Numerical forecasting product intelligent station correction method under small sample condition

Abstract

The invention discloses a numerical forecasting product intelligent station correction method under a small sample condition, and the method comprises the steps: obtaining the original data of a numerical forecasting product, and the original data comprise rainfall, temperature, wind speed and visibility; a graph neural network is constructed, a diffusion graph convolutional network is adopted as a basic module in a GNN layer of the graph neural network, an SE-ResNet layer is constructed, and the SE-ResNet layer is a residual block combined with an SE module; the SE-ResNet layer is fused into a graph neural network, and a correction model SE-ResGNN is obtained; and training the correction model SE-ResGNN, and correcting the original data through the trained correction model SE-ResGNN to obtain a correction result of the original data.

Description

Technical Field

[0001] The 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 become more mature, they are widely used in forecasting operations. Numerical forecast models can simulate changes in atmospheric systems, but due to model errors, observation data errors, local climate characteristics and complex terrain, their output data may differ from actual observations. Therefore, using post-processing technology to correct station data and improve the accuracy and reliability of station data is beneficial to disaster prevention and mitigation and has practical significance.

[0003] At present, the post-processing technology for station data mainly includes data assimilation, statistical methods, spatial interpolation and other methods. Due to the uneven distribution of stations and the complexity of data, different methods may be required for the correction of various data, which greatly increases the workload. In recent years, deep learning has developed rapidly, and existing deep learning models are rarely used in station data correction. Due to the complexity and specialness of data, it is necessary to propose 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 actual observation value and the forecast data, 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: 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 adopts a diffusion graph convolutional network as a basic module, constructing a 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 the correction result of the original data.

[0006] Optionally, the graph neural network includes six layers of diffusion graph convolutional networks connected in sequence, where 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 layer of diffusion graph convolutional network.

[0007] Optionally, in the SE-ResNet layer, global average pooling is performed on the input data of the input residual block, and the result of global average pooling is processed through a fully connected layer, a SiLU activation function, a fully connected layer, and a Sigmoid activation function connected in sequence to obtain channel attention weights; The input data is processed through a diffusion graph convolutional network and a SiLU activation function, the channel attention weights and the processing result are multiplied, and the multiplication result and the input data are concatenated to obtain the final output data of the SE-ResNet layer.

[0008] Optionally, the SE-ResNet layer is fused into the diffusion graph convolutional networks except the first and last layers.

[0009] Optionally, after obtaining the original data of the numerical weather prediction product, it further includes: Preprocessing the original data of the obtained numerical weather prediction product, where 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.

[0010] Optionally, the process of decomposing the wind speed includes: ,

[0011] where spd is the given wind speed, wdir is the wind direction, where u represents the zonal wind and v represents the meridional wind.

[0012] Optionally, the data processing process of the diffusion graph convolutional network includes:

[0013] where, and are the forward transfer matrix and the backward transfer matrix respectively, is the output of the l th layer, represents the Chebyshev polynomial of the kth order, and are the learning parameters of the l th layer.

[0014] Optionally, the correction model is trained using the mean square error as the loss function.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: Due to the complexity of site data forecasting, there are errors between the forecasts of existing numerical forecasting products for site data and the actual observed values. Graph neural networks (GNNs) have significant advantages in processing complex and structured data. Through its unique information transfer mechanism, it has strong capabilities in characterizing complex spatial dependence relationships. It can utilize graph structure information to improve the forecasting results, integrate information at different scales, improve the accuracy of forecasting through the message passing mechanism, and can also enhance the robustness of the prediction model and reduce the sensitivity to data noise. To further improve the performance of the GNN architecture, we have noticed the residual network (ResNet) that can train deeper networks. It is developed based on convolutional networks and is used to solve the possible degradation problems when the network depth is too deep. It can capture complex patterns in the forecasting data, making the network more stable during training and enhancing the generalization ability of the model. In addition, considering the particularity of the forecasting data, the attention mechanism of the SE module can effectively capture the information of site forecasting data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings: Figure 1 is the model structure of the SE-ResGNN according to the embodiment of the present invention; Figure 2 is the comparison chart of MSE before and after correction of surface air pressure, 2m temperature, 2m relative humidity, 10m meridional wind and zonal wind, and visibility variables according to the embodiment of the present invention; Figure 3 is the comparison chart of MSE before and after correction of 6h and 12h precipitation, 24h maximum and minimum 2m temperature according to the embodiment of the present invention; Figure 4 is the comparison chart of the 72h forecast data and the observed values of the surface air pressure before and after correction on December 4, 2023 according to the embodiment of the present invention; Figure 5 is the comparison chart of the 72h forecast data and the observed values of the 2m temperature before and after correction on December 4, 2023 according to the embodiment of the present invention; Figure 6 is the comparison chart of the 72h forecast data and the observed values of the meridional wind and zonal wind before and after correction on December 4, 2023 according to the embodiment of the present invention; Figure 7 is the comparison chart of the 72h forecast data and the observed values of visibility before and after correction on December 4, 2023 according to the embodiment of the present invention; Figure 8This is a comparison chart of the visibility 72-hour forecast data and the observed values before and after correction on June 6, 2023, for the embodiments of the present invention. Detailed implementation manners

[0017] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will detail this application with reference to the drawings and in conjunction with the embodiments.

[0018] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0019] The present invention discloses an intelligent site correction method for numerical weather prediction products under small sample conditions, including the following steps: S1, constructing a model, using graph neural networks (GNNs), and using diffusion graph convolutional networks (DGCNs) as the basic module for the GNN layer; S2, constructing an SE-ResNet layer, performing global average pooling on the input, using two fully connected layers to generate channel attention weights, and using the SiLU and Sigmoid activation functions respectively. Perform graph convolution operations on the input data in the GNN layer, multiply the output of the graph convolution by 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; S3, combining the SE-ResNet layer and the GNN layer to complete the model architecture; S4, processing the original data; S5, setting the batch size of the processed original data to 8, changing the shape to meet the input requirements of the graph convolution layer, and using it as the training data set; S6, training the SE-ResGNN model using the training data set; S7, inputting the original data processed in S5 into the trained model to obtain the correction result of the SE-ResGNN model. The present invention can achieve intelligent site correction of numerical weather prediction products under small sample conditions, improving the accuracy and reliability of the data.

[0020] The present invention proposes an intelligent site correction method for numerical weather prediction products under small sample conditions to achieve intelligent correction of site data, reduce the error from the actual observed values, and improve the accuracy of the forecast data (see Case 1 and Case 2 for details).

[0021] Technical solution: Provide an intelligent site correction method for numerical weather prediction products under small sample conditions. This method adds a residual block combined with the SE mechanism to the GNN network. Through the combination of graph convolution and residual connection, features are gradually extracted and the input data is reconstructed, achieving efficient feature learning and reconstruction on time series data, and enhancing the expression ability for complex time series data. The method includes the following specific steps: Step S1: Architecture model, using graph neural networks (GNNs), with the GNN layer using diffusion graph convolutional networks (DGCNs) as the basic module; Step S2: Architecture SE-ResNet layer, perform global average pooling on the input, use two fully connected layers to generate channel attention weights, and use the SiLU and Sigmoid activation functions respectively. Perform graph convolution operations on the input data in the GNN layer, multiply the output of the graph convolution by the attention weights of the SE module, and use the SiLU activation function. Add the input of the residual block to the activated output to obtain the final output result; Step S3: Combine the SE-ResNet layer and the GNN layer to complete the model architecture; Step S4: Process the original data; Step S5: Set the batch size of the processed original data to 8, change the shape to meet the input requirements of the graph convolution layer, and use it as the training dataset; Step S6: Use the training dataset to train the SE-ResGNN model; Step S7: Input the original data processed in S5 into the trained model to obtain the correction result of the SE-ResGNN model.

[0022] Furthermore, for step S1, use diffusion graph convolutional networks (DGCNs) as the basic module of the GNN architecture. The GNN network module is a six-layer network. Use the first five layers of DGCN to obtain a more generalized representation, and finally use the sixth layer of DGCN to output the reconstructed result.

[0023] Furthermore, for step S2, add an SE module on the basis of ResNet. The SE module is used to generate attention weights, and the SiLU and Sigmoid activation functions are used respectively. After performing graph convolution operations on the input data in the GNN layer, multiply the output of the graph convolution by the attention weights of the SE module, and use the SiLU activation function. Finally, add the input of the residual block to the activated output, that is, perform addition, to obtain the final output result.

[0024] Furthermore, for step S3, combine the SE-ResNet module and the GNN module to obtain the SE-ResGNN model.

[0025] Furthermore, in step S4, the method for processing the original data is to fill in null values and missing measurements, and use adjacent values for interpolation or direct filling. Among them, there are some abnormal values with negative values in the original precipitation data, and their absolute values are taken for repair. The filling and repair methods for variables such as air pressure, visibility, and temperature are similar. Decompose the wind into zonal wind u and meridional wind v; Furthermore, the loss function of this model adopts the Mean Square Error (MSE): (1) where is the predicted value, y is the target value. By using the above loss function and updating the best loss by recording the loss of the model on the test set and the recorded best loss, the best weights of the model can be obtained.

[0026] For the above technical solution, a detailed description will be given in combination with the relevant drawings: As Figure 1 shown, the present invention provides an intelligent station correction method for numerical prediction products under small sample conditions. This algorithm uses data such as precipitation, temperature, wind, visibility, etc., and uses the SE-ResGNN model to perform intelligent correction on numerical prediction products under small sample conditions. The method includes the following steps: Step S1: Build the architecture model. Use graph neural networks (GNNs), and use the diffusion graph convolutional network (DGCNs) as the basic module for the GNN layer.

[0027] To characterize the randomness of spatial and directional dependencies, the diffusion graph convolutional network (DGCNs) is used as the basic module for the GNN architecture: (2) where and are the forward transition matrix and the backward transition matrix respectively. The reason for using two transition matrices here is that the adjacency matrix may be asymmetric in a directed graph. represents the adjacency matrix, rowsum represents the row summation function.

[0028] In an undirected graph, . is the order of the diffusion convolution. The Chebyshev polynomials are used to approximate the convolution process in DGCN, and is defined recursively to obtain , . and are the learning parameters of the l -th layer, which are used to control how each node transforms the received information. is the output of the l -th layer, I represents the identity matrix, X represents the independent variable of the Chebyshev polynomial, represents the Chebyshev polynomial of the k-th order.

[0029] Different from 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 capturing neighborhood information and message passing directions also varies among different samples. The GNN network module is a six-layer network. The input of the first layer is , and then according to formula (2), using parameters and to obtain . Since the nodes in the first layer only pass 0 to their neighboring nodes, a single-layer GCN cannot obtain ideal features. Therefore, by adding another four layers of DGCN (i.e., ) to obtain a more generalized representation: (3) where and are the parameters of the second to fifth layers of DGCN, and is the non-linear activation function. Repeating the steps of formula (3), we obtain .

[0030] Finally, another layer of DGCN is used to output the reconstructed result: (4) where and are the learning parameters of the last layer.

[0031] Step S2: The architecture of the SE-ResNet layer, which is the Resblock in Figure 1 , performs global average pooling (AvgPool) on the input, uses two linear fully connected layers (Linear) to generate channel attention weights, and uses two activation functions, SiLU and Sigmoid respectively. In the GNN layer, graph convolution operations are performed on the input data, the output of the graph convolution is multiplied by the attention weights of the SE module, and the SiLU activation function is used. The input of the residual block is added to the activated output to obtain the final output result.

[0032] The SE-ResNet layer enhances the feature representation ability by adding an SE module as the weight for different channels to improve the features of different channels on top of the original residual connection extraction module, the GNN layer.

[0033] Based on the ResNet structure, a SE (squeeze-and-excitation) module is added to incorporate a self-attention mechanism into 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 for each channel to recalibrate the importance of each original channel. Among them, the squeeze part is used to embed global information and is achieved by generating channel statistics through global average pooling. The statistic is contracted through the spatial dimension of U , then z the formula for the c th element is: (5) Next, the purpose of the excitation part is to fully capture the dependencies related to the channels: (6) where is the SiLU function, , .

[0034] To limit the complexity of the model and improve its generalization ability, a limiting layer is formed around the non-linearity. This limiting layer has two fully connected layers, namely a dimensionality reduction layer, a SiLU, and a dimensionality increase layer. Both the dimensionality reduction layer and the dimensionality increase layer use the Linear layer. The final output of the SE module is obtained by activating to rescale : (7) where , represents and the per-channel multiplication between

[0035] Step S3: Combine the SE-ResNet layer with the GNN layer. For ease of understanding, the SE-ResNet layer and the GNN layer replace the DCGN in combination. This is an improvement based on the middle 4 DCGN layers. Residual connections and SE modules are added to the middle 4 DCGN. Among them, the output of the SE module is used as the weight and multiplied by the DCGN first, and the output result of the weighted DCGN layer 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.

[0036] Complete the model architecture.

[0037] Step S4: Process the original data.

[0038] Obtain the original data of precipitation, temperature, wind, visibility, etc. from numerical weather prediction products, fill in the null values and missing measurements in the data, and repair the outliers. The method for filling in null values and missing measurements is to use adjacent numerical values for interpolation or direct filling. Taking the 6-hour accumulated precipitation data as an example, if both the values before and after a null value or missing measurement exist, the average value of the two values before and after is used for filling; if the two values before and after do not exist, the value of the nearest time step is selected for direct filling. In addition, there are some outliers with negative values in the original precipitation data, and in this paper, their absolute values are taken for repair. The filling and repair methods for variables such as air pressure, visibility, and temperature are similar. In particular, to improve the correction effect of the model on wind direction and wind speed, this model decomposes the wind into zonal wind u and meridional wind v, and inputs them into the model for calculation respectively: (8) (9) where spd is the given wind speed and wdir is the wind direction.

[0039] Step S5: Set the batch size of the processed original data to 8, change the shape to meet the input requirements of the graph convolutional layer, and use it as the training dataset.

[0040] The training set in Step S4 contains 25 forecast file numbers and 50 station numbers. Set the batch size to 8 and shuffle randomly during training. Therefore, the original input shape is (8, 25, 50). First, adjust the dimension order of the input data to meet the input requirements of the graph convolutional layer, that is, adjust it to an array of size (8, 50, 25).

[0041] Step S6: Use the training dataset to train the SE-ResGNN model.

[0042] The total number of training epochs is set to 100, the ratio of the training set to the validation set is 0.9, the batch size batch_size is 8, and the learning rate of the model is 1e-4. During the training process of the model, calculate the forward and backward random walk matrices respectively, track the best loss and the corresponding model weights in real time, and update the model parameters according to the calculated gradients. In the validation stage, record the best loss of the model and save the epoch with the smallest loss value as the best model.

[0043] Step S7: Input the original data processed in S5 into the trained model to obtain the correction result of the SE-ResGNN model.

[0044] After saving the trained best model, the best model is used to correct the forecasts for each meteorological variable. By loading the model parameters, a random walk matrix is generated, and then the data is fed into the model. The corrected results output by the model correspond to the time one by one.

[0045] In particular, for the 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 formulas are as follows: (10) (11) Case 1: The following uses an actual case to verify the feasibility of a method for intelligent site correction of numerical weather prediction products under small sample conditions in the present invention.

[0046] A method for intelligent site correction of numerical weather prediction products under small sample conditions in the present invention can perform intelligent site correction on the results of numerical weather prediction products. For the correction model SE-ResGNN, taking the correction of surface air pressure, 2m temperature, 2m relative humidity, 10m meridional wind and zonal wind, visibility, 6h and 12h precipitation, 24h maximum and minimum 2m temperature as examples, the mean square error MSE is used for evaluation. The MSE of each variable is calculated respectively, which can reflect the correction effect of the model on each variable. Among them, the closer the MSE is to 0, the more significant the correction effect of the model on this variable. Figures 2 to 3 The MSE values before and after the correction of several variables are given. The columns in light gray are the MSE of the forecast values, and the columns in dark gray are the MSE of the corrected values. It can be seen that the model has effectively corrected each variable.

[0047] From Figure 2 It can be seen that compared with the original data, the error between the corrected surface air pressure and the observed value is significantly reduced, and the correction effect is remarkable. For the 2m air temperature, the mean square error is reduced by nearly half. For the surface 10m wind direction, the MSE of the meridional wind and the zonal wind are both reduced to half of the initial value before correction. The correction effect of visibility is more obvious. For these variables, the model has effectively corrected, and the corrected forecast values have effectively reduced the error with the observed values, and have good correction effects on variables such as surface air pressure, temperature, wind and visibility. Figure 3 The mean square error analysis of the forecast correction results of 6-hour and 12-hour precipitation, 24-hour maximum and minimum temperature is given. It can be seen that for precipitation, the forecast results of precipitation at different time periods have been effectively corrected, and the error between the forecast values of 6-hour and 12-hour precipitation and the observed values is reduced by more than half. For temperature, the correction effect of this model on low temperature is better, and it can be seen that the error of the 24-hour minimum temperature is significantly reduced.

[0048] Case 2: Next, an actual case is used to verify the accuracy and reliability of the comparison between the correction results of the correction model SE-ResGNN in an intelligent station correction method for numerical prediction products under small sample conditions and the observed values.

[0049] The correction model SE-ResGNN in an intelligent station correction method for numerical prediction products under small sample conditions in the present invention can perform intelligent station correction on the results of numerical prediction products. Here, taking the 72-hour forecast data of surface air pressure, 2-meter temperature, meridional wind and zonal wind on December 4, 2023, the 72-hour forecast data of visibility, and the 72-hour forecast data of visibility on June 6, 2023 as examples, Figures 4 to 8 The comparison between the 72-hour forecast data and the observed values before and after correction of surface air pressure, 2-meter temperature, zonal wind, meridional wind, visibility on December 4, 2023, and 24-hour precipitation on June 6, 2023 is given (the dark gray broken line is the forecast value, the light gray broken line is the corrected value, and the dotted broken line is the observed value).

[0050] From Figure 4 it can be seen that the forecast value of surface air pressure on December 4, 2023 is significantly lower than the observed value, and the change range within 72 hours is also larger than that of the observed value. After model correction, the forecast result has a very small difference from the observed value in terms of numerical size, and the change trend and range are also basically consistent with the observed value, indicating that the correction effect of the model is significant. For the 2-meter air temperature ( Figure 5 ), within the first 12 hours of the forecast, the corrected value is closer to the observed value. However, from 12 hours to about 36 hours, the corrected value of the model is on the low side, but it is still close to the observed value and the forecast value. After 36 hours, the original forecast value is significantly on the low side, while at this time, after model correction, the corrected value is significantly closer to the observed value. For the meridional wind and zonal wind ( Figure 6 ), the change range of the original forecast value is significantly too large. Especially for the meridional wind, the forecast value is significantly greater than the observed value within 24 hours and significantly on the low side at 36 hours, while the zonal wind is significantly on the high side after 20 hours. After model correction, both the numerical size and the change trend are closer to the observed value, and the overestimated and underestimated values of the meridional wind and zonal wind are significantly corrected, and the change range within 72 hours is more consistent with the observed value. For the visibility ( Figure 7 ), the forecast result before correction is significantly on the high side, reaching 25 km in most periods within 72 hours, with a large difference from the observed value, and there are also significant differences in the change trend compared with the observed value. After correction, the numerical size of the corrected value reaches a similar magnitude to the observed value, and the change range is also more consistent. The peaks at about 22 hours and 42 hours are also well presented, and the correction effect is significant. For the 24-hour precipitation ( Figure 8), it can be seen that the predicted value before correction is significantly higher, especially after 40h, obvious errors occur, while the corrected value is closer to the observed value. According to the above analysis, it can be seen that after the model corrects the prediction data, the corrected value is closer to the observed value, and the model corrects the obvious deviation of the prediction data in terms of numerical size and change range, and obvious correction effects are achieved for each prediction variable.

[0051] The above is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent station correction method for numerical prediction products under small sample conditions, characterized in that, Including: Obtain the original data of numerical weather prediction products, where the original data includes precipitation, temperature, wind speed, and visibility; Construct a graph neural network, where the GNN layer of the graph neural network uses a diffusion graph convolutional network as the basic module, and construct an SE-ResNet layer, where the SE-ResNet layer is a residual block combined with an SE module; integrate the SE-ResNet layer into the graph neural network to obtain a correction model SE-ResGNN; Train the correction model SE-ResGNN, and correct the original data through the trained correction model SE-ResGNN to obtain the correction result of the original data.

2. The method according to claim 1, wherein: The graph neural network includes six layers of diffusion graph convolutional networks connected in sequence, where 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 layer of diffusion graph convolutional network.

3. The method according to claim 1, wherein: In the SE-ResNet layer, perform global average pooling on the input data of the input residual block, and process the result of global average pooling through a fully connected layer, a SiLU activation function, a fully connected layer, and a Sigmoid activation function connected in sequence to obtain channel attention weights; Process the input data through a diffusion graph convolutional network and a SiLU activation function, multiply the channel attention weights and the processing result, and add the multiplied result to the input data to obtain the final output data of the SE-ResNet layer.

4. The method according to claim 1, wherein: The SE-ResNet layer is integrated into the diffusion graph convolutional networks except the first layer and the last layer.

5. The method according to claim 1, wherein: After obtaining the original data of numerical weather prediction products, it further includes: Preprocess the original data of the obtained numerical weather prediction products, where the preprocessing includes filling, and decompose the wind speed into zonal wind speed and meridional wind speed, and replace the input data of the wind speed with the decomposition result.

6. The method according to claim 1, wherein: The process of decomposing the wind speed includes: ， ； where spd is the given wind speed, wdir is the wind direction, where u represents the zonal wind and v represents the meridional wind.

7. The method according to claim 1, wherein: The data processing process of the diffusion graph convolutional network includes: ； Among them, and are the forward transition matrix and the backward transition matrix respectively, is the output of the l -th layer, represents the Chebyshev polynomial of the -th order, and l are the learning parameters of the 8. The method according to claim 1, wherein: Train the correction model using the mean square error as the loss function.

Images