FC(O3))

[0178] P WT =FC(O4)

[0179] Where FC(*) is the Linear layer conversion, SiLU(*) is the SiLU activation operation, Flatten(*) is the flattening operation, P1 is the feature extracted by the first fully connected activation layer, P2 is the feature extracted by the second fully connected activation layer, P3 is the feature extracted by the third fully connected activation layer, P4 is the feature extracted by the fourth fully connected activation layer, and P WT The final prediction of the network is extracted for multi-scale spatial features.

[0180] This multi-scale spatial feature extraction network takes typhoon wind field data and terrain data as input and outputs predictions based on wind-terrain correlation. This network is concise and effective, capturing the interactions between wind fields and terrain at different scales. It also leverages an attention mechanism to amplify the weights of key channels and spatial locations, generating an accurate near-surface wind field grid.

[0181] S3, the second typhoon wind field grid forecast; Figure 3 As shown, a site spatial location map and a site feature correlation map are established based on the meteorological observation site data. The two 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. The prediction of the site-free area is generated using real site observations to guide the generation of a high-quality near-surface wind field grid.

[0182] The spatial linkage network of meteorological observation sites includes spatial position correlation pipeline flow, feature correlation pipeline flow and the second fully connected prediction module. Spatial position correlation pipeline flow and feature correlation pipeline flow output site spatial information feature I SI The second fully connected prediction module uses the grid space information feature I SI Make a prediction and output the grid prediction value of the second typhoon wind field; specifically:

[0183] S31. Establish a site spatial location graph. Based on meteorological observation site data, treat each site as a node in the graph structure, and use the site's actual observation value, the typhoon wind field value corresponding to the site's location, and the terrain characteristic value corresponding to the site's location as site features. The graph modeling process uses the site's spatial location as the basis for constructing edges between nodes. If the Euclidean distance between any two nodes is less than the set threshold R, an edge is considered to exist between the two nodes. Edges play an important role in the subsequent message transmission mechanism.

[0184] S32. Establish a site feature correlation graph. Based on meteorological observation site data, treat each site as a node in the graph structure, and use the site's actual observation value, the typhoon wind field value corresponding to the site's location, and the terrain feature value corresponding to the site's location as site features. The graph modeling process uses site feature correlation 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 an edge is considered to exist between the two nodes.

[0185] S33. The spatial linkage network for meteorological observation sites includes two different pipeline flows for processing the two modeling graphs. The network takes the modeling graph and node features as input and outputs predictions based on the spatial linkage considerations of the sites. The spatial location pipeline flow formula for processing the spatial location modeling graph is as follows:

[0186] h1=GraphSage(Input)

[0187] h2=Dense(h1)

[0188] Among them, Input represents the network input, GraphSage(*) represents the GraphSage message aggregation layer, and Dense(*) represents the dense fully connected layer. Each GraphSage message aggregation layer consists of an average aggregator, a maximum pooling aggregator, and an adaptive weighted fusion layer. The calculation method is as follows:

[0189]

[0190] Pool i =Maxpooling({SiLU(W·h j +b)|j∈N(i)})

[0191] W m ,W p =Sofmtax(FC(Meanpooling(Mean i ),Meanpooling(Pool i )))

[0192] H i =W m Mean i +W p Pool i

[0193] Among them, Mean i is the average aggregate feature of the i-th node, N(i) is the number of neighbor nodes of the i-th node, h j is the feature of the jth neighbor node, Pool i is the maximum pooled aggregate feature of the i-th node, W and b represent the trainable weight and bias respectively, W m is the average aggregator weight, W p is the weight of the maximum pooling aggregator.

[0194] The message aggregation layer aggregates the average and maximum features of neighboring nodes, and can dynamically assign weights of the two features based on different nodes and different graph characteristics, effectively leveraging the spatial linkage between sites and improving the network's prediction performance.

[0195] The correlation pipeline flow for processing feature correlation modeling graphs consists of a GraphSage message aggregation layer and two graph attention message aggregation layers (GraphATtention, GAT), as follows:

[0196] l1=GraphSage(Input)

[0197] l2=GAT(l1)

[0198] l3=GAT(l3)

[0199] Here, Input represents the network input, GraphSage(*) represents the GraphSage message aggregation layer, and GAI(*) 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. The parameters of each attention head are independent, and the outputs of all attention heads are concatenated. The calculation method is as follows:

[0200]

[0201] Among them, K is the number of heads of multi-head attention, represents the l+1th layer embedded latent variable of the i-th node, σ(*) is the activation operation, N(i) is the number of neighbor nodes of the i-th node, is the normalized attention matrix, W k is the linear transformation weight matrix of the kth attention head.

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

[0203] S34, the second fully connected prediction module takes the outputs h2 and l3 of two different pipeline flows as input, projects the features into a high-dimensional space through four fully connected activation layers and one fully connected layer, and generates output based on site spatial linkage considerations. The network formula is as follows:

[0204] G1=SiLU(FC(Concatenate(h2,l3)))

[0205] G2=SiLU(FC(G1))

[0206] G3=SiLU(FC(G2))

[0207] G4=SiLU(FC(G3))

[0208] P G =FC(G4)

[0209] Among them, G1 is the feature extracted by the first fully connected-activated layer of the second fully connected prediction module, G2 is the feature extracted by the second fully connected-activated layer of the second fully connected prediction module, G3 is the feature extracted by the third fully connected-activated layer of the second fully connected prediction module, and G4 is the feature extracted by the fourth fully connected-activated layer of the second fully connected prediction module.

[0210] The network takes station data, wind field data, and terrain data as input and outputs predictions based on spatial linkage between the stations. This network is able to capture the implicit correlations between target grid points and stations, thereby using real station observations to guide the generation of predictions for areas without stations, resulting in high-quality near-surface wind field grids.

[0211] S4, fusion prediction of the final typhoon wind field grid; Figure 4 As shown in the figure, typhoon information I is obtained at the target grid point. T , Grid space information I SI , site density information I SD Input the adaptive weighted 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. Based on the first weight and the second weight, perform weighted fusion on the first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value to obtain the final typhoon wind field grid prediction value. Specifically:

[0212] The adaptive weighted network is used to generate the weights of the two prediction values ​​and perform weighted addition according to the weights to generate the final prediction. The adaptive weighted network has the following characteristics: The adaptive weighted network uses the typhoon information I at the target grid point T , Grid space information I SI , site density information I SD As input data. T Including typhoon intensity, typhoon center longitude and latitude, minimum air 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. SI It includes the latitude and longitude of the target grid point, the distance from the typhoon center, the azimuth 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. 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 I SD This includes the number of edges in the site spatial location graph and site feature correlation graph for the node 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:

[0213] D1=Silu(FC(Concatenate(I T ,I SI I SD ))

[0214] D2=Silu(FC(D1))

[0215] D3=Silu(FC(D2))

[0216] D4=Silu(FC(D3))

[0217] D5=Silu(FC(D4))

[0218] D F1 ,D F2 =SoftPlus(FC(D5))

[0219] O Final =(D F1 *P WT +D F2 *P G )

[0220] Among them, D1 is the feature extracted by the first fully connected-activated layer of the adaptive weighted network, D2 is the feature extracted by the second fully connected-activated layer of the adaptive weighted network, D3 is the feature extracted by the third fully connected-activated layer of the adaptive weighted network, D4 is the feature extracted by the fourth fully connected-activated layer of the adaptive weighted network, D5 is the feature extracted by the fifth fully connected-activated layer of the adaptive weighted network, and D F1 The first weight of the first typhoon wind field grid prediction value output by the multi-scale wind-terrain spatial correlation extraction network, D F2 The second weight of the second typhoon wind field grid prediction value output by the spatial linkage network of meteorological observation stations, P WT is the grid prediction value of the first typhoon wind field, P G is the grid prediction value of the second typhoon wind field, P Final For the final prediction.

[0221] The network can dynamically adjust the prediction value weights of the multi-scale spatial feature extraction network and the site spatial linkage network according to the geographical area and site distribution density of the target grid point and the typhoon conditions. It can not only improve the prediction performance of the network, but also improve the robustness of the network, ensure the stability of the prediction results, and generate high-quality near-surface typhoon wind fields.

[0222] The training process for extracting multi-scale wind-terrain spatial correlations based on convolutional neural networks includes model training, validation, and testing. The model training phase involves forward propagation and backpropagation, optimizing network parameters through a loss function to achieve model learning. During the model validation phase, the model's generalization is assessed using a five-fold cross-validation method, and model hyperparameters are adjusted based on the results. During the model testing phase, the trained model's performance is verified using independent validation samples to ensure robustness and accuracy on unknown data.

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

[0224] The proposed near-surface wind field downscaling method can efficiently and effectively perform downscaling tasks, providing strong technical support for research and application in related fields. Through this method, users can obtain clearer and more accurate wind field maps, enabling more effective scientific research and decision-making.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention 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 invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A typhoon wind field downscaling method for complex terrain areas based on deep learning, characterized by: The following steps are involved: S1. Data acquisition and standardization preprocessing; the acquired meteorological observation station data and low-resolution typhoon wind field data and high-resolution terrain data Standardization preprocessing to generate standardized data; S2, the first typhoon wind field grid forecast; the standardized wind field and terrain data are input into the multi-scale wind-terrain spatial feature to extract the typhoon wind field and terrain network, and the feature high-dimensional projection module is used to , multi-scale high-dimensional feature fusion module FM, the first fully connected prediction module generates wind speed and direction predictions based on wind-terrain correlation considerations; among them, the standardized wind field and terrain data input feature high-dimensional projection module , feature high-dimensional projection module Output wind field and terrain total high-dimensional features ; Overall high-dimensional characteristics of typhoon wind field and terrain Input the multi-scale feature fusion module FM and output the multi-scale wind-terrain high-dimensional fusion feature ; The first fully connected prediction module uses multi-scale wind-terrain high-dimensional fusion features Make predictions; S3, Grid forecast of the second typhoon wind field; Establish a site spatial location map and a site feature correlation map based on the meteorological observation site data, input the two site data maps and the standardized data obtained in step S1 into the meteorological observation site spatial feature extraction meteorological observation site spatial linkage network, and generate wind speed and wind direction forecasts considering the spatial linkage between meteorological observation sites through the spatial location correlation pipeline flow, feature correlation pipeline flow and the second fully connected prediction module; The spatial location correlation pipeline flow and feature correlation pipeline flow output site spatial information features. The second fully connected prediction module uses the grid space information features Make predictions; S4, fusion prediction of the final typhoon wind field grid; typhoon information at the target grid point , grid space information , site density information Input the adaptive weighted 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. 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 weightedly fused to obtain the final typhoon wind field grid prediction value.

2. The method for downscaling typhoon wind fields in complex terrain areas according to claim 1, characterized in that: The data normalization process in step S1 is: in, Represents standardized operations. Represents the input low-resolution typhoon wind field data, represents the input high-resolution terrain data, represents the standardized typhoon wind field data, Represents normalized terrain data.

3. The method for downscaling typhoon wind fields in complex terrain areas according to claim 2, characterized in that: Step S2 is specifically as follows: S21, feature high-dimensional projection module Taking 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 After splicing, the output is used 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: in, Represents the feature high-dimensional projection module, Represents the high-dimensional typhoon wind field characteristics after being processed by the feature high-dimensional projection module, Represents the high-dimensional terrain features after being processed by the feature high-dimensional projection module, Represents the splicing operation, Represents the overall high-dimensional characteristics of the typhoon wind field and terrain; Each feature high-dimensional projection module It consists of three convolution modules, two Maxpooling pooling layers and an attention layer. Each convolution module contains a convolution layer and a Silu activation layer. Each feature high-dimensional projection module The calculation process is: in, Representative feature extraction module Input, is the convolution operation, Activate the SiLU operation. It is the Batch Normalization batch normalization operation. is the maximum pooling operation, Represents the attention module processing operation, It is the feature high-dimensional projection module Output; S22, multi-scale feature fusion module FM with 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 the wind field and the terrain, and output multi-scale high-dimensional fusion features. ; The calculation formula of the multi-scale feature fusion module FM is as follows: in To input high-dimensional features, for The feature fusion operation of the module, is the feature fused by the first MSF layer, is the feature after weight adjustment of the first attention module, is the feature fused by the second MSF layer, is the feature after weight adjustment of the second attention module, The features fused by the third MSF layer The feature after weight adjustment of the third attention module is completed through three MSF layers and three attention modules to obtain the fused feature ; S23, the first fully connected prediction module uses multi-scale wind-terrain high-dimensional features As input, the features are fused through four fully connected activation layers and one fully connected layer Project it into high-dimensional space and generate multi-scale spatial feature extraction network predictions; the network formula is as follows: in For Linear layer conversion, Activate the SiLU operation. For the flattening operation, The features extracted by the first fully connected activation layer, The features extracted by the second fully connected activation layer, The features extracted by the third fully connected-activated layer, is the feature extracted by the fourth fully connected-activated layer, The final prediction of the network is extracted for multi-scale spatial features.

4. The method for downscaling typhoon wind fields in complex terrain areas according to claim 3, characterized in that: The attention module in step S21 includes the wind residual space attention mechanism WR-SA and the channel attention mechanism CA; the wind residual space attention mechanism WR-SA is designed to amplify the key spatial position weights, and the formula is as follows: in Represents the input grid data, Represents the input standardized typhoon wind field grid data, Represents taking the center value of the grid and copying it as The corresponding dimensions, is the convolution operation, for Activate the operation, represents the dot product operation, Represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, represents the weight of each grid position, Represents the grid position weight after two pooling operations, used for matching Dimensions and Dimension, Represents the output after processing by the spatial attention mechanism; The channel attention mechanism CA aims to amplify the key channel weights, and the formula is as follows: in Represents the input grid data, It is an average pooling operation, which is used to average the variables of each channel into a channel response value. For Linear layer conversion, Activate the SiLU operation. for Activate the operation, represents the dot product operation, Represents the channel response value corresponding to each channel, Represents the linear layer and The high-dimensional channel response features after activation operation processing, Represents the weight corresponding to each channel, Represents the output after processing by the channel attention mechanism.

5. The method for downscaling typhoon wind fields in complex terrain areas according to claim 4, characterized in that: In step S22, each feature fusion module MSF consists of three dilated convolution layers with different expansion rates, three SiLU activation layers, a convolution fusion module, and an attention module. Each convolution module contains a convolution layer, and the network formula is as follows: in represents a dilated convolution with a dilation rate of 1, represents a dilated convolution with a dilation rate of 2, represents a dilated convolution with a dilation rate of 3, is the local scale fusion feature after being processed by the dilation convolution module with a dilation rate of 1. is the medium-scale fusion feature after being processed by the dilation rate 2 dilation convolution module. It is the large-scale fusion feature after being processed by the dilation rate 3 dilation convolution module. It is the multi-scale high-dimensional fusion feature output by the MSF layer.

6. The method for downscaling typhoon wind fields in complex terrain areas according to claim 5, characterized in that: Step S3 is specifically as follows: S31, establishing a site spatial location map using the spatial locations of meteorological observation sites as a basis for constructing edges between nodes. If the longitude and latitude Euclidean distance between any two nodes is less than a set threshold R, it is considered that an edge exists between the two nodes; S32, establishing a site feature correlation graph based on the correlation of meteorological observation site features as the basis for constructing 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; S33, Meteorological Observation Site Spatial Linkage Network The network takes the modeling graph and node features as input and outputs a prediction based on the spatial linkage of the sites. The spatial location pipeline flow formula used to process the spatial location modeling graph is as follows: in, represents the network input, represent Message aggregation layer, represents a dense fully connected layer; each The message aggregation layer consists of an average aggregator, a maximum pooling aggregator, and an adaptive weighted fusion layer. The calculation method is as follows: in, is the average aggregate 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 maximum pooled aggregate feature of the i-th node, W and b represent the trainable weight and bias respectively, is the average aggregator weight, is the weight of the maximum pooling aggregator; The correlation pipeline flow for processing feature correlation modeling graphs consists of a Message aggregation layer and two GAT graph attention message aggregation layers, the formula is as follows: in, represents the network input, represent Message aggregation layer, Represents the graph attention message aggregation layer; each GAT layer contains a multi-head attention mechanism that can dynamically weight the messages transmitted by neighboring nodes according to the importance of neighboring nodes; the parameters of each attention head are independent, and the outputs of all attention heads are connected The calculation method is as follows: Among them, K is the number of heads of multi-head attention, represents the l+1th layer embedded latent variable of the i-th node, To activate the 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 kth attention head; S34, the second fully connected prediction module outputs two different pipeline flows and As input, the features are projected into a high-dimensional space through four fully connected activation layers and one fully connected layer, and the output is generated based on the site space linkage considerations; the network formula is as follows: in The 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, Features 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 according to claim 6, characterized in that: Step S4 is specifically as follows: the adaptive weighted network uses the typhoon information at the target grid point , grid space information , site density information As input data; the typhoon information Including typhoon intensity, typhoon center longitude and latitude, minimum air pressure near the typhoon center, two-minute average maximum sustained wind near the tropical cyclone center, two-minute average sustained wind speed, typhoon landing time; the grid spatial information Including the latitude and longitude of the target grid point, the distance from the typhoon center, the azimuth 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. 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 This includes the number of edges in the site spatial location graph and site feature correlation graph for the node 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, The features extracted from the first fully connected activation layer of the adaptive weighted network, The features extracted by the second fully connected activation layer of the adaptive weighted network, The features extracted by the third fully connected-activated layer of the adaptive weighted network, The features extracted from the fourth fully connected activation layer of the adaptive weighted network, The features extracted from the fifth fully connected activation layer of the adaptive weighted network, The first weight of the first typhoon wind field grid prediction value output by the multi-scale wind-terrain spatial correlation extraction network, The second weight of the second typhoon wind field grid prediction value output by the spatial linkage network of meteorological observation stations, is the grid prediction value of the first typhoon wind field, is the grid forecast value of the second typhoon wind field, For the final prediction.

8. A typhoon wind field downscaling system for complex terrain areas based on deep learning, characterized by: include: Data acquisition and standardization preprocessing module, which is responsible for obtaining meteorological observation station data and low-resolution typhoon wind field data from the data source and high-resolution terrain data , and perform standardized preprocessing on wind field and terrain data to generate standardized data; The first typhoon wind field grid prediction module inputs the standardized wind field and terrain data into the multi-scale wind-terrain spatial correlation extraction network, and generates the first typhoon wind field grid prediction value based on wind-terrain correlation considerations through the feature high-dimensional projection module PM, the multi-scale high-dimensional feature fusion module FM and the first fully connected prediction module; The second typhoon wind field grid prediction module will establish a site spatial location map and a site feature correlation map based on the meteorological observation site data, and input the standardized data obtained by the data acquisition and standardization preprocessing module into the meteorological observation site spatial feature extraction meteorological observation site 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 sites is generated; The fusion prediction final typhoon wind field grid module is used to calculate the typhoon information at the target grid point. , grid space information , site density information The adaptive weighted network is input to generate the weights of the first typhoon wind field grid prediction value and the second typhoon wind field grid prediction value, and the two prediction values ​​are weightedly fused to obtain the final typhoon wind field grid prediction value.

9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method for downscaling typhoon wind fields in complex terrain areas based on deep learning as described in any one of claims 1 to 7 when executing a computer program.

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