Deep learning-based strong convection strong wind short-term and imminent forecasting method and device
By optimizing and extending deep learning methods on the SwinTransformer architecture, and combining radar echoes and automatic weather station data, the accuracy problem of gale forecasting was solved, and more efficient short-term gale warnings were achieved.
Patent Information
- Application Number
- CN202511840412.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-10
AI Technical Summary
Existing technologies are insufficient to accurately predict severe weather events such as strong winds, and the complexity of meteorological scenarios makes forecasting difficult.
A deep learning-based method for short-term forecasting of severe convective winds is proposed. By optimizing and extending the SwinTransformer architecture, it combines radar echo data and automatic weather station data, and uses an encoder, decoder, feature fusion module, and MaxViT module for forecasting.
It improves the accuracy and efficiency of gale forecasting, enhances model performance, and provides more accurate short-term gale warning capabilities.
Smart Images
Figure CN121634345A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method and apparatus for short-term forecasting of severe convective winds based on deep learning. Background Technology
[0002] Severe weather events such as strong winds pose a serious threat to people's lives and property, and accurately predicting these weather events is a challenging task. Because meteorological scenarios are generally complex, accurate prediction is difficult. Therefore, how to accurately forecast short-term strong winds has become an important problem that urgently needs to be solved. Summary of the Invention
[0003] This invention provides a method and apparatus for short-term forecasting of severe convective winds based on deep learning. By optimizing and extending the SwinTransformer architecture, it accurately forecasts short-term severe winds using radar echo data and automatic weather station data.
[0004] This invention provides a method for short-term forecasting of severe convective winds based on deep learning, comprising:
[0005] Historical data is input into the encoder to obtain the encoding result of the historical data output by the encoder. The historical data includes historical radar echo data and historical automatic station data.
[0006] The encoding result is input into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module;
[0007] The encoding result and the intermediate features are input into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module;
[0008] The fused features are input into the decoder to obtain the decoding result of the fused features output by the decoder. The encoder and decoder include a Swing Transformer module.
[0009] Based on the decoding results, the short-term gale forecast results are obtained.
[0010] According to the present invention, a short-term forecasting method for severe convective winds based on deep learning is provided, wherein the encoder and the decoder each include multiple Swing Transformer modules, and the number of Swing Transformer modules in the encoder and the decoder are the same, and they are divided into N groups, where N is a positive integer.
[0011] According to the present invention, a short-term forecasting method for severe convective winds based on deep learning is provided, which inputs historical data into an encoder to obtain the encoding result of the historical data output by the encoder, including:
[0012] The historical data is sequentially input into multiple sets of Swin Transformer modules in the encoder to obtain the encoding result of the historical data output by each set of Swin Transformer modules.
[0013] According to a deep learning-based short-term forecasting method for severe convective winds provided by the present invention, the encoding result is input into a MaxViT module to obtain intermediate features of the historical data output by the MaxViT module, including:
[0014] The encoding result of the historical data output by the last group of Swin Transformer modules is input into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module.
[0015] According to a deep learning-based short-term forecasting method for severe convective winds provided by the present invention, the number of input feature fusion modules is the same as the number of groups of Swin Transformer modules in the encoder. The encoding result and the intermediate features are input into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module, including:
[0016] For the first-layer feature fusion module, the encoding result and the intermediate features are used as input to obtain the fused features output by the first-layer feature fusion module, and the fused features are input into the first group of SwinTransformer modules in the decoder;
[0017] For the nth layer feature fusion module, where n is a positive integer not equal to 1, the encoding result output by the (N+1-n)th group of Swin Transformer modules in the encoder and the decoding result output by the (n-1)th group of Swin Transformer modules in the decoder are used as inputs to obtain the fused feature output by the nth layer feature fusion module, and the fused feature is input into the nth group of Swin Transformer modules in the decoder.
[0018] According to a deep learning-based short-term forecasting method for severe convective winds provided by the present invention, the encoding result is input into a MaxViT module to obtain intermediate features of the historical data output by the MaxViT module, including...
[0019] The encoded result is processed sequentially through a first 1×1 convolutional layer, a channel-wise convolution, a channel attention mechanism, and a second 1×1 convolutional layer;
[0020] The input window attention mechanism is obtained by adding the output of the second 1×1 convolutional layer to the encoded result.
[0021] The output of the second 1×1 convolutional layer is added to the output of the window attention mechanism and then input into the grid attention mechanism;
[0022] The output of the window attention mechanism is added to the output of the grid attention mechanism and then input into the third 1×1 convolutional layer;
[0023] The output of the third 1×1 convolutional layer is added to the encoding result to obtain the intermediate feature.
[0024] According to a deep learning-based short-term forecasting method for severe convective winds provided by the present invention, the encoding result and the intermediate features are input into a feature fusion module to obtain the fused features of the historical data output by the feature fusion module, including:
[0025] After multiplying the two input data of the feature fusion module, the data is processed by adaptive average pooling and the fourth 1×1 convolutional layer in sequence and then input into the dual branch. The dual branch corresponds one-to-one with the two input data.
[0026] After passing through the fifth 1×1 convolutional layer and sigmoid layer in each branch, it is multiplied with the corresponding input data;
[0027] The results of multiplying the two branches are added together and then passed through a sixth 1×1 convolutional layer to obtain the fused features.
[0028] According to the deep learning-based short-term forecasting method for severe convective winds provided by the present invention, before inputting historical data into the encoder and obtaining the encoding result of the historical data output by the encoder, the method further includes:
[0029] The encoder, decoder, feature fusion module, and MaxViT module are trained as a whole;
[0030] The training loss function is:
[0031]
[0032]
[0033] in, The value of the loss function. The total number of regions divided, Let i be the weight of the i-th region. Encode the wind category for the i-th region. This represents the actual short-term gale results for the i-th region. For the short-term gale forecast result of the i-th region, when the wind force level is 7 or below... When the wind force is level 8 or 9, When the wind force is level 10 or above 1.
[0034] This invention also provides a deep learning-based short-term forecasting device for severe convective winds, comprising:
[0035] The encoding module is used to input historical data into the encoder to obtain the encoding result of the historical data output by the encoder. The historical data includes historical radar echo data and historical automatic station data.
[0036] A conversion module is used to input the encoding result into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module;
[0037] A fusion module is used to input the encoding result and the intermediate features into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module;
[0038] A decoding module is used to input the fused features into a decoder to obtain the decoding result of the fused features output by the decoder. The encoder and decoder include a Swing Transformer module.
[0039] The forecast module is used to obtain short-term gale forecast results based on the decoding results.
[0040] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the deep learning-based short-term forecasting method for severe convective winds as described above.
[0041] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep learning-based short-term forecasting method for severe convective winds as described above.
[0042] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based short-term forecasting method for severe convective winds as described above.
[0043] The present invention provides a method and apparatus for short-term forecasting of severe convective winds based on deep learning. It optimizes and extends the SwinTransformer architecture, combining an encoder, decoder, feature fusion module, and MaxViT module for short-term wind forecasting. The encoder and MaxViT module fully extract features from historical data, and the output of the encoder is used as the input of the MaxViT module. The feature fusion module fuses the information extracted by the encoder and MaxViT module, enhancing model performance. The model structure is simple, improving the accuracy and efficiency of short-term wind forecasting. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0045] Figure 1 This is a flowchart illustrating the short-term forecasting method for severe convective winds based on deep learning provided by the present invention.
[0046] Figure 2 This is a schematic diagram of the model of the deep learning-based short-term forecasting method for severe convective winds provided by the present invention.
[0047] Figure 3 This is a schematic diagram of the MaxViT module (multi-axis vision converter) in the model of the short-term forecasting method for strong convective winds based on deep learning provided by this invention.
[0048] Figure 4 This is a schematic diagram of the feature fusion module in the model of the short-term forecasting method for severe convective winds based on deep learning provided by this invention.
[0049] Figure 5 This is a schematic diagram of the structure of the short-term forecasting device for severe convective winds based on deep learning provided by the present invention. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0051] The following is combined with Figure 1This invention describes a deep learning-based short-term forecasting method for severe convective winds, comprising:
[0052] Step 101: Input historical data into the encoder to obtain the encoding result of the historical data output by the encoder. The historical data includes historical radar echo data and historical automatic station data.
[0053] Step 102: Input the encoding result into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module;
[0054] Step 103: Input the encoding result and the intermediate features into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module;
[0055] Step 104: Input the fused feature into the decoder to obtain the decoding result of the fused feature output by the decoder. The encoder and decoder include the Swing Transformer module.
[0056] Step 105: Based on the decoding results, obtain the short-term gale forecast results.
[0057] Historical data can be obtained through S-band weather radar and automatic weather stations.
[0058] Before using the model for short-term gale forecasting, the model is trained, and its structure is as follows: Figure 2 As shown.
[0059] Multi-source observation data, including automatic weather station data and radar-based data from 2022 to 2024, can be used. First, the maximum wind speed variable of the automatic weather station is interpolated to a unified grid, and events with a maximum wind speed of level 8 or higher at a 10-minute resolution within two hours are selected, accounting for 10% or more of the total number of pixels in the grid. At the same time, missing values and abnormal records are removed, and the radar-based data is accurately aligned with the automatic station observation data through spatiotemporal matching.
[0060] Based on wind force levels, maximum wind speeds were divided into three categories: level 7 and below, level 8-9 (strong winds), and level 10 and above (strong winds), assigned values of 0, 0.5, and 1 respectively. For the radar base data, the polar coordinate system was converted to a latitude-longitude grid, and the reflectivity intensity was normalized to the [0, 1] interval. The final dataset consisted of 5624 samples, including 5069 training samples and 555 test samples. The automatic weather station data in the dataset had a time resolution of 10 minutes, the radar data had a time resolution of 6 minutes, and both had a spatial resolution of 250 meters.
[0061] The model input consists of 10 consecutive frames of radar echo images and 6 consecutive frames of strong wind grid data within one hour, with a radar echo time resolution of 6 minutes and a strong wind time resolution of 10 minutes. The model output is 6 frames of strong wind images for the next hour (with a time resolution maintained at 10 minutes). Stochastic gradient descent (SGD) is used as the optimizer during training, with an initial learning rate of 0.01 and a total of 150 training epochs. Mean squared error (MSE) is used as the loss function, and a dynamic weighting mechanism is combined to optimize the training process. A single CPU extrapolation takes only 3 seconds. Tests show that its average accuracy is 0.59, providing strong technical support for improving the ability to issue severe weather warnings.
[0062] This embodiment optimizes and extends the Swin Transformer architecture, combining an encoder, decoder, feature fusion module, and MaxViT module for short-term gale forecasting. The encoder and MaxViT module fully extract features from historical data, with the encoder output serving as the input to the MaxViT module. The feature fusion module then fuses the information extracted by the encoder and MaxViT module, enhancing model performance. The model structure is simple, improving the accuracy and efficiency of short-term gale forecasting.
[0063] Based on the above embodiments, in this embodiment, the encoder and the decoder each include multiple Swing Transformer modules, and the number of Swing Transformer modules in the encoder and the decoder is the same, and they are divided into N groups, where N is a positive integer.
[0064] like Figure 2 As shown, the encoder and decoder each contain 6 Swing Transformer modules, each with a window size of 5, connected by skip connections, with 16 input channels and 6 output channels. The Swing Transformer modules in both the encoder and decoder are divided into three groups.
[0065] Based on the above embodiments, such as Figure 2 As shown, in this embodiment, historical data is input into the encoder to obtain the encoding result of the historical data output by the encoder, including:
[0066] The historical data is sequentially input into multiple sets of Swin Transformer modules in the encoder to obtain the encoding result of the historical data output by each set of Swin Transformer modules.
[0067] Based on the above embodiments, such as Figure 2As shown, in this embodiment, the encoding result is input into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module, including:
[0068] The encoding result of the historical data output by the last group of Swin Transformer modules is input into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module.
[0069] Based on the above embodiments, such as Figure 2 As shown, in this embodiment, the number of input feature fusion modules is the same as the number of groups of Swin Transformer modules in the encoder. The encoding result and the intermediate features are input into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module, including:
[0070] For the first-layer feature fusion module, the encoding result and the intermediate features are used as input to obtain the fused features output by the first-layer feature fusion module, and the fused features are input into the first group of SwinTransformer modules in the decoder;
[0071] For the nth layer feature fusion module, where n is a positive integer not equal to 1, the encoding result output by the (N+1-n)th group of Swin Transformer modules in the encoder and the decoding result output by the (n-1)th group of Swin Transformer modules in the decoder are used as inputs to obtain the fused feature output by the nth layer feature fusion module, and the fused feature is input into the nth group of Swin Transformer modules in the decoder.
[0072] Based on the above embodiments, such as Figure 3 As shown, in this embodiment, the encoding result is input into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module, including...
[0073] The encoded result is processed sequentially through a first 1×1 convolutional layer, a channel-wise convolution, a channel attention mechanism, and a second 1×1 convolutional layer;
[0074] The input window attention mechanism is obtained by adding the output of the second 1×1 convolutional layer to the encoded result.
[0075] The output of the second 1×1 convolutional layer is added to the output of the window attention mechanism and then input into the grid attention mechanism;
[0076] The output of the window attention mechanism is added to the output of the grid attention mechanism and then input into the third 1×1 convolutional layer;
[0077] The output of the third 1×1 convolutional layer is added to the encoding result to obtain the intermediate feature.
[0078] The converter in this model is an optimization and extension of the MaxViT architecture. To significantly improve computational efficiency, the original feedforward neural network structure was removed. Meanwhile, to further enhance model performance, convolutional layers and skip connections were incorporated. In summary, MaxViT data is progressively processed through 1×1 convolutions, channel-wise convolutions, channel attention mechanisms, and another 1×1 convolution, then sequentially through Block-SA (window attention) and Grid-SA (grid attention) modules, finally passing through a 1×1 convolution and adding it to the initial input to obtain the final output.
[0079] Based on the above embodiments, such as Figure 4 As shown, in this embodiment, the encoding result and the intermediate features are input into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module, including:
[0080] After multiplying the two input data of the feature fusion module, the data is processed by adaptive average pooling and the fourth 1×1 convolutional layer in sequence and then input into the dual branch. The dual branch corresponds one-to-one with the two input data.
[0081] After passing through the fifth 1×1 convolutional layer and sigmoid layer in each branch, it is multiplied with the corresponding input data;
[0082] The results of multiplying the two branches are added together and then passed through a sixth 1×1 convolutional layer to obtain the fused features.
[0083] The feature fusion module merges data from the encoder, MaxViT module, and decoder, and inputs the fused data into the decoder.
[0084] The two input data are first multiplied and then subjected to adaptive average pooling to obtain the result. Then, the input is adjusted by a 1×1 convolution and becomes a double branch. Each branch is convolved and then enters a sigmoid layer and is multiplied with the original input. Finally, the results of the two branches are added together and then adjusted by convolution before being output.
[0085] Based on the above embodiments, this embodiment further includes the following step before inputting historical data into the encoder and obtaining the encoding result of the historical data output by the encoder:
[0086] The encoder, decoder, feature fusion module, and MaxViT module are trained as a whole;
[0087] The training loss function is:
[0088]
[0089]
[0090] in, The value of the loss function. The total number of regions divided, Let i be the weight of the i-th region. Encode the wind category for the i-th region. This represents the actual short-term gale results for the i-th region. For the short-term gale forecast result of the i-th region, when the wind force level is 7 or below... When the wind force is level 8 or 9, When the wind force is level 10 or above 1.
[0091] The optimizer chosen for model training was SGD, with a weight decay of 0.0001 and a momentum of 0.9. The loss function was a weighted MSE loss, with a weight of 1 for regions categorically encoded as 0 and 10 for other regions.
[0092] The following describes the deep learning-based short-term forecasting device for severe convective winds provided by this invention. The deep learning-based short-term forecasting device for severe convective winds described below can be referred to in correspondence with the deep learning-based short-term forecasting method for severe convective winds described above.
[0093] like Figure 5 As shown, the device includes an encoding module 501, a conversion module 502, a fusion module 503, a decoding module 504, and a prediction module 505, wherein:
[0094] The encoding module 501 is used to input historical data into the encoder to obtain the encoding result of the historical data output by the encoder. The historical data includes historical radar echo data and historical automatic station data.
[0095] The conversion module 502 is used to input the encoding result into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module;
[0096] The fusion module 503 is used to input the encoding result and the intermediate features into the feature fusion module to obtain the fused features of the historical data output by the feature fusion module;
[0097] The decoding module 504 is used to input the fused features into the decoder to obtain the decoding result of the fused features output by the decoder. The encoder and decoder include the Swing Transformer module.
[0098] The forecast module 505 is used to obtain short-term gale forecast results based on the decoding results.
[0099] This embodiment optimizes and extends the Swin Transformer architecture, combining an encoder, decoder, feature fusion module, and MaxViT module for short-term gale forecasting. The encoder and MaxViT module fully extract features from historical data, with the encoder output serving as the input to the MaxViT module. The feature fusion module then fuses the information extracted by the encoder and MaxViT module, enhancing model performance. The model structure is simple, improving the accuracy and efficiency of short-term gale forecasting.
[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A deep learning-based severe convective gale short-impending forecast method, characterized in that, The method comprises the following steps: inputting historical data into an encoder to obtain an encoding result of the historical data output by the encoder, the historical data comprising historical radar echo data and historical automatic station data; inputting the encoding result into a MaxViT module to obtain intermediate features of the historical data output by the MaxViT module; inputting the encoding result and the intermediate features into a feature fusion module to obtain fusion features of the historical data output by the feature fusion module; inputting the fusion features into a decoder to obtain a decoding result of the fusion features output by the decoder, the encoder and the decoder comprising Swin Transformer modules; obtaining a short-impending gale prediction result according to the decoding result.
2. The deep learning-based severe convective gale short-impending forecast method according to claim 1, characterized in that, The encoder and the decoder each comprise a plurality of Swin Transformer modules, and the number of Swin Transformer modules in the encoder and the decoder is the same, and the Swin Transformer modules are divided into N groups, N being a positive integer.
3. The deep learning-based severe convective gale short-impending forecast method according to claim 2, characterized in that, The method comprises the following steps of inputting historical data into an encoder to obtain an encoding result of the historical data output by the encoder: inputting the historical data into a plurality of groups of Swin Transformer modules in the encoder in sequence to obtain an encoding result of the historical data output by each group of Swin Transformer modules.
4. The deep learning-based severe convective gale short-impending forecast method according to claim 3, characterized in that, The method comprises the following steps of inputting the encoding result into a MaxViT module to obtain intermediate features of the historical data output by the MaxViT module: inputting the encoding result of the historical data output by the last group of Swin Transformer modules into the MaxViT module to obtain the intermediate features of the historical data output by the MaxViT module.
5. The deep learning-based severe convective gale short-impending forecast method according to claim 3, characterized in that, The number of input feature fusion modules is the same as the number of groups of Swin Transformer modules in the encoder, and the method comprises the following steps of inputting the encoding result and the intermediate features into a feature fusion module to obtain fusion features of the historical data output by the feature fusion module: for a first layer feature fusion module, the encoding result and the intermediate features are taken as inputs to obtain fusion features output by the first layer feature fusion module, and the fusion features are input into a first group of Swin Transformer modules in the decoder; for an nth layer feature fusion module, n being a positive integer not equal to 1, the encoding result output by an N+1-nth group of Swin Transformer modules in the encoder and the decoding result output by an (n-1)th group of Swin Transformer modules in the decoder are taken as inputs to obtain fusion features output by the nth layer feature fusion module, and the fusion features are input into an nth group of Swin Transformer modules in the decoder.
6. The deep learning-based severe convective gale short-impending forecast method according to any one of claims 1-5, characterized in that, The method comprises the following steps of inputting the encoding result into a MaxViT module to obtain intermediate features of the historical data output by the MaxViT module: The encoding result is sequentially processed through a first 1*1 convolution layer, a channel-by-channel convolution, a channel attention mechanism, and a second 1*1 convolution layer; The output of the second 1*1 convolution layer is added to the encoding result and input into a window attention mechanism; The output of the second 1*1 convolution layer is added to the output of the window attention mechanism and input into a grid attention mechanism; The output of the window attention mechanism is added to the output of the grid attention mechanism and input into the third 1*1 convolution layer; The output of the third 1*1 convolution layer is added to the encoding result to obtain the intermediate feature.
7. The deep learning-based severe convective gale short-impending forecast method according to any one of claims 1-5, characterized in that, The encoding result and the intermediate feature are input into a feature fusion module to obtain a fusion feature of the historical data output by the feature fusion module, including: After the two input data of the feature fusion module are multiplied, they are sequentially processed through adaptive average pooling and a fourth 1*1 convolution layer and then input into a double-branch, which corresponds to the two input data one-to-one; After sequentially passing through a fifth 1*1 convolution layer and a sigmoid layer in each branch, the output is multiplied by the corresponding input data; The multiplication results of the double-branch are added and then input into a sixth 1*1 convolution layer to obtain the fusion feature.
8. The deep learning-based severe convective gale short-impending forecast method according to any one of claims 1-5, characterized in that, Before the historical data is input into an encoder to obtain an encoding result of the historical data output by the encoder, the method further includes: The encoder, the decoder, the feature fusion module, and the MaxViT module are trained as a whole; The loss function of the training is: ; ; wherein, is a loss function value, is a total amount of divided regions, is a weight of the i-th region, is a gale category code of the i-th region, is an actual short-impending gale result of the i-th region, is a short-impending gale prediction result of the i-th region, when the wind force level is 7 or less , when the wind force level is 8 and 9 , when the wind force level is 10 or more 1.
9. A deep learning-based severe convective gale short-impending forecast device, characterized by, The method includes: An encoding module is configured to input historical data into an encoder to obtain an encoding result of the historical data output by the encoder, the historical data including historical radar echo data and historical automatic station data; A conversion module is configured to input the encoding result into a MaxViT module to obtain an intermediate feature of the historical data output by the MaxViT module; A fusion module is configured to input the encoding result and the intermediate feature into a feature fusion module to obtain a fusion feature of the historical data output by the feature fusion module; A decoding module is configured to input the fusion feature into a decoder to obtain a decoding result of the fusion feature output by the decoder, the encoder and the decoder including a Swin Transformer module; A prediction module is configured to obtain a short-impending gale prediction result according to the decoding result.
10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the deep learning-based short-impending severe convective gale prediction method of any one of claims 1 to 8. The processor executes the program to implement the deep learning-based short-impending severe convective gale prediction method of any one of claims 1 to 8.