An intelligent site correction method for numerical forecast products under small sample conditions

Abstract

The present invention discloses a method for intelligent site correction of numerical forecast products under small sample conditions, comprising: obtaining original data of the numerical forecast product, wherein the original data includes precipitation, temperature, wind speed and visibility; constructing a graph neural network, wherein the GNN layer of the graph neural network adopts a diffusion graph convolutional network as a basic module, and constructing an SE-ResNet layer, wherein the SE-ResNet layer is a residual block combined with an SE module; integrating the SE-ResNet layer into the graph neural network to obtain a correction model SE-ResGNN; training the correction model SE-ResGNN, and correcting the original data using the trained correction model SE-ResGNN to obtain a correction result of the original data.

Description

Technical Field

[0001] The present invention belongs to the technical field of numerical forecast correction, and in particular relates to an intelligent site correction method for numerical forecast products under small sample conditions. Background Art

[0002] As numerical forecast products mature, they are gaining widespread application in forecasting operations. Numerical forecast models can simulate changes in atmospheric systems, but due to model errors, observational data errors, local climate characteristics, and complex terrain, their output data may differ from actual observations. Therefore, using post-processing techniques to correct station data and improve its accuracy and reliability is beneficial for disaster prevention and mitigation and has practical significance.

[0003] Currently, post-processing techniques for station data primarily include data assimilation, statistical methods, and spatial interpolation. Due to the uneven distribution of stations and the complexity of the data, correcting various types of data may require different methods, significantly increasing the workload. Deep learning has advanced rapidly in recent years, but existing deep learning models are rarely applied to station data correction. Given the complexity and unique nature of the data, it is necessary to develop an intelligent station correction algorithm for numerical forecast products under small sample conditions to improve the accuracy of forecast data. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes an intelligent site correction method for numerical forecast products under small sample conditions, which realizes the intelligent correction of site data, reduces the error between the site data and the actual observation value, and improves the accuracy of the forecast data to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides a method for intelligent site correction of numerical forecast products under small sample conditions, comprising:

[0006] Obtaining raw data of numerical forecast products, wherein the raw data includes precipitation, temperature, wind speed and visibility;

[0007] Constructing a graph neural network, wherein the GNN layer of the graph neural network uses a diffuse graph convolutional network as a basic module, constructing an SE-ResNet layer, wherein the SE-ResNet layer is a residual block combined with an SE module; integrating the SE-ResNet layer into the graph neural network to obtain a revised model SE-ResGNN;

[0008] The correction model SE-ResGNN is trained, and the original data is corrected by the trained correction model SE-ResGNN to obtain a correction result of the original data.

[0009] Optionally, the graph neural network includes six layers of diffusion graph convolutional networks connected in sequence, wherein more generalized features are obtained through the first five layers of diffusion graph convolutional networks, and the more generalized features are reconstructed through the last layers of diffusion graph convolutional networks.

[0010] Optionally, in the SE-ResNet layer, global average pooling is performed on the input data of the input residual block, and the result of the global average pooling is processed by a fully connected layer, a SiLU activation function, a fully connected layer, and a Sigmoid activation function connected in sequence to obtain a channel attention weight;

[0011] The input data is processed by the diffusion graph convolutional network and the SiLU activation function, the channel attention weight and the processing result are multiplied, and the multiplication result is spliced ​​with the input data to obtain the final output data of the SE-ResNet layer.

[0012] Optionally, the SE-ResNet layers are fused into the diffusion graph convolutional network except the first and last layers.

[0013] Optionally, after obtaining the raw data of numerical forecast products, the following steps are also included:

[0014] The original data for obtaining the numerical forecast product is preprocessed, wherein the preprocessing includes filling and decomposing the wind speed into latitudinal wind speed and longitudinal wind speed, and replacing the input data of the wind speed with the decomposition result.

[0015] Optionally, the process of decomposing the wind speed includes:

[0016] ,

[0017]

[0018] Where spd is the given wind speed, wdir is the wind direction, u represents zonal wind, and v represents meridional wind.

[0019] Optionally, the data processing of the diffusion graph convolutional network includes:

[0020]

[0021] in, and are the forward transfer matrix and the backward transfer matrix respectively, For the l The output of the layer, represents the k-th order Chebyshev polynomial, and For the l The learning parameters of the layer.

[0022] Optionally, the revised model is trained using mean square error as a loss function.

[0023] Compared with the prior art, the present invention has the following advantages and technical effects:

[0024] Due to the complexity of station data forecasts, existing numerical forecast products exhibit discrepancies between their forecasts and actual observations. Graph neural networks (GNNs) offer significant advantages in processing complex and structured data. Through their unique message passing mechanism, they are powerful in representing complex spatial dependencies. They can leverage graph structure to improve forecasts, integrating information from different scales and enhancing forecast accuracy through message passing. They can also enhance the robustness of prediction models and reduce sensitivity to data noise. To further improve the performance of GNN architectures, we have focused on Residual Networks (ResNets), which enable deeper training. These ResNets are based on convolutional networks and address the degradation issues that can occur with excessive network depth. They can capture complex patterns in forecast data, making the network more stable during training and improving model generalization. Furthermore, given the unique nature of forecast data, the attention mechanism of the SE module effectively captures information from station forecast data. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0026] Figure 1 The model structure of SE-ResGNN according to an embodiment of the present invention;

[0027] Figure 2 This is a comparison chart of the MSE before and after correction of the surface pressure, 2-meter temperature, 2-meter relative humidity, 10-meter meridional wind and zonal wind, and visibility variables in an embodiment of the present invention;

[0028] Figure 3 This is a comparison chart of the MSE before and after correction of the 6-hour and 12-hour precipitation and the 24-hour highest and lowest 2-meter temperatures in an embodiment of the present invention;

[0029] Figure 4 This is a comparison chart of the 72-hour surface air pressure forecast data and the observed values ​​before and after the correction on December 4, 2023, according to an embodiment of the present invention;

[0030] Figure 5 This is a comparison chart of the 72-hour forecast data and observed values ​​of the 2m temperature before and after the correction on December 4, 2023, according to an embodiment of the present invention;

[0031] Figure 6A comparison chart of the 72-hour forecast data and observed values ​​of the meridional wind and zonal wind before and after the correction on December 4, 2023, according to an embodiment of the present invention;

[0032] Figure 7 This is a comparison chart of the 72-hour visibility forecast data and observed values ​​before and after the correction on December 4, 2023, according to an embodiment of the present invention;

[0033] Figure 8 This is a comparison chart of the 72-hour visibility forecast data and observation values ​​before and after the correction on June 6, 2023, according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0035] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0036] The present invention discloses a method for intelligent site correction of numerical forecast products under small sample conditions, comprising the following steps: S1, an architecture model, using graph neural networks (GNNs), and the GNN layer using diffuse graph convolutional networks (DGCNs) as the basic module; S2, an architecture SE-ResNet layer, performing global average pooling on the input, using two fully connected layers to generate channel attention weights, and using SiLU and Sigmoid activation functions respectively, performing graph convolution operations on the input data in the GNN layer, and multiplying the output of the graph convolution with the attention weight of the SE module , and using the SiLU activation function, the input of the residual block is added to the activated output to obtain the final output result; S3, the SE-ResNet layer is combined with the GNN layer to complete the model architecture; S4, the raw data is processed; S5, the batch size of the processed raw data is set to 8, and the shape is changed to meet the input requirements of the graph convolution layer, which is used as the training data set; S6, the SE-ResGNN model is trained using the training data set; S7, the raw data processed in S5 is input into the trained model to obtain the SE-ResGNN model correction result. This invention can realize intelligent site correction of numerical forecast products under small sample conditions, improving the accuracy and reliability of the data.

[0037] This paper proposes an intelligent site correction method for numerical forecast products under small sample conditions to achieve intelligent correction of site data, reduce the error between the site data and the actual observation value, and improve the accuracy of the forecast data (see Cases 1 and 2 for details).

[0038] Technical Solution: This paper proposes an intelligent site correction method for numerical forecast products under small sample conditions. This method incorporates a residual block combined with an SE mechanism into a GNN network. By combining graph convolution and residual connections, it gradually extracts features and reconstructs input data. This method achieves efficient feature learning and reconstruction on time series data, enhancing the ability to express complex time series data. The method includes the following specific steps:

[0039] Step S1: Architect the model using graph neural networks (GNNs), with the GNN layer using diffuse graph convolutional networks (DGCNs) as the basic module;

[0040] Step S2: Build the SE-ResNet layer, perform global average pooling on the input, use two fully connected layers to generate channel attention weights, and use SiLU and Sigmoid activation functions respectively. Perform graph convolution on the input data in the GNN layer, multiply the output of the graph convolution with the attention weight of the SE module, and use the SiLU activation function to add the input of the residual block to the activated output to obtain the final output result;

[0041] Step S3: Combine the SE-ResNet layer with the GNN layer to complete the model architecture;

[0042] Step S4: processing the original data;

[0043] Step S5: Set the batch size of the processed original data to 8 and change the shape to meet the input requirements of the graph convolution layer as the training dataset;

[0044] Step S6: Train the SE-ResGNN model using the training dataset;

[0045] Step S7: Input the original data processed in S5 into the trained model to obtain the SE-ResGNN model correction result.

[0046] Furthermore, for step S1, a diffuse graph convolutional network (DGCN) is used as the basic module of the GNN architecture. The GNN network module is a six-layer network. The first five layers of DGCN are used to obtain a more generalized representation, and the sixth layer of DGCN is used to output the reconstruction result.

[0047] Furthermore, in step S2, a SE module is added to the ResNet layer. The SE module generates attention weights and uses two activation functions, SiLU and Sigmoid. After the GNN layer performs a graph convolution operation on the input data, the output of the graph convolution is multiplied by the attention weights of the SE module and activated using the SiLU function. Finally, the input of the residual block is added to the activated output, i.e., the sum is performed to obtain the final output.

[0048] Furthermore, for step S3, the SE-ResNet module is combined with the GNN module to obtain the SE-ResGNN model.

[0049] Furthermore, in step S4, the method for processing the original data is to fill in the null and missing values, using adjacent values ​​for interpolation or direct filling. Among them, the original precipitation data contains some negative outliers, and their absolute values ​​are taken for repair. The filling and repair methods for variables such as air pressure, visibility, and temperature are similar. The wind is decomposed into zonal wind u and meridional wind v;

[0050] Furthermore, the loss function of this model uses Mean Square Error (MSE):

[0051] (1)

[0052] in is the predicted value, y is the target value. Using the above loss function, we can update the best loss by recording the loss of the model on the test set and the best loss recorded to obtain the optimal weight of the model.

[0053] The above technical solution is described in detail with reference to the accompanying drawings:

[0054] like Figure 1 As shown, the present invention provides an intelligent site correction method for numerical forecast products under small sample conditions. The algorithm uses precipitation, temperature, wind, visibility and other data and utilizes the SE-ResGNN model to perform intelligent correction on numerical forecast products under small sample conditions. The method includes the following steps:

[0055] Step S1: Architecture model, using graph neural networks (GNNs), and the GNN layer uses diffuse graph convolutional networks (DGCNs) as the basic module.

[0056] In order to characterize the randomness of spatial and directional dependencies, diffuse graph convolutional networks (DGCNs) are used as the basic module of the GNN architecture:

[0057] (2)

[0058] in, and The two transfer matrices are forward transfer matrix and backward transfer matrix respectively. The reason for using two transfer matrices here is that the adjacency matrix may be asymmetric in a directed graph. represents the adjacency matrix, rowsum Represents a row sum function.

[0059] In an undirected graph, . is the order of the diffusion convolution. The Chebyshev polynomial is used to approximate the convolution process in DGCN, and it is defined recursively. ,get , . and For the l The learned parameters of a layer control how each node transforms the information it receives. For the l The output of the layer, I represents the identity matrix, X represents the independent variable of the Chebyshev polynomial, represents the k-th order Chebyshev polynomial.

[0060] Unlike other traditional GNNs that use fixed spatial structures, in the GNN network used in this paper, each sample has its own subgraph structure. Therefore, the adjacency matrix that captures neighborhood information and message passing direction There are also differences in different samples. The GNN network module is a network consisting of six layers. The input of the first layer is , then according to formula (2), use the parameter and get Since the nodes in the first layer only pass 0 to their neighboring nodes, the GCN in one layer cannot obtain the ideal features. Therefore, by adding another four layers of DGCN (i.e. ) to obtain a more generalized representation:

[0061] (3)

[0062] in and are the parameters of the second to fifth layers of DGCN, is a nonlinear activation function. Repeat the steps of formula (3) to obtain .

[0063] Finally, another layer of DGCN is used to output the reconstructed result:

[0064] (4)

[0065] in and is the learning parameter of the last layer.

[0066] Step S2: Architecture SE-ResNet layer Figure 1 The Resblock in the GNN performs global average pooling (AvgPool) on the input, uses two linear fully connected layers (Linear) to generate channel attention weights, and uses SiLU and Sigmoid activation functions respectively. Graph convolution is performed on the input data in the GNN layer, and the output of the graph convolution is multiplied by the attention weight of the SE module. The SiLU activation function is used to add the input of the residual block to the activated output to obtain the final output result.

[0067] The SE-ResNet layer adds an SE module as the weight of different channels to the original residual connection extraction module GNN layer to improve the features of different channels and enhance the expressiveness of features.

[0068] A SE (squeeze-and-excitation) module is added to the ResNet structure to add a self-attention mechanism to the residual structure. When the model inputs multiple elements, the SE module can select the importance of each channel through squeeze and excitation, and obtain weight coefficients on each channel to recalibrate the importance of each original channel. Among them, the squeeze part is used to embed global information, which is achieved by generating channel statistics through global average pooling. Statistics Through the spatial dimension of U To shrink, z No. c The calculation formula for each element is:

[0069] (5)

[0070] Next, the purpose of the excitation part is to fully capture the channel-related dependencies:

[0071] (6)

[0072] in is the SiLU function, , .

[0073] In order to limit the complexity of the model and improve the generalization ability, the model forms a restriction layer around the nonlinearity. The restriction layer has two fully connected layers, namely a dimensionality reduction layer, a SiLU layer and a dimensionality increase layer. The dimensionality reduction layer and the dimensionality increase layer both use Linear layers. The final output of the SE module is obtained by activating Rescale Obtained:

[0074] (7)

[0075] in , express and Channel-by-channel multiplication between .

[0076] Step S3: Combine the SE-ResNet layer with the GNN layer. For ease of understanding, the SE-ResNet layer is combined with the GNN layer to replace the DCGN. This is an improvement based on the middle 4 DCGN layers. Residual connections and SE modules are added to the middle 4 DCGNs. The output of the SE module is first multiplied with the DCGN as a weight, and the weighted DCGN layer output result is added to the DCGN input data through the residual connection structure to generate the final output of the SE-ResNet layer, providing input data for the next DGCN.

[0077] Complete the model architecture.

[0078] Step S4: Process the original data.

[0079] Obtain the original data of precipitation, temperature, wind, visibility, etc. of the numerical forecast product, fill in the empty values ​​and missing values ​​of the data, and repair the outliers. The method of filling in empty values ​​and missing values ​​is to use adjacent values ​​for interpolation or direct filling. Taking the 6-hour cumulative precipitation data as an example, if the two values ​​before and after a certain empty value or missing value exist, the average value of the two values ​​before and after is used to fill it. If the two values ​​before and after do not exist, the value of the most recent time is selected to fill it directly. In addition, there are some outliers with negative values ​​in the original precipitation data. This paper takes their absolute values ​​for repair. The filling and repair methods of variables such as air pressure, visibility, and temperature are similar. In particular, in order to improve the model's correction effect on wind direction and wind speed, this model decomposes the wind into zonal wind u and longitudinal wind v, which are input into the model for calculation respectively:

[0080] (8)

[0081] (9)

[0082] Among them, spd is the given wind speed and wdir is the wind direction.

[0083] Step S5: Set the batch size of the processed original data to 8 and change the shape to meet the input requirements of the graph convolution layer as the training dataset.

[0084] The training set in step S4 contains 25 forecast files and 50 stations. The batch size is set to 8, and random shuffling is performed during training, resulting in an original input shape of (8, 25, 50). First, adjust the dimensional order of the input data to meet the input requirements of the graph convolution layer, that is, adjust it to an array of size (8, 50, 25).

[0085] Step S6: Train the SE-ResGNN model using the training dataset.

[0086] The training set was 100 epochs, with a training set to validation set ratio of 0.9, a batch size of 8, and a learning rate of 1e-4. During model training, the forward and backward random walk matrices were calculated, the optimal loss and the corresponding model weights were tracked in real time, and the model parameters were updated based on the calculated gradients. During the validation phase, the optimal loss of the model was recorded, and the epoch with the lowest loss was saved as the optimal model.

[0087] Step S7: Input the original data processed in S5 into the trained model to obtain the SE-ResGNN model correction result.

[0088] After saving the best trained model, we use it to revise the forecast for each meteorological variable. We load the model parameters, generate a random walk matrix, and then pass the data into the model. The model outputs a one-to-one correspondence between the corrected results and the time.

[0089] In particular, for wind forecast correction, the decomposed u and v need to be restored to the wind speed spd and wind direction wdir forecast correction results finally output by the model. The calculation formula is as follows:

[0090] (10)

[0091] (11)

[0092] Case 1:

[0093] The feasibility of the intelligent site correction method for numerical forecast products under small sample conditions in the present invention is verified using an actual case.

[0094] The present invention proposes an intelligent site correction method for numerical forecast products under small sample conditions. This method can perform intelligent site corrections on numerical forecast products. Using the SE-ResGNN correction model, the corrections for surface pressure, 2-meter temperature, 2-meter relative humidity, 10-meter meridional and zonal winds, visibility, 6-hour and 12-hour precipitation, and 24-hour maximum and minimum 2-meter temperatures are evaluated using the mean square error (MSE). The MSE for each variable is calculated to reflect the model's correction effect on each variable. The closer the MSE is to 0, the more significant the model's correction effect on that variable. Figures 2 to 3 The MSE values ​​of several variables before and after correction are given. The light gray columns are the MSE of the predicted values, and the dark gray columns are the MSE of the corrected values. It can be seen that the model has effectively corrected each variable.

[0095] from Figure 2 As can be seen, compared to the original data, the error between the surface pressure and the observations after correction is significantly reduced, demonstrating a significant correction effect. For the 2-meter temperature, the mean square error is reduced by nearly half. For the wind direction at 10 meters above the ground, both the mean square error (MSE) for meridional and zonal winds are reduced to half of their initial values ​​before correction. The correction effect on visibility is even more pronounced. The model effectively corrects these variables, and the revised forecasts significantly reduce the error between the forecasts and the observations. The correction effect is particularly strong for variables such as surface pressure, temperature, wind, and visibility. Figure 3 The mean square error analysis of the revised 6-hour and 12-hour precipitation forecasts, as well as the 24-hour maximum and minimum temperature forecasts, is presented. It shows that for precipitation, the forecasts for different time periods are effectively corrected, with the errors in the 6-hour and 12-hour precipitation forecasts relative to the observed values ​​reduced by more than half. Regarding temperature, the model provides a more effective correction for low temperatures, with a significant reduction in the error in the 24-hour minimum temperature.

[0096] Case 2:

[0097] The following is an actual case to verify the accuracy and reliability of the correction results of the correction model SE-ResGNN in the intelligent site correction method of numerical forecast products under small sample conditions in the present invention and the comparison with the observed values.

[0098] The correction model SE-ResGNN in the intelligent site correction method of numerical forecast products under small sample conditions in the present invention can perform intelligent site correction on the results of numerical forecast products. Here, the 72-hour surface pressure forecast data, 72-hour 2-meter temperature forecast data, 72-hour meridional wind and zonal wind forecast data, 72-hour visibility forecast data on December 4, 2023, and the 72-hour visibility forecast data on June 6, 2023 are taken as examples. Figures 4 to 8Comparisons of the 72-hour forecast data and observed values ​​before and after correction are given for surface pressure, 2-meter temperature, zonal wind, meridional wind, visibility on December 4, 2023, and 24-hour precipitation on June 6, 2023 (dark gray lines are forecast values, light gray lines are revised values, and dashed lines are observed values).

[0099] Depend on Figure 4 It can be seen that the forecast value of surface pressure on December 4, 2023 is significantly lower than the observed value, and the change range within 72 hours is also larger than the observed value. After the model correction, the forecast results are very close to the observed values ​​in terms of numerical value, and the change trend and range are basically consistent with the observed values, indicating that the model correction effect is significant. Figure 5 ), within the first 12 hours of the forecast, the revised value is closer to the observed value, while the revised value of the model from 12 hours to 36 hours is smaller, but still close to the observed value and the forecast value. After 36 hours, the original forecast value is significantly lower, but after the model correction, the revised value is obviously closer to the observed value. Figure 6 ), the original forecast values ​​fluctuated significantly, especially the meridional wind, which was significantly greater than the observed value within 24 hours and significantly lower after 36 hours, while the zonal wind was significantly higher after 20 hours. After model correction, the numerical values ​​and change trends are closer to the observed values, and the larger and smaller values ​​of the meridional and zonal winds have been significantly corrected, and the change range within 72 hours is more consistent with the observed values. Figure 7 ), the forecast results before the correction were significantly higher, reaching 25km in most periods within 72 hours, which was significantly different from the observed values, and the change trend was also quite different from the observed values. After the correction, the numerical value of the correction value reached a similar order of magnitude as the observed value, and the change range was more consistent. The peak values ​​around 22 hours and 42 hours also performed well, and the correction effect was significant. For 24-hour precipitation ( Figure 8 ), it can be seen that the pre-correction forecasts were significantly higher, with significant errors occurring especially after 40 hours. The revised values ​​are now much closer to the observed values. Based on the above analysis, the model's correction of the forecast data brings the revised values ​​closer to the observed values. The model corrects for significant deviations in both numerical value and variation, achieving significant correction results for all forecast variables.

[0100] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for intelligent site correction of numerical forecast products under small sample conditions, characterized by: include: Obtaining raw data of numerical forecast products, wherein the raw data includes precipitation, temperature, wind speed and visibility; Constructing a graph neural network, wherein the GNN layer of the graph neural network uses a diffuse graph convolutional network as a basic module, constructing an SE-ResNet layer, wherein the SE-ResNet layer is a residual block combined with an SE module; integrating the SE-ResNet layer into the graph neural network to obtain a revised model SE-ResGNN; The correction model SE-ResGNN is trained, and the original data is corrected by using the trained correction model SE-ResGNN to obtain a correction result of the original data; The graph neural network includes six layers of diffusion graph convolutional networks connected in sequence, wherein the first five layers of diffusion graph convolutional networks obtain more generalized features, and the last layer of diffusion graph convolutional networks reconstructs the more generalized features; In the SE-ResNet layer, global average pooling is performed on the input data of the input residual block, and the result of the global average pooling is processed by a fully connected layer, a SiLU activation function, a fully connected layer, and a Sigmoid activation function connected in sequence to obtain a channel attention weight; The input data is processed by a diffusion graph convolutional network and a SiLU activation function, the channel attention weight and the processing result are multiplied, and the multiplication result is added to the input data to obtain the final output data of the SE-ResNet layer; The SE-ResNet layers are fused into the diffusion graph convolutional network except the first and last layers; The data processing process of the diffusion graph convolutional network includes: in, and are the forward transfer matrix and the backward transfer matrix respectively, H l+1 is the output of the lth layer, T k represents the k-th order Chebyshev polynomial, and is the learning parameter of the lth layer; After obtaining the raw data of numerical forecast products, the following is also included: Preprocessing the raw data of the numerical forecast product, wherein the preprocessing includes filling and decomposing the wind speed into zonal wind speed and meridional wind speed, and replacing the input data of the wind speed with the decomposition result; The process of decomposing the wind speed includes: Where spd is the given wind speed, wdir is the wind direction, u represents the zonal wind, and v represents the meridional wind; The revised model is trained using mean square error as the loss function.

Images