A thunderstorm gale approaching forecast method and system based on unet multi-modal integration

Abstract

The application discloses a thunderstorm gale approaching prediction method and system based on UNET multi-modal integration, including specific extraction of thunderstorm gale samples, and construction of a data set for thunderstorm gale events; local data enhancement of thunderstorm gale samples in the data set is implemented for the spatiotemporal evolution characteristics of severe convective systems; a plurality of prediction models based on the UNET network and embedded with different attention mechanisms are constructed; a loss function that fuses extreme gale false alarm, false alarm penalty and weighted wind speed is used to train each prediction model; an integrated strategy for reducing false alarm risk is used to integrate the outputs of each prediction model to generate a final thunderstorm gale integrated prediction result.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting, and in particular relates to a method and system for nowcasting thunderstorms and strong winds based on UNET multimodal integration. Background Technology

[0002] Thunderstorms and strong winds refer to convective winds (usually defined as maximum wind speeds ≥17.2 m / s) that accompany lightning. Due to their sudden onset, rapid development, short duration, and significant localization, they have always been a challenging and crucial aspect of meteorological forecasting. Accurate forecasting of thunderstorms and strong winds is of great importance for disaster prevention and mitigation, and for protecting people's lives and property.

[0003] In recent years, neural network models, represented by the UNET network, have been widely used in radar echo extrapolation and precipitation nowcasting due to their advantages of low computational cost, simple architecture, and ability to effectively capture multi-scale spatial characteristics of weather systems. However, directly applying the UNET network to thunderstorm and gale nowcasting still faces two major technical challenges:

[0004] First, the sample size for thunderstorm and strong wind events is extremely sparse, resulting in a severe imbalance. Thunderstorms and strong winds, as extreme severe convective events, occur with a very low probability in historical observation data, leading to a significant disparity between positive and negative samples. During model training, the model is easily dominated by majority class (non-windy weather) samples, making it difficult to learn the localized characteristics of thunderstorms and strong winds and their complex correspondence with strong radar echoes. This results in serious false alarms and missed predictions in the forecast results.

[0005] Secondly, traditional two-dimensional convolutional architecture models are insufficient in modeling the spatiotemporal evolution of severe convective systems. These models often lack detail in their severe convective forecasts because, spatially, hazardous weather fields typically constitute a small proportion of the overall field. These models may focus on the overall field state while neglecting local details. Furthermore, the formation, dissipation, and rapid movement of thunderstorms and strong winds is a complex dynamic process involving the interaction of systems at different scales. Existing traditional models excel at extracting spatial patterns, but their inherent network structure limits their ability to capture rapidly evolving features in the temporal dimension. They struggle to effectively characterize the dynamic processes of convective system development, dissipation, movement, and inter-scale energy transfer, leading to underreporting of hazardous wind areas. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a method and system for nowcasting thunderstorms and gales based on UNET multimodal integration.

[0007] The technical solution of the present invention is as follows:

[0008] A method for nowcasting thunderstorms and strong winds based on UNET multimodal ensemble, including...

[0009] Based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data, and terrain height data, and by considering the joint screening of peak wind speed and lightning occurrence, specific extraction of thunderstorm wind samples is carried out to construct a dataset for thunderstorm wind events.

[0010] Local data augmentation was performed on the thunderstorm and strong wind samples in the dataset to target the spatiotemporal evolution characteristics of severe convective systems;

[0011] Multiple forecasting models based on the UNET network and embedded with different attention mechanisms are constructed. The forecasting models adaptively focus on key spatiotemporal features related to the generation and dissipation of thunderstorms and strong winds using the embedded attention mechanisms.

[0012] Each forecast model was trained using a loss function that combines the penalty weights of false alarms and missed reports of extreme gale events. The loss function applies a higher penalty weight to missed reports of thunderstorm gale events than to false alarms and assigns higher weights to thunderstorm gale samples with high wind speeds.

[0013] An integrated strategy aimed at reducing the risk of missed forecasts was adopted, which combined the outputs of various forecast models to generate the final integrated forecast result for thunderstorms and gales.

[0014] Furthermore, the specific method for constructing a dataset for thunderstorm wind events, based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data, and terrain data, and by considering the joint screening of peak wind speed and lightning occurrence, includes the following:

[0015] Based on the observation time of the radar composite reflectivity mosaic data, identify potential high-wind stations in the ground minute instantaneous wind speed station observation data where the maximum ground wind speed observation data in the past hour is greater than or equal to the first threshold.

[0016] Within a preset range centered on the potential high-wind station, if lightning is determined to have occurred based on lightning location data, or if the radar combined reflectivity factor in the radar combined reflectivity factor mosaic data is greater than or equal to the second threshold, then the radar combined reflectivity factor mosaic data, the ground minute-by-minute instantaneous wind speed station observation data, and the terrain height data corresponding to the potential high-wind station are used as thunderstorm high-wind samples and added to the dataset; the features in the thunderstorm high-wind samples are the radar combined reflectivity factor mosaic data minute-by-minute over the past hour, the ground maximum wind speed observation data over the past hour, and the terrain height data, all labeled as ground maximum wind speed observation data for the next hour.

[0017] Furthermore, the specific method for performing local data augmentation on the concentrated thunderstorm and strong wind samples in the dataset to target the spatiotemporal evolution characteristics of severe convective systems includes:

[0018] Perform geometric transformations on features and labels, including random translation, random rotation, and random mirroring.

[0019] The radar composite reflectivity factor mosaic data, ground maximum wind speed observation data, and ground maximum wind speed observation data in the label are randomly scaled, the radar composite reflectivity factor mosaic data in the feature are graded intensity perturbation, and the feature is randomly noise-added.

[0020] Furthermore, the attention mechanism includes at least the attention gate mechanism, the CBAM attention mechanism, the multi-head attention mechanism, and the spatial pyramid attention mechanism.

[0021] Furthermore, the weighted loss function that enables the forecast model to emphasize the impact of extreme wind events is constructed based on the mean absolute error (MAE), and at least simultaneously incorporates tiered observation weights, missed report penalty weights, and false alarm penalty weights:

[0022] The graded observation weights are assigned based on the interval in which the observed wind speed value is located. The higher the threshold of the interval to which the observed wind speed belongs, the greater the graded observation weight is assigned to the thunderstorm wind sample.

[0023] The missed detection penalty weight is a high-multiplier penalty weight applied when the forecast model misses a wind speed exceeding the preset strong wind threshold.

[0024] The false alarm penalty weight is a penalty weight that is lower than the missed alarm penalty weight when the forecast model makes a false alarm for wind speeds exceeding the preset strong wind threshold.

[0025] Furthermore, the expression for the weighted loss function that makes the forecast model take into account the impact of extreme wind events is:

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] In the formula, and These represent the predicted maximum wind speed output by the forecast model and the corresponding observed maximum wind speed, respectively; N represents the total number of thunderstorm wind samples at all grid points across all forecast times. Based on maximum wind speed observation data Hierarchical observation weights; and These are the first and second maximum wind speed thresholds, respectively; The penalty weight for unreported extreme winds; The penalty weight for reporting extreme wind speeds; , , The maximum wind speeds are weighted for levels 1, 2, and 3, respectively. Weight for underreporting; The weights are for false alarms; the weight parameters in the weighted loss function are adaptively adjusted in multiple training rounds. Based on the forecast model's performance in forecasting thunderstorm and strong wind events in previous training rounds, the values ​​of the maximum wind speed weights (levels 1, 2, and 3), the missed alarm weights, and the false alarm weights in the weighted loss function are dynamically adjusted.

[0031] Furthermore, the specific method for generating the final integrated forecast product of thunderstorms and gales by combining the outputs of various forecast models using an integrated strategy aimed at reducing the risk of underreporting includes:

[0032] For each grid point in the forecast space, the maximum wind speed value predicted by all forecast models at that grid point is taken as the final integrated forecast wind speed for that grid point.

[0033] A thunderstorm and gale nowcasting system based on UNET multimodal ensemble includes a dataset construction module, a local data augmentation module, a forecast model construction module, a training module, and an ensemble module;

[0034] The dataset construction module is used to extract specific samples of thunderstorm wind events based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data and terrain height data, and by considering the joint screening of wind speed peak and lightning occurrence.

[0035] The local data augmentation module is used to perform local data augmentation on the thunderstorm and strong wind samples in the dataset, focusing on the spatiotemporal evolution characteristics of the severe convective system.

[0036] The forecast model construction module is used to construct multiple forecast models based on the UNET network and embedded with different attention mechanisms. The forecast models adaptively focus on key spatiotemporal features related to the generation and dissipation of thunderstorms and strong winds using the embedded attention mechanisms.

[0037] The training module is used to train each forecast model using a loss function that integrates the false alarms of extreme gale events and the weighted wind speed with penalties for missed reports. The loss function applies a higher penalty weight to missed reports of thunderstorm gale events than to false alarms and assigns higher weights to thunderstorm gale samples with high wind speeds.

[0038] The integration module is used to adopt an integration strategy aimed at reducing the risk of missed reports, and integrate the outputs of various forecast models to generate the final integrated forecast result for thunderstorms and gales.

[0039] An electronic device includes a memory and a processor, the memory storing a computer program, the processor being configured to invoke and run the computer program stored in the memory to perform the method as described in any of the preceding methods.

[0040] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] This invention provides a method and system for nowcasting thunderstorms and strong winds based on UNET multimodal ensemble. This method solves the problems of extremely low positive sample ratio and sample imbalance by constructing a multi-source data joint screening method for specific extraction of thunderstorm and strong wind samples, significantly expanding the effective training sample. Specifically, this invention no longer relies on single hourly maximum wind speed observations, but combines minute-by-minute peak instantaneous wind speeds, lightning location data, and radar combined reflectivity factors to construct a more refined sample screening standard. By using minute-by-minute maximum values ​​instead of hourly values, and using combined reflectivity as an effective supplement to lightning data, the number of thunderstorm and strong wind samples is significantly increased. This design effectively expands the thunderstorm and strong wind sample library, enabling the forecast model to capture more key feature information related to thunderstorms and strong winds, laying a data foundation for the accurate training of subsequent models.

[0043] This invention addresses the problem of insufficient generalization ability of models for localized and diverse thunderstorm and gale events by implementing local data augmentation based on the spatiotemporal evolution characteristics of severe convective systems, thereby enhancing the model's robustness. For key local areas where thunderstorms and gales occur, this invention designs an augmentation process including stochastic geometric transformations (translation, rotation, and flipping), intensity scaling and perturbation, and random noise addition. This design enriches the diversity of thunderstorm and gale samples by simulating the spatial movement, morphological changes, intensity fluctuations, and observation errors of convective systems in the actual atmosphere. This forces the forecast model to learn more essential and robust feature representations, thereby suppressing noise interference and significantly improving the model's forecasting and generalization capabilities for different types of thunderstorm and gale events.

[0044] This invention addresses the shortcomings of traditional spatial models in modeling local details and rapid dynamic processes by embedding multiple attention mechanisms into the UNET network, thereby enhancing the ability to capture key spatiotemporal features of thunderstorms and strong winds. The invention constructs various forecasting models, each embedding different attention mechanisms such as attention gates, CBAM, multi-head attention, and spatial pyramids. These mechanisms work synergistically, enabling the forecasting models to adaptively focus on key spatiotemporal features related to the generation, dissipation, and evolution of thunderstorms and strong winds.

[0045] This invention addresses the problems of low attention to high-impact, low-probability events and severe underreporting in forecast models by designing a loss function that integrates underreporting penalties for extreme gale warnings and underreporting, along with weighted wind speed. It effectively balances underreporting and underreporting. This design abandons traditional mean squared error or mean absolute error loss functions, employing a loss function that constrains the model in three dimensions: hierarchical observation weights assign higher weights to thunderstorm gale samples with higher wind speeds, forcing the model to focus on extreme gale events; penalty weights are applied to underreporting, significantly improving the model's sensitivity to thunderstorm gale events; and penalty weights are applied to underreporting, preventing the model from indiscriminately expanding the forecast range to reduce underreporting.

[0046] This invention employs an integrated strategy aimed at reducing the risk of missed predictions, combining the outputs of various attention models to further improve the reliability and accuracy of the final forecast product. The invention integrates the forecast results of multiple forecast models embedded with different attention mechanisms, taking the maximum value of the predicted wind speed from all models as the final result for each forecast grid point. This design combines the advantages of different attention mechanisms, enabling the final integrated forecast product to maintain sensitivity to areas of strong winds while significantly outperforming any single model or traditional method.

[0047] In this invention, the weighted loss function is adaptively adjusted by setting weight parameters, allowing the weights of graded observations, missed reporting penalties, and false alarms to be dynamically optimized based on the model's forecast performance in each training round. Compared to a loss function with fixed weights, this method achieves self-optimization of the model for thunderstorm and gale forecasting without the need for repeated manual parameter tuning: when severe missed reports occur, the weights of high wind speeds and the missed reporting penalties are automatically increased; when false alarms intensify, the false alarm penalties are automatically increased. This allows the forecasting model to continuously balance its sensitivity to thunderstorms and gales and its forecast accuracy during training, effectively solving the problem that fixed weights are difficult to adapt to different thunderstorm and gale scenarios. This further improves the adaptability of the loss function to thunderstorm and gale forecasting tasks, enabling the trained forecasting model to better meet the forecasting needs for extreme weather in actual meteorological forecasting. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the nowcasting method for thunderstorms and strong winds based on UNET multimodal ensemble.

[0049] Figure 2 This is a schematic diagram showing the distribution of ground-based wind speed stations within the target area;

[0050] Figure 3 Flowchart of the data augmentation mechanism;

[0051] Figure 4 This is a schematic diagram of the UNET network structure;

[0052] Figure 5 This is a mosaic of hourly radar composite reflectivity from 19:00 on July 16, 2025 to 03:00 on July 17, 2025. Figure 5 Figures a through i represent the period from 19:00 on the 16th to 03:00 on the 17th, respectively.

[0053] Figure 6 This is a UNET-ENSEM 1-hour maximum wind speed forecast and corresponding station observation maps for the period from 19:00 on July 16, 2025 to 03:00 on July 17, 2025. The color-coded values ​​represent the UNET-ENSEM 1-hour maximum wind speed forecast, and the red dots indicate the locations of stations with observed maximum wind speeds ≥17.2 m / s for the corresponding 1-hour period. Figure 6 Figures a through i represent the period from 20:00 on the 16th to 04:00 on the 17th, respectively.

[0054] Figure 7 Same as the UNET-1 1-hour maximum wind speed forecast and corresponding station observation map for the maximum wind speed from 19:00 on July 16, 2025 to 03:00 on July 17, 2025; the color is the UNET-ENSEM 1-hour maximum wind speed forecast value, and the red dots are the locations of stations with observed maximum wind speeds ≥17.2 m / s for the corresponding 1-hour period. Figure 6 Figures a through i represent the period from 20:00 on the 16th to 04:00 on the 17th, respectively.

[0055] Figure 8 Same as the UNET-2 1-hour maximum wind speed forecast and corresponding station observation map for the maximum wind speed from 19:00 on July 16, 2025 to 03:00 on July 17, 2025; the color is the UNET-ENSEM 1-hour maximum wind speed forecast value, and the red dots are the locations of stations with observed maximum wind speeds ≥17.2 m / s for the corresponding 1-hour period. Figure 6 Figures a through i represent the period from 20:00 on the 16th to 04:00 on the 17th, respectively. Detailed Implementation

[0056] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0057] Example 1:

[0058] This invention provides a method for nowcasting thunderstorms and strong winds based on UNET multimodal ensemble, such as... Figure 1 As shown, including

[0059] Based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data, and terrain height data, and by considering the joint screening of peak wind speed and lightning occurrence, specific extraction of thunderstorm wind samples is carried out to construct a dataset for thunderstorm wind events in order to address the problem of its extremely low positive sample ratio.

[0060] Generally, the horizontal resolution of the radar composite reflectivity factor mosaic data can reach 1 km, the temporal resolution can reach 6 min, and the latitude and longitude range can be selected according to the target, for example, the latitude and longitude range can be selected as 12.2-54.2°N, 73-135°E.

[0061] The time resolution of the above-mentioned ground-based instantaneous wind speed station observation data can reach 1 minute;

[0062] The time resolution of the aforementioned lightning location data can reach 6 seconds;

[0063] The horizontal resolution of the above terrain height data can reach 30m.

[0064] Local data augmentation is performed on the thunderstorm and strong wind samples in the dataset to target the spatiotemporal evolution characteristics of strong convective systems. By simulating the movement, rotation, intensity changes and observation errors of convective systems, augmented samples are generated to improve the model's ability to generalize and recognize thunderstorm and strong wind features.

[0065] Multiple forecasting models based on the UNET network and embedded with different attention mechanisms are constructed. The forecasting models use the embedded attention mechanisms to adaptively focus on key spatiotemporal features related to the generation and dissipation of thunderstorms and strong winds, so as to overcome the shortcomings of traditional spatial models in modeling local details and rapid dynamic processes.

[0066] Each forecast model was trained using a loss function that combines the penalty weights of false alarms and missed reports of extreme gale events. The loss function imposes a higher penalty weight on missed reports of thunderstorm gale events than on false alarms and assigns higher weights to thunderstorm gale samples with high wind speeds, thereby forcing the model to learn and pay attention to low-probability but high-impact weather events such as thunderstorm gale events.

[0067] An integrated strategy aimed at reducing the risk of missed forecasts was adopted, which combined the outputs of various forecast models to generate the final integrated forecast result for thunderstorms and gales.

[0068] Furthermore, based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data, and topographic data, and by considering the joint screening of peak wind speed and lightning occurrence, specific methods are used to extract thunderstorm wind samples and construct a dataset for thunderstorm wind events.

[0069] Based on the observation time of the radar composite reflectivity mosaic data, identify potential high-wind stations in the ground minute instantaneous wind speed station observation data where the maximum ground wind speed observation data in the past hour is greater than or equal to the first threshold.

[0070] Within a preset range centered on a potential high-wind station, if lightning is determined to have occurred based on lightning location data, or if the radar combined reflectivity factor in the radar combined reflectivity factor mosaic data is greater than or equal to the second threshold, then the radar combined reflectivity factor mosaic data, the ground minute-by-minute instantaneous wind speed station observation data, and the terrain height data corresponding to the potential high-wind station are used as thunderstorm high-wind samples and added to the dataset. The features in the thunderstorm high-wind samples, also known as the "feature field," are the radar combined reflectivity factor mosaic data minute-by-minute over the past hour, the ground maximum wind speed observation data over the past hour, and the terrain height data, all labeled as ground maximum wind speed observation data for the next hour.

[0071] The following example illustrates the construction of a dataset:

[0072] This example selects Anhui Province and its surrounding areas (29.0-35.05°N, 114.5-120°E) as the target region. It uses national mosaic data of radar combined reflectivity factors and minute-by-minute instantaneous wind speed station observation data (including national and regional stations, totaling over 9000 stations) output from the China Meteorological Administration's Severe Weather Alert and Nowcasting System (SWAN3.0) from April to September 2019-2025. Figure 2 The dataset was constructed using lightning location data from the China Meteorological Administration's ADTD (Advanced Direction-finding and Time-of-arrival Detection) and terrain height data from the DEM (Digital Elevation Model). Detailed data information is shown in Table 1.

[0073] Table 1

[0074] SWAN 3.0 Radar Combined Reflectivity Factor National Mosaic Data 12.2–54.2°N, 73–135°E 1km×1km 6 minutes Ground-based minute-by-minute instantaneous wind speed station observations 29.0—35.05°N, 114.5—120.0°E \ 1 minute China Meteorological Administration ADTD lightning location data 29.0—35.05°N, 114.5—120.0°E \ 6 seconds Terrain height data 29–36°N, 114–121°E 30m×30m \

[0075] The steps for building the dataset are as follows:

[0076] 1) Based on the observation time of the radar composite reflectivity mosaic data (0, 6, 12, 18, 24, 30, 36, 42, 48, 54 minutes per hour), the maximum instantaneous wind speed of each station in the past hour is calculated as the maximum wind speed of that station in the past hour.

[0077] 2) For the target area (29.0-35.05°N, 114.5-120°E), the criteria for selecting thunderstorm wind samples are: "at least two stations in the target area have observed maximum wind speeds ≥ 8 (maximum wind speed ≥ 17.2 m / s) in the past hour, and lightning activity has been observed within a 50 km radius of these stations." If the lightning data file is missing or contains outliers, the combined reflectivity factor is used to replace lightning activity. Current research indicates a close relationship between lightning activity and the combined reflectivity factor and echo top height. A combined reflectivity factor ≥ 35 dBZ and an echo top height exceeding the −10°C layer are considered the "switch" for lightning outbreaks. Lightning activity is most concentrated in the 40-50 dBZ range during the mature stage. Therefore, this method uses: "at least two stations in the target area have observed maximum wind speeds ≥ 8 (maximum wind speed ≥ 17.2 m / s) in the past hour, and the radar combined reflectivity factor within a 50 km radius of these stations is ≥ 45 dBZ" as the selection criterion. The screening results showed a total of 15,386 thunderstorm and strong wind samples within the target area from April to September 2019 to 2025. Each sample's features included radar combined reflectivity factor mosaic data (10 in total) for every 6 minutes of the past hour, surface maximum wind speed observation data for the past hour, and terrain height data. The label for each sample was surface maximum wind speed observation data for the next hour. For example, if thunderstorm and strong winds occurred in the past hour up to 08:00 (Beijing time) on June 1, 2025, the sample's features would be radar combined reflectivity factor mosaic data for every 6 minutes of the past hour from 06:00 to 07:00 on June 1, 2025, surface maximum wind speed observation data for the past hour from 06:00 to 07:00, and terrain height data for the past hour. The corresponding label would be surface maximum wind speed observation data for the past hour from 07:00 to 08:00. After sample screening, a bilinear interpolation algorithm was used to uniformly interpolate the combined reflectivity factor, maximum wind speed observation, and terrain height data to the target area, forming a grid of 606×551 (latitudinal × longitudinal) points with a spatial resolution of 1 km.

[0078] 3) Based on the chronological order, a total of 9,664 samples from April to September 2019 to 2023 were selected as the training set, 2,417 samples from April to September 2024 were selected as the validation set, and 3,305 samples from April to September 2025 were selected as the test set.

[0079] Furthermore, since the area of ​​thunderstorms and strong winds in the thunderstorm and strong wind samples is extremely small compared to the target area, the method of this invention employs a local data augmentation mechanism for the areas where thunderstorms and strong winds occur to expand the thunderstorm and strong wind samples and improve the model's forecasting and generalization capabilities.

[0080] The specific methods for performing local data augmentation on the thunderstorm and strong wind samples in the dataset, focusing on the spatiotemporal evolution characteristics of severe convective systems, include:

[0081] Before local data augmentation, the dataset is generally divided into a training set for model training, a validation set for model validation, and a test set for model testing.

[0082] Geometric transformations, including random translation, random rotation, and random mirror flipping, are performed on all features and labels of thunderstorm and strong wind samples in the training set to simulate the spatial movement and morphological changes of convective systems.

[0083] like Figure 3 As shown, specifically, based on the random numbers output by the system, the features and labels belonging to the thunderstorm and strong wind samples are randomly translated (i.e., the entire field is randomly translated in all directions, up, down, left, and right, to simulate the spatial movement of the convective system), randomly rotated (with the center of the target area as the origin, the entire field is rotated at random angles to simulate convective systems with different directions), and randomly mirrored (the entire field is randomly flipped horizontally or vertically to enhance the symmetry of the samples).

[0084] The radar composite reflectivity factor mosaic data, ground maximum wind speed observation data, and ground maximum wind speed observation data in the label are randomly scaled, the radar composite reflectivity factor mosaic data in the feature are graded intensity perturbation, and noise is randomly added to all features.

[0085] The following example illustrates random scaling, graded intensity perturbation, and random noise addition:

[0086] like Figure 3 As shown, based on the surface maximum wind speed observation data for the next hour in the labels, areas with maximum wind speeds ≥17.2 m / s (maximum wind force ≥ level 8) are identified, and all features except terrain height data within the corresponding areas are extracted. Using the random numbers output by the system, random scaling (i.e., multiplying the feature field and labels simultaneously by a random scaling factor to simulate the expansion and contraction of the convective system), graded intensity perturbation (i.e., multiplying different intensity ranges in the radar combined reflectivity factor mosaic data, including <15 dBZ, 15~25 dBZ, 25~35 dBZ, 35~45 dBZ, and >45 dBZ, by different random scaling factors, while keeping other features and labels unchanged, to enhance model robustness), and random noise addition (i.e., adding random noise to the feature field to simulate the impact of radar and ground observation errors on the characteristics of high-wind areas) are performed sequentially.

[0087] Furthermore, the enhanced thunderstorm wind samples were standardized, while the thunderstorm wind samples in the validation and test sets were not augmented, but only standardized.

[0088] Furthermore, the method of this invention builds multiple prediction models based on the UNET network, which integrate different attention mechanisms. These prediction models are all composed of modules such as the UNET main network, attention mechanisms, and loss functions.

[0089] UNET main network structure as follows Figure 4 As shown in the figure, the blue / white background boxes represent feature information (i.e., feature-map), where F, H, W, and D represent the dimensionality of the feature information; the blue arrows represent 3×3 convolutions used for feature extraction; the pink arrows represent skip-connections used for fusing high- and low-dimensional feature information; the yellow arrows represent max-pooling operations used to reduce dimensionality; the orange arrows represent upsampling used to restore data dimensionality; and the green arrows indicate that the feature information of this layer is input into the attention mechanism. The network follows an encoder-decoder structure: the encoder part is responsible for feature extraction, consisting of convolutional and downsampling operations; the decoder part, consisting of upsampling and skip-connection layers, is responsible for restoring the original resolution of the feature information. Since this method focuses on the generation and dissipation of strong convective systems that cause thunderstorms and strong winds, an attention mechanism is added to the UNET structure. This allows the neural network to select feature information highly correlated with thunderstorms and strong winds for processing, obtaining more effective information, suppressing noise, and improving model efficiency. Figure 2 When some low-dimensional feature information extracted during the downsampling process is sent to the right decoding part through the jump connection layer, it is first filtered by the attention mechanism. Similarly, the high-dimensional feature information in the upsampling process is also filtered by the attention mechanism. Then, the high- and low-dimensional feature information that have passed the filtering are fused.

[0090] Attention mechanisms include at least the attention gate mechanism, the CBAM attention mechanism, the multi-head attention mechanism, and the spatial pyramid attention mechanism.

[0091] The expression for the attention gate mechanism is: The two inputs are features from a shallow, high-resolution layer (i.e., input feature x) and features from a deep, low-resolution layer (i.e., gated signal g). x and g are adjusted to have the same number of channels through a 1×1 convolution and then summed element-wise. ), and then through the ReLU activation function ( ), 1×1 convolution and the Sigmoid function ( ) to obtain the attention coefficient matrix ( The feature information is then restored to its original size using a resampling module, and finally the attention coefficient matrix is ​​used to perform weighted calculations on the feature information of this layer.

[0092] The CBAM attention mechanism consists of two parts: spatial attention and channel attention.

[0093] The expression for channel attention is: After the feature field F is input, it first enters the channel attention, and global average pooling (AvgPool, GAP) and global max pooling (MaxPool, GMP) are performed based on the width and height of the feature field. Then, it passes through a multilayer perceptron (MLP) to obtain the channel attention weights, and then passes through the sigmoid function ( The normalized attention weights are obtained and then multiplied channel by channel to be weighted onto the original input feature field, thus completing the recalibration of the original features by the channel attention.

[0094] The expression for spatial attention is: To obtain attention features in the spatial dimension, the feature field F output by channel attention is also subjected to GAP and GMP based on its width and height, transforming the feature dimension from H×W to 1×1, and then passed through a convolution with a 7×7 kernel. After applying the ReLU activation function to reduce the dimensionality of the feature field, it is then boosted back to its original dimensionality after a single convolution, and finally activated by the Sigmoid activation function. The standardized feature field is merged with the feature field output by the channel attention, thereby completing the recalibration of the feature field in both spatial and channel dimensions. In the formula, F is the input feature field with shape C×H×W, AvgPool(F) is the average pooling across channels at each spatial location, resulting in a feature field of shape 1×H×W. MaxPool(F) is the max pooling across channels at each spatial location, resulting in a feature field of shape 1×H×W. For the splicing operation, two 1×H×W feature fields are spliced ​​together in the channel dimension to obtain 2×H×W. For one The convolutional layer is used to fuse spatial information and output a single-channel feature field 1×H×W. The final spatial attention mask, with a shape of 1×H×W, is multiplied point-by-point with the input feature field to enhance important spatial regions.

[0095] The expression for the multi-head attention mechanism is as follows:

[0096]

[0097]

[0098] This mechanism begins with a linear projection and head-by-head processing of the input query (Q), key (K), and value (V). First, Q, K, and V are projected onto multiple subspaces using independent weight matrices, resulting in several sets of lower-dimensional submatrices, each corresponding to an attention head (see the above expression). Then, within each attention head, the dot product of query W and key K is calculated to obtain the original attention score. Then divide by the square root of the key vector dimension ( The scale is applied to mitigate the vanishing gradient problem. If masking is required, a mask can be applied to the scaled fraction matrix to obscure irrelevant positions. Next, the fraction matrix is ​​softmax normalized to obtain the attention weights at each position. Finally, the weight matrix is ​​compared with the corresponding value vector. The outputs of each attention head are weighted and summed to generate the output of each head (see the expression below). Finally, the outputs of all attention heads are concatenated along the feature dimension and fused through a linear projection to restore the original input dimension, forming the final result of the multi-head attention module.

[0099] The expression for the spatial pyramid attention mechanism is: This mechanism first performs multi-scale parallel processing on the input feature field F, using pooling windows of different sizes (such as global pooling, region pooling, etc.). Extract multi-scale spatial context information; then transform and upsample the features at each scale. To the original size, and then through splicing ( ) or by addition ( These multi-scale features; then, convolutional layers (typically 1×1 convolutions) and activation functions ( A comprehensive spatial attention weight field is generated; finally, this weight field is multiplied element-wise with the original feature field, enabling the model to adaptively focus on both local details and global structure at the same time, thereby enhancing its ability to represent multi-scale visual patterns.

[0100] Furthermore, the weighted loss function that makes the forecast model emphasize the impact of extreme wind events is constructed based on the mean absolute error (MAE), and at least simultaneously incorporates tiered observation weights, missed report penalty weights, and false alarm penalty weights:

[0101] The graded observation weights are assigned based on the interval in which the observed wind speed value is located. The higher the threshold of the interval to which the observed wind speed belongs, the greater the graded observation weight is assigned to the thunderstorm wind sample.

[0102] The missed detection penalty weight is a high-multiplier penalty weight applied when the forecast model misses a wind speed exceeding the preset strong wind threshold.

[0103] The false alarm penalty weight is a penalty weight that is lower than the missed alarm penalty weight when the forecast model makes a false alarm for wind speeds exceeding the preset strong wind threshold.

[0104] Furthermore, loss functions are used to calculate the gradient of model parameters and optimize the model in deep learning methods. Currently, the most commonly used loss functions are root mean square error (MSE) and mean absolute error (MAE). However, the MSE loss function generally makes the forecast tend to be smooth; MAE tends to focus on the overall forecast bias of the wind speed field, thereby reducing the attention to low-probability events such as thunderstorms and strong winds, resulting in a large number of missed reports of thunderstorms and strong winds. Therefore, the method of this invention adopts a weighted MAE loss function (Equation 7-10) based on wind force level, missed report penalty, and false report penalty, increasing the model's attention to thunderstorms and strong winds, and improving the model's forecasting ability by applying different penalty strengths to false reports and missed reports of thunderstorms and strong winds.

[0105] The expression for the weighted loss function in this invention, which makes the forecast model take into account the impact of extreme wind events, is as follows:

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] In the formula, and These represent the predicted maximum wind speed output by the forecast model and the corresponding observed maximum wind speed, respectively; N represents the total number of thunderstorm wind samples at all grid points across all forecast times. Based on maximum wind speed observation data Hierarchical observation weights; and These are the first and second maximum wind speed thresholds, respectively; generally, and Take the wind speeds corresponding to level 7 and level 8 wind forces, and assign weights of 1, 10, and 20 to the three intervals of wind speed < 13.9 m / s (wind force < level 7), 13.9 m / s ≤ wind speed < 17.2 m / s (level 7 < wind force < level 8), and wind speed ≥ 17.2 m / s (wind force ≥ level 8), respectively; The penalty weight for missed reporting of extreme strong winds is 20 times the weight when a missed report occurs; This is the penalty weight for false alarms of extreme gales. When a false alarm occurs, a penalty of 10 times is applied. This is mainly to prevent the model from significantly increasing the forecast range of thunderstorms and gales in order to reduce the penalty for false alarms, so as to improve the accuracy of the model's forecast of thunderstorms and gales. , , The maximum wind speeds are weighted for levels 1, 2, and 3, respectively. Weight for underreporting; The weights are for false alarms; the weight parameters in the weighted loss function are adaptively adjusted in multiple training rounds. Based on the forecast model's performance in forecasting thunderstorm and strong wind events in previous training rounds, the values ​​of the maximum wind speed weights (levels 1, 2, and 3), the missed alarm weights, and the false alarm weights in the weighted loss function are dynamically adjusted.

[0111] Furthermore, the adaptive adjustment process of the weight parameters described above will be illustrated with an example below. This example involves two rounds of training. In the first round, if the forecast model significantly misses large-scale thunderstorm and strong wind events, the loss function value will be large. In this case, the forecast model will automatically calculate the gradient during backpropagation and increase the gradient based on maximum wind speed observation data in the next round of training. In the graded observation weights, the weight value a3 corresponds to the high wind speed level (e.g., ≥17.2 m / s), and the missed report weight b is increased. By increasing the penalty for missed thunderstorm winds, the forecast model is forced to pay more attention to thunderstorm winds. In the second round of training, if the forecast model can predict thunderstorm winds, but the forecast range is far beyond the actual observation range (i.e., serious false alarms), the loss function value, although lower than in the first round, remains at a high level. At this time, the forecast model will maintain or further optimize the values ​​of a3 and b in the next round of training, and increase the false alarm penalty coefficient c to balance forecast accuracy and reduce false alarms.

[0112] Through the above multiple rounds of iterative training, the weight parameters (a1, a2, a3, b, c) in the loss function can be adaptively adjusted to their optimal values, enabling the forecast model to control the false alarm rate and improve the overall forecast accuracy while ensuring its sensitivity to thunderstorm and strong wind events.

[0113] Furthermore, specific methods for generating the final integrated forecast product of thunderstorms and gales by adopting an integration strategy aimed at reducing the risk of underreporting include:

[0114] For each grid point in the forecast space, the maximum wind speed value predicted by all forecast models at that grid point is taken as the final integrated forecast wind speed for that grid point.

[0115] Example 2:

[0116] The present invention provides a thunderstorm and gale nowcasting system based on UNET multimodal integration, comprising a dataset construction module, a local data augmentation module, a forecast model construction module, a training module, and an integration module;

[0117] The dataset construction module is used to extract specific samples of thunderstorm wind events based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data and terrain height data, and by considering the joint screening of peak wind speed and lightning occurrence.

[0118] The local data augmentation module is used to perform local data augmentation on the thunderstorm and strong wind samples in the dataset, taking into account the spatiotemporal evolution characteristics of the severe convective system.

[0119] The forecast model building module is used to build a variety of forecast models based on the UNET network and embedded with different attention mechanisms. The forecast models use the embedded attention mechanisms to adaptively focus on key spatiotemporal features related to the generation and dissipation of thunderstorms and strong winds.

[0120] The training module is used to train each forecast model using a loss function that integrates the penalty weighted wind speed for missed reports and the penalty for missed reports of thunderstorm wind events. The loss function applies a higher penalty weight to missed reports of thunderstorm wind events than to missed reports and assigns higher weights to thunderstorm wind samples with high wind speeds.

[0121] The integration module is used to combine the outputs of various forecast models to generate the final integrated forecast result for thunderstorms and gales, employing an integration strategy aimed at reducing the risk of missed reports.

[0122] Example 3:

[0123] An electronic device includes a memory and a processor. The memory stores a computer program, and the processor is used to invoke and run the computer program stored in the memory to perform the methods of any of the above embodiments.

[0124] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above embodiments.

[0125] Application Examples:

[0126] This example verifies the effectiveness of the method of the present invention. The forecast product of the present invention is named UNET-ENSEM, and two control experiments are set up. The first experiment (named UNET-1) uses the same forecasting scheme as UNET-ENSEM (i.e., an ensemble forecasting scheme based on the UNET multi-attention mechanism), but the training and validation sets used are different. The specific differences are detailed in Table 2. This control group is mainly used to compare the impact of different dataset construction methods on the final forecast results. The second experiment (named UNET-2) uses the same training and validation sets as UNET-ENSEM, but the forecasting scheme uses the classic UNET algorithm to construct the forecasting model. The specific algorithm differences are detailed in Table 3. This control group is set up to compare the impact of different forecasting schemes on the final results.

[0127] Table 2. Details of the differences between the datasets used in UNET-1 and UNET-ENSEM

[0128] Maximum wind speed observation data Ground hourly maximum wind speed observation The maximum value of the instantaneous ground wind speed observed minute by minute over the past hour is recorded. Thunderstorm and strong wind sample screening criteria At least two stations in the target area have recorded maximum wind speeds of ≥8 (maximum wind speed ≥17.2 m / s) in the past hour, and lightning activity has been observed within a 50 km radius of these stations. At least two stations within the target area have recorded maximum wind speeds of ≥17.2 m / s (maximum wind force ≥ level 8) in the past hour, and lightning activity or a combined reflectivity factor ≥45 dBZ has been observed within a 50 km radius of these stations. Number of training set samples 3546 9664 Number of validation set samples 887 2417

[0129] Table 3. Details of the differences between the forecasting schemes used in UNET-2 and UNET-ENSEM

[0130] Data Augmentation Not adopted Local data augmentation mechanism for thunderstorm and strong wind areas Attention mechanism Not adopted Four mechanisms are employed: attention gate, CBAM attention, multi-head attention, and spatial pyramid attention. loss function Mean Absolute Error (MAE) Weighted MAE based on wind force level, missed alarm penalty, and false alarm penalty Multi-model integration The UNET model was not used; it directly outputs forecast products. Based on multi-model forecast products, the maximum forecast wind speed is taken for each grid point to form the integrated forecast product.

[0131] Based on the past 1-hour maximum wind speed observation data of Anhui Province (including national stations and automatic stations, totaling 974 stations), this method uses indicators such as TS score, false alarm rate (FAR), and false negative rate (FNR) to evaluate the 1-hour maximum wind speed forecast output by each product (Formulas 11-13).

[0132]

[0133]

[0134]

[0135] NA, NB, and NC represent the number of stations that reported correctly, falsely, and missed, respectively. Considering that this method focuses on the forecasting capability of each product for thunderstorms with winds of force 8 or higher, the forecasting performance for maximum wind speeds ≥17.2 m / s was primarily tested. Table 4 shows the test results of each product for thunderstorms with winds of force 8 or higher in the next hour for the test set (i.e., 3305 samples from April to September 2025). It can be seen that UNET-ENSEM achieved a TS score of 0.202 and an FNR of 0.671 for thunderstorms with winds of force 8 or above in the next hour, which are 300% higher and 29% lower than UNET-1, respectively. This indicates that, under the same forecasting scheme, using the maximum instantaneous wind speed per minute over the past hour instead of the maximum wind speed per hour, and using a combined reflectivity factor ≥45dBZ instead of the lightning activity-related thunderstorm wind sample selection criterion, effectively expanded the number of training samples, which helps the model capture more thunderstorm wind-related feature information. Further comparison of the test results of UNET-ENSEM and UNET-2 revealed that UNET-2 has almost no forecasting ability for thunderstorms with winds of force 8 or above (TS score of only 0.003), indicating that, under the same dataset, the UNET regression model is unable to capture the formation and dissipation process of thunderstorm winds, while the ensemble forecast product based on the UNET multi-attention mechanism can significantly improve the forecasting ability for thunderstorm winds.

[0136] Table 4. Test results of various products for thunderstorms with winds of level 8 or above in the next hour, from April to September 2025.

[0137] UNET-1 0.05 0.515 0.948 UNET-2 0.003 0.157 0.997 UNET-ENSEM 0.202 0.657 0.671

[0138] On the evening of July 16, 2025, influenced by the northerly airflow behind the 500hPa trough and the southward movement of the 850hPa cold shear line, Anhui Province experienced a relatively strong thunderstorm and gale process from north to south. Starting from 19:00 on the 16th, using radar composite reflectivity factor mosaics taken every 6 minutes over the past hour, ground maximum wind speed station observations over the past hour, and topographic height data, the forecast effectiveness of the above products for level 8 thunderstorm and gale in the next hour was compared.

[0139] The results showed that at 19:00 on the 16th, there were two east-west squall lines at the borders of Anhui and Henan, and Anhui and Jiangsu. The squall line in Henan had a looser overall structure, while the one in Jiangsu was stronger, with the highest combined reflectivity factor exceeding 60 dBZ. Figure 5 (See Figure a). At 20:00, UNET products forecasted gale-force winds of level 8 or higher at the border of Anhui and Jiangsu provinces, with the affected area significantly larger than observed. Figure 6 (See Figure a). Subsequently, between 20:00 and 22:00, the squall line within Henan Province gradually weakened and dissipated as it moved southward, while the squall line within Jiangsu Province continued to develop and move southward. Figure 5(Figures b to d) The thunderstorm and strong wind area remained in the rear part of the squall line as it moved southward. The UNET-ENSEM forecast of the strong wind area was larger than expected, and its movement was slightly slower than observed. Figure 6 (Figures b to d). From 23:00 onwards, the western segment of the squall line continuously merged and intensified with newly formed cells from the Dabie Mountains, gradually moving southwestward, while the eastern segment of the squall line gradually weakened and dissipated. Figure 5 (Images from e and i) The gale area predicted by UNET-ENSEM is also gradually moving southwest. Figure 5 (See e-map and i-map). Overall, UNET-ENSEM has a better forecasting effect on the evolution of thunderstorms and strong winds, while UNET-1 and UNET-2 have almost no forecasting ability for thunderstorms and strong winds. Figures 7-8 ).

[0140] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for nowcasting thunderstorms and strong winds based on UNET multimodal ensemble, characterized in that, include Based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data, and terrain height data, and by considering the joint screening of peak wind speed and lightning occurrence, specific extraction of thunderstorm wind samples is carried out to construct a dataset for thunderstorm wind events. Local data augmentation was performed on the thunderstorm and strong wind samples in the dataset to target the spatiotemporal evolution characteristics of severe convective systems; Multiple forecasting models based on the UNET network and embedded with different attention mechanisms are constructed. The forecasting models adaptively focus on key spatiotemporal features related to the generation and dissipation of thunderstorms and strong winds using the embedded attention mechanisms. Each forecast model was trained using a loss function that combines the penalty weights of false alarms and missed reports of extreme gale events. The loss function applies a higher penalty weight to missed reports of thunderstorm gale events than to false alarms and assigns higher weights to thunderstorm gale samples with high wind speeds. An integrated strategy aimed at reducing the risk of missed forecasts was adopted, which combined the outputs of various forecast models to generate the final integrated forecast result for thunderstorms and gales; The expression for the weighted loss function that makes the forecast model take into account the impact of extreme wind events is: ； ； ； ； In the formula, and These represent the predicted maximum wind speed output by the forecast model and the corresponding observed maximum wind speed, respectively; N represents the total number of thunderstorm wind samples at all grid points across all forecast times. Based on maximum wind speed observation data Hierarchical observation weights; and These are the first and second maximum wind speed thresholds, respectively; The penalty weight for unreported extreme winds; The penalty weight for reporting extreme wind speeds; , , The maximum wind speeds are weighted for levels 1, 2, and 3, respectively. Weight for underreporting; The weights for false alarms are set as follows: The weight parameters in the weighted loss function are adaptively adjusted in multiple training rounds. Based on the forecasting performance of the forecasting model for thunderstorm and strong wind events in previous training rounds, the values ​​of the maximum wind speed weights of levels 1, 2, and 3, the false alarm weights, and the missing report weights in the weighted loss function are dynamically adjusted. The specific method for generating the final integrated forecast product of thunderstorms and gales by combining the outputs of various forecast models using an integrated strategy aimed at reducing the risk of missed reports includes: For each grid point in the forecast space, the maximum wind speed value predicted by all forecast models at that grid point is taken as the final integrated forecast wind speed for that grid point.

2. The thunderstorm and gale nowcasting method based on UNET multimodal ensemble as described in claim 1, characterized in that, The specific method for constructing a dataset for thunderstorm and gale events, based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data, and terrain data, and by considering the joint screening of peak wind speed and lightning occurrence, includes the following: Based on the observation time of the radar composite reflectivity mosaic data, identify potential high-wind stations in the ground minute instantaneous wind speed station observation data where the maximum ground wind speed observation data in the past hour is greater than or equal to the first threshold. Within a preset range centered on the potential high-wind station, if lightning is determined to have occurred based on lightning location data, or if the radar combined reflectivity factor in the radar combined reflectivity factor mosaic data is greater than or equal to the second threshold, then the radar combined reflectivity factor mosaic data, the ground minute-by-minute instantaneous wind speed station observation data, and the terrain height data corresponding to the potential high-wind station are used as thunderstorm high-wind samples and added to the dataset; the features in the thunderstorm high-wind samples are the radar combined reflectivity factor mosaic data minute-by-minute over the past hour, the ground maximum wind speed observation data over the past hour, and the terrain height data, all labeled as ground maximum wind speed observation data for the next hour.

3. The method for nowcasting thunderstorms and strong winds based on UNET multimodal ensemble as described in claim 2, characterized in that, The specific methods for performing local data augmentation on thunderstorm and strong wind samples in the dataset, focusing on the spatiotemporal evolution characteristics of severe convective systems, include: Perform geometric transformations on features and labels, including random translation, random rotation, and random mirroring. The radar composite reflectivity factor mosaic data, ground maximum wind speed observation data, and ground maximum wind speed observation data in the label are randomly scaled, the radar composite reflectivity factor mosaic data in the feature are graded intensity perturbation, and the feature is randomly noise-added.

4. The thunderstorm and gale nowcasting method based on UNET multimodal ensemble as described in claim 3, characterized in that, The attention mechanisms mentioned include at least the attention gate mechanism, the CBAM attention mechanism, the multi-head attention mechanism, and the spatial pyramid attention mechanism.

5. The method for nowcasting thunderstorms and strong winds based on UNET multimodal ensemble as described in claim 1, characterized in that, The weighted loss function that makes the forecast model pay attention to the impact of extreme wind events is constructed based on the mean absolute error (MAE) and incorporates at least three levels of observation weights, missed report penalty weights, and false alarm penalty weights simultaneously: The graded observation weights are assigned based on the interval in which the observed wind speed value is located. The higher the threshold of the interval to which the observed wind speed belongs, the greater the graded observation weight is assigned to the thunderstorm wind sample. The missed detection penalty weight is a high-multiplier penalty weight applied when the forecast model misses a wind speed exceeding the preset strong wind threshold. The false alarm penalty weight is a penalty weight that is lower than the missed alarm penalty weight when the forecast model makes a false alarm for wind speeds exceeding the preset strong wind threshold.

6. A thunderstorm and gale nowcasting system based on UNET multimodal integration, characterized in that, It includes a dataset construction module, a local data augmentation module, a prediction model construction module, a training module, and an ensemble module; The dataset construction module is used to extract specific samples of thunderstorm wind events based on radar combined reflectivity factor mosaic data, ground minute-by-minute instantaneous wind speed station observation data, lightning location data and terrain height data, and by considering the joint screening of wind speed peak and lightning occurrence. The local data augmentation module is used to perform local data augmentation on the thunderstorm and strong wind samples in the dataset, focusing on the spatiotemporal evolution characteristics of the severe convective system. The forecast model construction module is used to construct multiple forecast models based on the UNET network and embedded with different attention mechanisms. The forecast models adaptively focus on key spatiotemporal features related to the generation and dissipation of thunderstorms and strong winds using the embedded attention mechanisms. The training module is used to train each forecast model using a loss function that integrates the false alarms of extreme gale events and the weighted wind speed with penalties for missed reports. The loss function applies a higher penalty weight to missed reports of thunderstorm gale events than to false alarms and assigns higher weights to thunderstorm gale samples with high wind speeds. The integration module is used to adopt an integration strategy aimed at reducing the risk of missed reports, and integrate the outputs of various forecast models to generate the final integrated forecast result of thunderstorm and gale. The expression for the weighted loss function that makes the forecast model take into account the impact of extreme wind events is: ； ； ； ； In the formula, and These represent the predicted maximum wind speed output by the forecast model and the corresponding observed maximum wind speed, respectively; N represents the total number of thunderstorm wind samples at all grid points across all forecast times. Based on maximum wind speed observation data Hierarchical observation weights; and These are the first and second maximum wind speed thresholds, respectively; The penalty weight for unreported extreme winds; The penalty weight for reporting extreme wind speeds; , , The maximum wind speeds are weighted for levels 1, 2, and 3, respectively. Weight for underreporting; The weights for false alarms are set as follows: The weight parameters in the weighted loss function are adaptively adjusted in multiple training rounds. Based on the forecasting performance of the forecasting model for thunderstorm and strong wind events in previous training rounds, the values ​​of the maximum wind speed weights of levels 1, 2, and 3, the false alarm weights, and the missing report weights in the weighted loss function are dynamically adjusted. The specific method for generating the final integrated forecast product of thunderstorms and gales by combining the outputs of various forecast models using an integrated strategy aimed at reducing the risk of missed reports includes: For each grid point in the forecast space, the maximum wind speed value predicted by all forecast models at that grid point is taken as the final integrated forecast wind speed for that grid point.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor being configured to invoke and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5 above.

Images