A deep learning-based near-surface wind field downscaling method

Abstract

The application relates to a deep learning-based near-surface wind field downscaling method, which comprises the following steps: S1, acquiring site data, wind field data and terrain data, and a feature extraction module outputs total high-dimensional features of the wind field and the terrain; S2, the total high-dimensional features of the wind field and the terrain are input into a multi-scale feature fusion module to output multi-scale fusion features; S3, a first near-surface wind field grid prediction value is output; S4, a second fully-connected prediction module outputs a second near-surface wind field grid prediction value; S5, adaptive weighting analysis data is acquired, the adaptive weighting analysis data is input into an adaptive weighting network, and a final surface wind field grid prediction value is obtained based on a first weight and a second weight. Compared with the prior art, the application has the advantages of improving the accuracy and efficiency of near-surface wind field downscaling.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a deep learning-based method for downscaling near-surface wind fields. Background Technology

[0002] Downscaling of near-surface wind fields is crucial for disaster prevention and wind energy utilization. Especially in the absence of high-resolution near-surface wind field grids, downscaling these grids allows for the rapid and efficient generation of high-quality, high-resolution near-surface wind field grids. This enables timely wind speed forecasting, laying a solid foundation for wind power generation and facilitating more efficient wind energy utilization. Furthermore, wind forecasting allows for timely warnings of wind-related disasters, reducing losses caused by these events.

[0003] However, current traditional downscaling methods are limited by speed and effectiveness. In particular, the accuracy of wind field downscaling under highly complex terrain conditions is not high, the downscaling speed is slow, the quality and availability of near-surface wind fields are not high, and it is impossible to obtain downscaling results in a timely and accurate manner. Summary of the Invention

[0004] This invention provides a deep learning-based method for downscaling near-surface wind fields, aiming to address the problems of slow processing speed and unsatisfactory downscaling effects in existing technologies. This method achieves rapid and accurate downscaling of near-surface wind fields, providing high-quality, high-resolution near-surface wind fields suitable for disaster prevention and wind energy utilization.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] Compared with the prior art, the present invention has the following beneficial effects:

[0007] This invention captures the spatial correlation between wind and topography and the spatial linkage between meteorological stations through two different network structures. Then, an adaptive weighted network is used to dynamically weight the predictions of the two networks to obtain the final wind field downscaling result. It comprehensively considers the correlation between near-surface wind field and topography and the spatial linkage between stations, and automatically adapts to geographical location and station distribution density. This not only improves the accuracy of wind field downscaling, but also enhances the robustness of the downscaling effect. It has important application value for disaster prevention and wind energy utilization, and can meet the near-surface wind field downscaling tasks in various scenarios.

[0008] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0009] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0010] Figure 1 This is a flowchart of the present invention;

[0011] Figure 2 This is a structural diagram of the multi-scale spatial feature extraction network of the present invention;

[0012] Figure 3 This is a structural diagram of the site spatial linkage network of the present invention;

[0013] Figure 4 This is a structural diagram of an adaptive weighted network. Detailed Implementation

[0014] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0015] This invention provides a deep learning-based method for downscaling near-surface wind fields, the flowchart of which is shown below. Figure 1As shown, this invention aims to address the problems of slow processing speed and unsatisfactory downscaling effects in existing technologies. This method achieves rapid and accurate downscaling of near-surface wind fields. Firstly, it provides high-quality, high-resolution near-surface wind fields, applicable to disaster prevention and wind energy utilization. It utilizes a multi-scale spatial feature extraction network based on convolutional neural networks to extract the spatial correlation between the wind field grid and the terrain grid. This network takes the wind field grid and terrain grid centered on the target grid point as input and outputs predicted wind speed and direction for the target grid point. Secondly, it utilizes a site spatial linkage network based on graph neural networks to extract the spatial correlation between stations. This network takes station data as input and outputs predicted wind speed and direction for the target grid point. Based on the spatial location information of the target grid point and the station distribution density of its area, an adaptive weighted network is used to weight and sum the predictions based on the two different considerations, generating the final wind field grid. This grid comprehensively considers the correlation between the near-surface wind field and the terrain, as well as the spatial linkage between stations, automatically adapting to geographical location and station distribution density.

[0016] The method of the present invention includes the following steps:

[0017] S1. Acquire site data, wind field data, and terrain data. Input the wind field data and terrain data into a multi-scale spatial feature extraction network. The multi-scale spatial feature extraction network consists of a feature extraction module, a multi-scale feature fusion module, and a first fully connected prediction module. After the wind field data and terrain data are input into the feature extraction module, the feature extraction module outputs the total high-dimensional features F of the wind field and terrain. total ;

[0018] S2, the total high-dimensional features of the wind field and terrain F total Input multi-scale feature fusion module, output multi-scale fused feature F fused ;

[0019] S3, Multi-scale fusion feature F fused Input the first fully connected prediction module, and the first fully connected prediction module outputs the first near-surface wind field grid prediction value;

[0020] S4. Construct a spatial location map of station data and a correlation map of station data features based on the station data. Input the spatial location map of station data, the correlation map of station data features, and the corresponding station features into a station spatial linkage network. The station spatial linkage network includes a spatial location pipeline flow, a correlation pipeline flow, and a second fully connected prediction module. The outputs of the spatial location pipeline flow and the correlation pipeline flow are input to the second fully connected prediction module. The second fully connected prediction module outputs the second near-surface wind field grid prediction value.

[0021] S5. Obtain adaptive weighted analysis data, input the adaptive weighted analysis data into the adaptive weighted network, generate the first weight corresponding to the first near-surface wind field grid prediction value and the second weight corresponding to the second near-surface wind field grid prediction value, and obtain the final surface wind field grid prediction value based on the first weight and the second weight.

[0022] In this invention, a multi-scale spatial feature extraction network is first used to generate predictions based on wind-topography correlation. This network takes a wind field grid and a topography grid centered on the target grid point as input, and outputs predicted wind speed and direction for the target grid point. The multi-scale spatial feature extraction network is used to generate predictions based on wind-topography correlation. Specifically, the multi-scale spatial feature extraction network has the following characteristics: it consists of a feature extraction module, a multi-scale feature fusion module, and a fully connected prediction module.

[0023] The feature extraction module takes wind field data and terrain data as input, processes them into high-dimensional features, and then concatenates these high-dimensional features as input to the multi-scale feature fusion module. The calculation formula for the feature extraction module is as follows:

[0024] F W =FE(Input) W )

[0025] F T =FE(Input) T )

[0026] F total =Concat(F W ,F T )

[0027] Where FE(*) represents the feature extraction module, and Input W F represents the input wind field data. W Input represents the high-dimensional wind field features after processing by the feature extraction module. T F represents the input terrain data. T This represents the high-dimensional terrain features after processing by the feature extraction module. Concat(*) represents the concatenation operation. total This represents the overall characteristics after merging the high-dimensional features of the wind field and the high-dimensional features of the terrain.

[0028] Each feature extraction (FE) module consists of three convolutional modules, two mean pooling layers, and an attention layer. Each convolutional module contains a convolutional layer and a Silu activation layer. The network formula is as follows:

[0029] F1=MeanPooling(BN(Silu(Conv(Input))))

[0030] F2=MeanPooling(BN(Silu(Conv(F1))))

[0031] F3 = BN(Silu(Conv(F2)))

[0032] F output =Att(Silu(Conv(F3)))

[0033] Where Input represents the input grid data, Conv(*) is the convolution operation, Silu(*) is the Silu activation operation, BN(*) is the Batch Normalization operation, MeanPooling(*) is the average pooling operation, and Att(*) represents the attention mechanism processing operation.

[0034] Attention mechanisms are divided into channel attention mechanisms and spatial attention mechanisms. Channel attention (CA) aims to amplify the weights of key channels, as shown in the following formula:

[0035] A C1 =MeanPooling(Input)

[0036] A C2 =Silu(FC(A1))

[0037] A C3 =Sigmoid(FC(A2))

[0038] F out =A C3 ⊙Input

[0039] Where Input represents the input grid data, MeanPooling(*) is the average pooling operation, used to average the variables of each channel into a single channel response value, FC(*) is the Linear Transformation, Silu(*) is the Silu activation operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, and A C1 A represents the channel response value corresponding to each channel. C2 A represents the high-dimensional channel response features after processing with a linear layer and Silu activation operations. C3 F represents the weight corresponding to each channel. out This represents the output after processing by the channel attention mechanism. 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.

[0040] Spatial Attention (SA) aims to amplify the weights of key spatial locations, as shown in the following formula:

[0041] A S1 =Input W -Center(Input W )

[0042] A S2 =Sigmoid(Conv(A) S1 ))

[0043] A S3 =MeanPooling(A S2 )+Maxpooling(A S2 )

[0044] F out =A S3 ⊙Input

[0045] Where Input represents the input grid data, Input W This represents the input wind field grid data. Center(*) represents taking the grid center value and copying it to the corresponding dimension. Conv(*) is the convolution operation, Silu(*) is the Silu activation operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, and A S1 A represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid. S2 A represents the weight of each grid location. S3 This represents the grid position weights after two pooling operations, used for matching A. S2 Dimensions and Input Dimensions, F out This represents the output after processing by the spatial attention mechanism. This 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's performance.

[0046] The module receives wind field data and terrain data as input and outputs the corresponding high-dimensional feature F. total This network design is simple and efficient, capable of processing wind field 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 wind field and terrain.

[0047] The multi-scale feature fusion module uses high-dimensional features F totalAs input, the high-dimensional wind field features and high-dimensional terrain features are fused through three multi-scale fusion (MSF) layers and one attention layer to capture the correlation between wind field and terrain, and output the multi-scale fused feature F. fused This is used as the input to the fully connected prediction module. The calculation formula for the multi-scale feature fusion module is as follows:

[0048] T1 = MSF(F total )

[0049] T2 = MSF(T1)

[0050] T3 = MSF(T2)

[0051] F fused =Att(T3)

[0052] Where F total For the input high-dimensional features, MSF(*) represents the feature fusion operation of the MSF module. T1 represents the features fused by the first MSF module, T2 represents the features fused by the second MSF module, and T3 represents the features fused by the third MSF module. The feature fusion is completed through the three MSF modules. After the attention layer dynamically adjusts the channel weights and spatial position weights, the fused feature F is obtained. fused .

[0053] Each Feature Fusion Module (MSF) consists of three dilated convolutional modules with different dilation rates, a convolutional fusion module, and an attention module. Each convolutional module contains convolutional layers. The network formula is as follows:

[0054] S 1.1 =BN(Silu(D_Conv) d＝1 (Input)))

[0055] S 1.2 =BN(Silu(D_Conv) d＝2 (Input)))

[0056] S 1.3 =BN(Silu(D_Conv) d＝3 (Input)))

[0057] S2 = Conv(Concat(S 1.1 ,S 1.2 ,S 1.3 ))

[0058] S out =Att(S2)

[0059] Where D_Conv d＝1D_Conv represents a dilated convolution with an inflation rate of 1. d＝2 D_Conv represents a dilated convolution with an inflation rate of 2. d＝3 S represents a dilated convolution with an inflation rate of 3. 1.1 S represents the local scale fusion feature after processing by a dilated convolutional module with a dilation rate of 1. 1.2 S represents the medium-scale fused features after processing by a dilated convolutional module with an inflation rate of 2. 1.3 S1 represents the large-scale fused feature after processing by a dilated convolutional module with an inflation rate of 3, and S2 represents the multi-scale fused feature after fusion by a convolutional fusion module. out This refers to the multi-scale fusion features output by the MSF layer.

[0060] The module receives high-dimensional features F total As input, and output multi-scale fused features F fused This module utilizes dilated convolution to capture wind field-terrain correlations at different scales and fuses them, enabling the network to identify changes in wind behavior and subtle variations in terrain, thereby improving the network's predictive capabilities.

[0061] The fully connected prediction module uses high-dimensional features F fused As input, the features F are fused through four fully connected-activation layers and one fully connected layer. fused The projection is onto a high-dimensional space, and a multi-scale spatial feature extraction network is generated for prediction. The network formula is as follows:

[0062] O1=Silu(FC(Flatten(F fused )))

[0063] O2 = Silu(FC(O1))

[0064] O3 = Silu(FC(O2))

[0065] O4 = Silu(FC(O3))

[0066] O WT =FC(O4)

[0067] Where FC(*) represents the Linear Transformation, Silu(*) represents the Silu Activation Operation, Flatten(*) represents the Flatten Operation, O1 represents the features extracted from the first fully connected-activation layer, O2 represents the features extracted from the second fully connected-activation layer, O3 represents the features extracted from the third fully connected-activation layer, O4 represents the features extracted from the fourth fully connected-activation layer, and O... WT This is the final prediction for the multi-scale spatial feature extraction network.

[0068] This network takes wind field and topographic data as input and outputs predictions based on wind-topographic correlation. The network design is simple yet effective, capturing the interaction between wind field and topography at different scales and generating an accurate near-surface wind field grid.

[0069] The structure of a multi-scale spatial feature extraction network is as follows: Figure 2 As shown, the structure of the multi-scale spatial feature extraction network includes: a feature extraction module for extracting high-dimensional wind field features and high-dimensional terrain features; a multi-scale feature fusion module for capturing the spatial correlation between wind and terrain at different spatial scales; and a fully connected prediction module. Each feature extraction module contains multiple convolutional layers capable of capturing features at different levels; pooling layers to reduce feature dimensionality while retaining important feature information; normalization layers to ensure the stability of the network training process, preventing distribution shifts and gradient explosion / vanishing phenomena; activation layers to achieve complex nonlinear feature extraction, enhancing the network's expressive power; and attention layers to enhance the model's focus on important channels and spatial locations. Each multi-scale convolutional fusion module contains multiple multi-scale fusion layers based on dilated convolutions, capable of capturing wind-terrain correlations at different spatial scales, as well as pooling layers, normalization layers, activation layers, and attention layers. The fully connected prediction module contains multiple linear layers and activation layers, capable of capturing high-dimensional nonlinear relationships between variables. Based on this, a multi-scale spatial feature extraction network is constructed using a feature extraction module, a multi-scale feature fusion module, and a fully connected prediction module to capture the correlation between wind field and terrain, and output wind speed and wind direction predictions.

[0070] Secondly, the site data is first modeled into two types of graphs based on spatial location and feature correlation. Then, a site spatial linkage network is used to generate predictions based on site spatial linkage considerations. The site spatial linkage network has the following characteristics: it employs two different pipeline processing methods to process the graphs constructed in the two different ways. Finally, a fully connected prediction module is used to merge the outputs of the two pipelines for prediction, generating predictions based on site spatial linkage considerations.

[0071] The specific steps for constructing a spatial location map and a correlation map of site data features based on the aforementioned site data are as follows:

[0072] Establish a spatial location map of the station data. Using station data as the data source, each station is considered a node in the graph structure. The actual observed values ​​of the station, the wind field value corresponding to the station's location, and the terrain value corresponding to the station's location are used as station features. The graph modeling process uses the spatial location of the 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 an edge is considered to exist between the two nodes. Edges play an important role in the subsequent message passing mechanism.

[0073] Establish a correlation graph for site data features. Using site data as the data source, each site is considered a node in the graph structure. The actual observed value of the site, the wind field value corresponding to the site's location, and the terrain value corresponding to the site's location are used as site features. The graph modeling process uses the correlation of site features as the basis for constructing edges between nodes. If the cosine similarity of features between any two nodes is greater than a set threshold T, then an edge is considered to exist between the two nodes.

[0074] The multi-scale spatial feature extraction network includes two different pipeline flows to process two types of modeling graphs. The network takes the modeling graph and node features as input and outputs predictions based on site spatial linkage considerations. The formula for the spatial location pipeline flow used to process the spatial location modeling graph is as follows:

[0075] h1 = GraphSage(Input)

[0076] h2 = Dense(h1)

[0077] Where 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 max-pooling aggregator, and an adaptive weighted fusion layer, calculated as follows:

[0078]

[0079] Pool i =Maxpooling({Silu(W·h) j +b)|j∈N(i)})

[0080] W m W p =Softmax(fc(Meanpooling(Mean i Meanpooling(Pool) i )))

[0081] H i =W m ·Mean i +W p ·Pool i

[0082] Among them, Mean i Let h be the average aggregation feature of the i-th node, N(i) be the number of neighboring nodes of the i-th node, and h be the average aggregation feature of the i-th node. j For the characteristics of the j-th neighbor node, Pool iLet W be the max-pooling aggregated feature of the i-th node, and W and b represent the trainable weights and biases, respectively. m W represents the weight of the average aggregator. p The weights assigned to the max-pooling aggregator. This message aggregation layer aggregates the average and maximum features from neighboring nodes and can dynamically allocate the weights of these two features based on different nodes and graph characteristics, effectively leveraging the spatial interoperability between sites and improving the network's predictive performance.

[0083] The correlation pipeline used to handle feature correlation modeling graphs consists of a GraphSage message aggregation layer and two GraphAttention message aggregation layers (GraphATtention, GAT), as shown in the following formula:

[0084] l1 = GraphSage(Input)

[0085] l2 = GAT(l1)

[0086] l3 = GAT(l3)

[0087] 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:

[0088]

[0089] Where K is the number of heads receiving multi-head attention. Let represent the (l+1)th level embedded latent variable of the i-th node, || represent the Concat operation, σ(*) represent the activation operation, and N(i) represent the number of neighboring nodes of the i-th node. For the standardized attention matrix, W k Let be the linear transformation weight matrix of the k-th attention head. Through this multi-head attention mechanism, the network can discover the potential connections between nodes and their neighboring nodes, effectively leveraging the spatial interoperability between sites.

[0090] The second fully connected prediction module takes the outputs h2 and l3 from 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 an output based on site spatial linkage considerations. The network formula is as follows:

[0091] G1 = Silu(FC(Concat(h2,l3)))

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

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

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

[0095] O G =FC(G4)

[0096] Where FC(*) represents the Linear Layer Transformation, Silu(*) represents the Silu Activation Operation, Flatten(*) represents the Flatten Operation, G1 represents the features extracted from the first fully connected-activation layer, G2 represents the features extracted from the second fully connected-activation layer, G3 represents the features extracted from the third fully connected-activation layer, G4 represents the features extracted from the fourth fully connected-activation layer, and O represents the features extracted from the fourth fully connected-activation layer. G This is the final prediction for the spatial interconnection network of sites.

[0097] This network receives site data, wind field data, and topographic data as input and outputs predictions based on spatial correlation considerations between sites. It captures the implicit correlations between sites, thereby using real site observations to guide the generation of predictions for site-free areas, resulting in a high-quality near-surface wind field grid.

[0098] The structure of the site spatial linkage network is as follows Figure 3 As shown, the site spatial linkage network includes: a geographic location modeling graph and a feature correlation modeling graph to represent the spatial relationships between sites, capable of modeling multiple meteorological stations from both geographic location and feature correlation perspectives; two different pipeline flow processing methods to process the graphs constructed by the two different methods, maximizing the discovery of correlations between sites and avoiding the introduction of additional noise; a fully connected prediction module to extract high-dimensional features, realize the high-dimensional transformation of low-dimensional features, enhance the network's expressive power, and output wind speed and wind direction predictions; based on this, the site spatial linkage network is constructed using two site modeling graphs, two different pipeline flow processing methods, and the fully connected prediction module to capture the spatial correlations between sites and output wind speed and wind direction predictions.

[0099] Finally, an adaptive weighted network is used to generate the weights for the two predicted values ​​above, and these weights are then summed to generate the final prediction. The adaptive weighted network has the following characteristics: it uses the latitude and longitude of the target grid point, the number of surrounding stations, timestamps, wind field data, and terrain data as input data. The network consists of five fully connected activation layers, one linear layer, and one SoftPlus activation layer. The network formula is as follows:

[0100] D1 = Silu(FC(Input))

[0101] D2 = Silu(FC(D1))

[0102] D3 = Silu(FC(D2))

[0103] D4 = Silu(FC(D3))

[0104] D5 = Silu(FC(D4))

[0105] D F1 D F2 =SoftPlus(FC(D5))

[0106] O Final =(D F1 *O WT +D F2 *O G )

[0107] Where FC(*) represents the Linear Layer Transformation, Silu(*) represents the Silu Activation Operation, D1 represents the features extracted from the first fully connected-activation layer, D2 represents the features extracted from the second fully connected-activation layer, D3 represents the features extracted from the third fully connected-activation layer, D4 represents the features extracted from the fourth fully connected-activation layer, D5 represents the features extracted from the fifth fully connected-activation layer, and D... F1 D is used to extract network prediction weights for multi-scale spatial feature extraction. F2 Site spatial linkage network predicted value weight, O WT For multi-scale spatial feature extraction network predictions, O G For the predicted value of the site spatial linkage network, O Final This is the final prediction.

[0108] This network can dynamically adjust the weights of multi-scale spatial feature extraction network predictions and the weights of site spatial linkage network predictions based on the geographical region of the target grid and the density of site distribution. This not only improves the prediction performance of the network but also enhances its robustness, ensuring the stability of the prediction results and generating high-quality near-surface wind fields.

[0109] The training process for lightweight deep learning networks includes data standardization, model training, and model validation. The data standardization stage standardizes the model input to improve training speed and performance stability while reducing overfitting. The model training stage includes forward propagation and backpropagation, optimizing network parameters through a loss function to achieve the model's learning process. The model validation stage includes cross-validation and performance evaluation to ensure the model's robustness and accuracy on unknown data.

[0110] The structure of an adaptive weighted network is as follows Figure 4As shown, the adaptive weighted network structure includes: a fully connected layer for extracting high-dimensional features and a high-dimensional transformation for low-dimensional features; and an activation layer for extracting complex nonlinear features, enhancing the network's expressive power. Based on this, using the latitude and longitude, wind field data, terrain data, timestamps, and number of surrounding stations at the target grid point as input data, an adaptive weighted network is constructed using fully connected and activation layers. This network predicts the weights of the multi-scale spatial feature extraction network and the station spatial linkage network for the target grid point. Then, the prediction results from the two networks are weighted and summed to generate the final wind speed and direction predictions. This achieves adaptive consideration of different geographical locations and different station distribution densities.

[0111] Based on predictions using a multi-scale spatial feature extraction network and a site spatial linkage network, an adaptive weighted network is used to dynamically calculate the corresponding weights, thereby obtaining wind speed and direction predictions for the target grid points. This method comprehensively considers the correlation between wind field and topography, as well as the linkage between sites, and can adaptively consider the geographical location and site distribution density of different grid points, resulting in significant improvements in accuracy and robustness.

[0112] The entire downscaling network of this invention includes a multi-scale spatial feature extraction network, a site spatial linkage network, and an adaptive weighting network. The training process of the downscaling network includes: a standardization stage to improve the training speed and performance stability of the model while reducing overfitting; a model training stage, including forward propagation and backpropagation, to optimize network parameters through a loss function to realize the model's learning process; and a model validation stage, including cross-validation and performance evaluation, to ensure the robustness and accuracy of the model on unknown data.

[0113] The near-surface wind field downscaling method of this invention can be applied to regional wind speed downscaling in wind energy assessment. By processing wind field data under different geographical regions and station distribution densities through a deep learning network, it quickly and accurately generates high-resolution wind field maps, improving the accuracy and robustness of the wind field grid. This is of great significance for wind energy development, wind power station site selection, and other fields. The near-surface wind field downscaling method of this invention has advantages such as fast processing speed, good downscaling effect, and strong robustness, and can meet the near-surface wind field downscaling tasks in various scenarios. Furthermore, the implementation of this method does not require complex hardware support and can be easily deployed on existing computing platforms, providing an efficient wind field downscaling tool for the energy and meteorological fields.

[0114] The specific embodiments of this invention provide those skilled in the art with a novel solution for downscaling near-surface wind fields, contributing to the development of meteorological data downscaling techniques. The scope of protection is defined by the claims. The implementation of this method will significantly improve the quality and usability of near-surface wind fields under complex terrain conditions, thus playing a crucial role in disaster prevention and wind energy utilization. By providing higher-quality wind field grids, more accurate scientific research and decision-making can be conducted, which is essential for improving efficiency and safety in these fields. Furthermore, the downscaling technique of this invention can also be applied to other fields, such as environmental protection and sustainable development, where high-quality wind field grids are equally crucial. In summary, this invention provides a powerful and flexible tool for improving near-surface wind field quality, opening up new possibilities for various applications.

[0115] The near-surface wind field downscaling method of this invention not only efficiently completes the downscaling task but also significantly improves the quality of wind field data, providing strong technical support for research and applications in related fields. Through this invention, users can obtain clearer and more accurate wind field maps, thereby enabling more effective scientific research and decision-making.

[0116] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for downscaling near-surface wind fields, characterized in that, The method includes the following steps: S1. Acquire site data, wind field data, and terrain data. Input the wind field data and terrain data into a multi-scale spatial feature extraction network, which consists of a feature extraction module. The module consists of a multi-scale feature fusion module and a first fully connected prediction module. After wind field data and terrain data are input into the feature extraction module, the feature extraction module outputs the total high-dimensional features of the wind field and terrain. ; S2, the total high-dimensional features of the wind field and terrain Input multi-scale feature fusion module, output multi-scale fused features ; S3, Multi-scale Fusion Features Input the first fully connected prediction module, and the first fully connected prediction module outputs the first near-surface wind field grid prediction value; S4. Construct a spatial location map of station data and a correlation map of station data features based on the station data. Input the spatial location map of station data, the correlation map of station data features, and the corresponding station features into a station spatial linkage network. The station spatial linkage network includes a spatial location pipeline flow, a correlation pipeline flow, and a second fully connected prediction module. The outputs of the spatial location pipeline flow and the correlation pipeline flow are input to the second fully connected prediction module. The second fully connected prediction module outputs the second near-surface wind field grid prediction value. S5. Obtain adaptive weighted analysis data, input the adaptive weighted analysis data into the adaptive weighted network, generate the first weight corresponding to the first near-surface wind field grid prediction value and the second weight corresponding to the second near-surface wind field grid prediction value, and obtain the final surface wind field grid prediction value based on the first weight and the second weight.

2. The near-surface wind field downscaling method based on deep learning according to claim 1, characterized in that, The total high-dimensional features of the wind field and topography for: in, Represents the feature extraction module. This represents the input wind field data. This represents the high-dimensional wind field features after processing by the feature extraction module. Represents the input terrain data, This represents the high-dimensional terrain features after processing by the feature extraction module. This represents a splicing operation. It represents the overall high-dimensional characteristics of wind field and topography.

3. The near-surface wind field downscaling method based on deep learning according to claim 2, characterized in that, Each feature extraction module It consists of three convolutional modules, two meanpooling layers, and an attention layer. Each convolutional module contains a convolutional layer and a Silu activation layer, as well as a feature extraction module. The calculation process is as follows: in, Representative feature extraction module Input, For convolution operations, For Silu activation operation, For Batch Normalization, For average pooling operation, This represents the attention mechanism processing operation. For feature extraction module The output.

4. The near-surface wind field downscaling method based on deep learning according to claim 3, characterized in that, The attention mechanism processing operations include spatial attention mechanism (SA) and channel attention mechanism (CA).

5. The near-surface wind field downscaling method based on deep learning according to claim 4, characterized in that, The multi-scale fusion features for: in, This represents the overall high-dimensional characteristics of the wind field and topography. for Feature fusion operation of the module Features of the first MSF module fusion Features fused for the second MSF module, This is a feature fused to the third MSF module.

6. The near-surface wind field downscaling method based on deep learning according to claim 5, characterized in that, First near-surface wind field grid prediction value for: in, For linear layer transformation, For Silu activation operation, For flattening operation, The features extracted from the first fully connected-activation layer of the first fully connected prediction module. The features extracted from the second fully connected-activation layer of the first fully connected prediction module. The features extracted from the third fully connected-activation layer of the first fully connected prediction module. These are the features extracted from the fourth fully connected-activation layer of the first fully connected prediction module.

7. The near-surface wind field downscaling method based on deep learning according to claim 1, characterized in that, Output of spatial location pipeline flow for: in, The input to the pipeline represents the spatial location of the flow. 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. The calculation process of the message aggregation layer is as follows: =Softmax(fc( ( ) )) 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; Output of correlated pipeline flow for: in, The input to the correlated pipeline flow is represented. represent Message aggregation layer The representative graph attention message aggregation layer, each GAT layer contains a multi-head attention mechanism.

8. A near-surface wind field downscaling method based on deep learning according to claim 7, characterized in that, Second near-surface wind field grid prediction value for: in For linear layer transformation, For Silu activation operation, 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.

9. A near-surface wind field downscaling method based on deep learning according to claim 8, characterized in that, The adaptive weighted analysis data includes the latitude and longitude of the target grid point, the number of surrounding stations, timestamps, wind field data, and terrain data. The wind field data and terrain data are wind field grids and terrain grids centered on the target grid point. The near-surface wind field grid prediction values ​​are the wind speed and wind direction prediction values ​​of the target grid point.

10. A near-surface wind field downscaling method based on deep learning according to claim 1, characterized in that, Final surface wind field grid prediction values for: in, For linear layer transformation, For Silu activation operation, 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 assigned to the first near-surface wind field grid prediction value output by the multi-scale spatial feature extraction network. The second weight of the second near-surface wind field grid prediction value output by the site spatial linkage network. This is the predicted value for the first near-surface wind field grid. This is the second near-surface wind field grid prediction value.

Images