Minute-level-oriented thunderstorm strong wind nowcasting method and system

Abstract

The invention relates to the technical field of meteorological prediction, and discloses a minute-level thunderstorm strong wind nowcasting method and system, and the method comprises the steps: obtaining multi-source meteorological observation data, and carrying out the preprocessing of the multi-source meteorological observation data, and obtaining the preprocessed data; performing time continuity judgment and fusing the processed data to obtain a training sample; processing the training samples based on a sliding window, dividing the training samples into a training set, a verification set and a test set, constructing a neural network model, and training the neural network model to obtain a forecasting model; real-time a frames of observation data are processed based on the forecasting model, a frames of wind speed forecasting images with the resolution ratio of 1 kilometer in the future are obtained, and the wind speed forecasting images comprise a wind speed continuous distribution diagram and wind area falling area forecasting results larger than or equal to 8.0 m / s and larger than or equal to 17.2 m / s. According to the invention, the comprehensive perception capability of strong convection characteristics is improved, and minute-level strong wind nowcasting is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a minute-level thunderstorm gale nowcasting method and system. Background Art

[0002] Thunderstorms and gales are sudden, highly destructive, and disastrous weather phenomena triggered by severe convective weather systems. They are often accompanied by heavy rainfall, lightning, and hail. They are characterized by short onset, small spatial scale, rapid movement, and strong localization. Especially in the small and medium-scale convective systems that are frequently active in my country during summer, extreme wind speeds can develop within minutes, leading to serious consequences such as damaged buildings, broken trees, and maritime transport accidents. These are a major cause of high-incidence meteorological disasters.

[0003] With the advancement of meteorological technology, multi-source observation methods such as radar observations, ground-based automatic weather stations, and lightning location systems have gradually improved, providing richer data support for thunderstorm and gale monitoring and warning. However, existing thunderstorm and gale nowcasting technologies still face numerous challenges. On the one hand, traditional extrapolation methods such as TITAN (Tracking Radar Echoes) and SCIT (Storm Identification and Tracking) rely primarily on radar echo morphology for storm tracking, but lack direct modeling of wind speed itself and its spatial evolution, making it unable to accurately reflect the formation and evolution of wind field structure. On the other hand, although current numerical models have improved in scale accuracy, they still suffer from insufficient timeliness and lagging update frequency in terms of minute-level timeliness, rapid rolling updates, and spatial positioning of extreme wind zones.

[0004] In addition, multi-source observation data are inconsistent in terms of spatial resolution, time step, and observation accuracy, making data fusion difficult. Traditional forecasting methods often cannot fully explore the correlation between factors such as radar, wind speed, and lightning, resulting in low accuracy in identifying areas with extremely high wind speeds. Especially when the high wind threshold is high (such as ≥17.2m / s), the missed reporting rate is significantly high.

[0005] Therefore, it is necessary to design a minute-level thunderstorm and gale nowcasting method and system to solve the problems existing in current technology. Summary of the Invention

[0006] In view of this, the present invention proposes a minute-level thunderstorm and gale nowcasting method and system, aiming to solve the current problems of lack of direct modeling of wind speed itself and its spatial evolution, inability to accurately reflect the formation and evolution of wind field structure, insufficient timeliness and delayed update frequency.

[0007] In one aspect, the present invention provides a minute-level thunderstorm and gale nowcasting method, comprising: Acquiring multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data, the multi-source meteorological observation data including radar echo data, maximum wind observation data from a ground automatic weather station, lightning location data, and terrain elevation data, the preprocessing including noise processing and consistency processing; Performing temporal continuity judgment on the pre-processed data and fusing the processed data to obtain training samples; Processing the training samples based on a sliding window, dividing the training samples into a training set, a validation set, and a test set in chronological order, constructing a neural network model with an encoder-decoder structure, and training the neural network model based on the training set, the validation set, and the test set to obtain a prediction model; Based on the forecast model, the real-time a-frame observation data is processed to obtain a wind speed forecast image with a resolution of 1 km for the future a-frame, wherein the wind speed forecast image includes a continuous wind speed distribution map and prediction results of wind zones ≥8.0 m / s and ≥17.2 m / s.

[0008] Furthermore, obtaining multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data includes: The radar echo data, the maximum wind observation data from the ground automatic weather station, and the lightning location data are combined with the terrain elevation data, synchronously registered in the time and space dimensions, and uniformly cropped to the same spatial scale and time resolution within the target area; the radar echo data has a time resolution of 6 minutes and a spatial resolution of 1 kilometer; When performing noise processing on the radar echo data, the process includes: setting the values of the radar echo data with echo intensity lower than 10 dBZ to zero, and removing image frames with clutter and anomalies to generate a radar echo image; When the maximum wind observation data of the ground automatic weather station is subjected to consistency processing, the method includes: gridding the maximum wind observation data of the discrete stations into grid data consistent with the radar echo data by using an inverse distance weighted interpolation method, and generating a maximum wind image; When the lightning location data is processed for consistency, it includes: accumulating in time periods of 10 minutes, and assigning a radius expansion value with the observation point as the center in space to generate a lightning density map.

[0009] Furthermore, when determining the temporal continuity of the pre-processed data and fusing the pre-processed data to obtain training samples, the method includes: Based on the check function, the spatial proportion of the extremely high wind image in the pre-processed data that is greater than the high wind threshold is checked, and the time point information is recorded to generate a pkl file; Record the data with time difference of adjacent data less than or equal to ten minutes in the table file; The training sample is formed by combining the time period in the table with the radar echo image and the lightning density map at the same time.

[0010] Furthermore, processing the training samples based on a sliding window and dividing the training samples into a training set, a validation set, and a test set in chronological order includes: All the training samples are divided into a training set, a validation set, and a test set in chronological order, with a ratio of 8:1:1.

[0011] Furthermore, when constructing a neural network model with an encoder-decoder structure, the encoder includes: a multi-scale patch embedding module for dividing the radar echo image, the maximum wind image, and the lightning density map in the training sample into image blocks and extracting multi-scale features through a convolutional neural network; A learnable position encoding module is used to add encoding vectors representing temporal and spatial position information to the embedded image blocks, enhancing the model's ability to recognize temporal changes and spatial distributions. The spatial-temporal attention module includes the window attention mechanism, the transfer window attention mechanism, and the time dimension attention mechanism, which are used to extract local image features, cross-region features, and time series dependencies respectively; Parallel spatiotemporal convolution modules are used to enhance the encoder's ability to model local structural information and improve model generalization.

[0012] Furthermore, the decoder is used to receive the spatiotemporal feature map output by the encoder, and includes: The upsampling reconstruction module is used to restore the low-resolution feature map to the original spatial resolution through deconvolution or nearest neighbor interpolation methods; Feature fusion module, used to splice and fuse encoder output features at multiple scales; The Patch de-embedding module is used to restore the fused features to the image structure and output the wind speed forecast image of the next a frames, with each frame interval of 10 minutes and a spatial resolution of 1 km.

[0013] Furthermore, when the neural network model is trained based on the training set, validation set and test set, the following steps are included: The weighted mean square error is used as the main loss function, where high loss weights are set for grid points with wind speed ≥ 17.2 m / s; A perceptual loss function is introduced to compare the feature distribution of the predicted image with the actual image to keep the spatial structure of the wind area clear; Add a lightning density consistency loss term to enhance the spatial overlap between the predicted wind area and the actual high-incidence lightning area; The Adam optimizer is used in combination with the cosine annealing algorithm to adjust the learning rate. The model is trained and verified by alternately using the training set and the verification set, and the model parameters with the highest key success index on the verification set are selected as the prediction model.

[0014] Furthermore, when the training samples are processed based on a sliding window and the training samples are divided into a training set, a validation set, and a test set in chronological order, the method further includes: The sliding window length is 36 frames, the input sequence consists of the first 18 frames of multi-source image data, and the prediction sequence is the last 18 frames of maximum wind images.

[0015] Furthermore, a wind speed forecast image with a resolution of 1 km in the future a frame is obtained, and the output of the wind speed forecast image includes: Continuous value wind speed forecast graph, used to display the wind speed value predicted for each grid point at each time in the future; The binary map of wind speed areas ≥8.0m / s is used to represent the distribution of strong wind areas; A binary map of the extreme wind drop zone of ≥17.2m / s is used to indicate the warning area of the extreme wind zone; Multi-channel image sequences are stored by frame number, supporting dynamic image display and comparison and overlay with radar real-time conditions, and are used for display on the Web-GIS platform and push to mobile warning interfaces.

[0016] Compared with the existing technology, the beneficial effects of the present invention are as follows: through the thunderstorm and gale nowcasting method with minute-level timeliness and identification of extreme wind areas as the core goals, it breaks through the limitations of traditional extrapolation methods and numerical forecasts in terms of timeliness and identification of strong wind areas. By utilizing multi-source observation data such as radar echoes, ground extreme winds, lightning locations and terrain elevation, after noise processing and consistency alignment, high-quality and unified standard training sample construction is achieved, and the comprehensive perception ability of strong convective characteristics is improved; a deep neural network model with an encoder-decoder structure is adopted, combined with multi-scale patch embedding and spatial-temporal attention mechanism, to capture the "birth-dynamic-change" evolution characteristics of the extreme wind area. During the training process, the high-temporal and spatial resolution samples generated based on the sliding window and the weighted loss strategy enable the model to have higher discrimination ability under the wind speed threshold of ≥17.2 m / s; through real-time rolling updates every 10 minutes, a 1 km resolution wind speed distribution map and strong wind area prediction results are output frame by frame for the next hour, realizing minute-level strong wind nowcasting.

[0017] On the other hand, the present application also provides a minute-level thunderstorm gale nowcasting system for applying the above-mentioned minute-level thunderstorm gale nowcasting method, comprising: an acquisition unit configured to acquire multi-source meteorological observation data and pre-process the multi-source meteorological observation data to obtain processed data, wherein the multi-source meteorological observation data includes radar echo data, maximum wind observation data from a ground automatic weather station, lightning location data, and terrain elevation data, and the pre-processing includes noise processing and consistency processing; a judgment unit configured to perform time continuity judgment on the processed data and fuse the processed data to obtain training samples; a processing unit configured to process the training samples based on a sliding window, divide the training samples into a training set, a validation set, and a test set in chronological order, construct a neural network model with an encoder-decoder structure, train the neural network model based on the training set, the validation set, and the test set, and obtain a prediction model; The forecasting unit is configured to process the real-time a-frame observation data based on the forecasting model to obtain a wind speed forecast image with a resolution of 1 km for the future a-frame, wherein the wind speed forecast image includes a continuous wind speed distribution map and prediction results of wind zones ≥8.0 m / s and ≥17.2 m / s.

[0018] It is understandable that the above-mentioned minute-level thunderstorm gale nowcasting method and system have the same beneficial effects and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings: Figure 1 A flowchart of a minute-level thunderstorm and gale nowcasting method provided by an embodiment of the present invention; Figure 2 A technical roadmap for the minute-level thunderstorm and gale nowcasting method provided by an embodiment of the present invention; Figure 3 A test score chart for the minute-level thunderstorm and gale nowcasting method provided by an embodiment of the present invention; Figure 4 A schematic diagram of prediction results of a minute-level thunderstorm gale nowcasting method provided by an embodiment of the present invention; Figure 5 This is a functional block diagram of a minute-level thunderstorm and gale nowcasting system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0021] In some embodiments of the present application, see Figure 1 As shown in FIG, a minute-level thunderstorm gale nowcasting method includes: S100: Acquire multi-source meteorological observation data and pre-process the multi-source meteorological observation data to obtain pre-processed data. The multi-source meteorological observation data includes radar echo data, maximum wind observation data from ground automatic weather stations, lightning location data, and terrain altitude data. The pre-processing includes noise processing and consistency processing.

[0022] S200: Performing temporal continuity judgment on the pre-processed data and fusing the processed data to obtain training samples.

[0023] S300: Processing the training samples based on the sliding window, dividing the training samples into a training set, a validation set, and a test set in chronological order, constructing a neural network model with an encoder-decoder structure, training the neural network model based on the training set, the validation set, and the test set to obtain a prediction model.

[0024] S400: Process the real-time a-frame observation data based on the forecast model to obtain a wind speed forecast image with a resolution of 1 km for the future a-frame. The wind speed forecast image includes a continuous wind speed distribution map and prediction results for wind zones ≥8.0 m / s and ≥17.2 m / s.

[0025] Specifically, "a frame" refers to the number of consecutive wind speed forecast frames in the future. The default value is a=6, which means that the wind speed image within the next hour is predicted (each frame is 10 minutes apart, for a total of 6 frames). However, in actual deployment, it can be flexibly set according to business needs, a∈[1,12].

[0026] Specifically, in step S100, radar echoes, maximum ground winds, lightning locations, and terrain elevation data are collected, subjected to noise removal and spatial-temporal consistency processing, and a high-quality dataset with a standardized structure and temporal continuity is constructed. In step S200, temporal continuity judgment logic is introduced, and data segments with persistent weather evolution characteristics are identified through a sliding time window. These data segments are then integrated to form multi-channel training samples for model training. In step S300, a sliding window technique is used to divide the training samples into training, validation, and test sets. A neural network model with an encoder-decoder structure is constructed. The encoder of the model integrates multi-scale patch embedding and a joint spatial-temporal attention mechanism to fully extract wind speed and convective evolution characteristics in severe convective systems. The decoder combines attention-guided temporal information to reconstruct future wind field images. Finally, in step S400, the model is deployed to the business system, which inputs real-time a-frame observation images and outputs a 10-minute, 1-kilometer resolution wind speed forecast sequence for the next a-frame (e.g., 60 minutes). This directly obtains continuous wind speed distribution and predictions of strong wind locations of 8.0 m / s and 17.2 m / s.

[0027] Specifically, see Figure 2 The figure shows the complete technical roadmap of the proposed model. The technical process includes the acquisition and standardized preprocessing of multi-source observation data, temporal continuity screening and sample fusion, sliding window sample generation, deep neural network model training based on an encoder-decoder structure, and the continuous forecasting of multi-frame wind speed images in the future. This clarifies the overall technical logic of the system for the task of minute-level thunderstorm and gale nowcasting.

[0028] It is understandable that this embodiment improves the ability to identify extreme wind drop areas in thunderstorms and strong winds and the responsiveness of minute-level forecasts. By introducing high-frequency observation data and spatial interpolation processing, the problem of differences in multi-source data at the spatiotemporal scale is resolved; the temporal continuity judgment and sliding window sample construction method enhance the model's perception of the evolution of strong winds with strong localization and high suddenness; the encoder-decoder based deep learning model accurately models the evolution trend of wind speed at different scales through a spatial-temporal attention mechanism, and combined with a wind speed intensity weighted loss strategy, improves the model's recognition accuracy in high-threshold wind zones (≥17.2m / s); this embodiment has the ability to warn of extreme wind drop areas with a minute-level update frequency, 1km accuracy, and a high hit rate, improving the accuracy of short-term forecast services and the foresight of disaster prevention and control.

[0029] In some embodiments of the present application, multi-source meteorological observation data is acquired and preprocessed to obtain the preprocessed data, including combining radar echo data, maximum wind observation data from ground-based automatic weather stations, and lightning location data with terrain elevation data, performing simultaneous registration in the spatiotemporal dimensions, and uniformly cropping the data to the same spatial scale and temporal resolution within the target area. The radar echo data has a temporal resolution of 6 minutes and a spatial resolution of 1 kilometer.

[0030] Specifically, when performing noise processing on radar echo data, it includes: setting the values of the radar echo data with echo intensity lower than 10dBZ to zero, and removing image frames with clutter and anomalies to generate a radar echo image.

[0031] When the maximum wind observation data of the ground automatic weather station is processed for consistency, it includes: gridding the maximum wind observation data of the discrete stations into grid data consistent with the radar echo data through the inverse distance weighted interpolation method, and generating a maximum wind image.

[0032] Specifically, consistency processing of lightning location data includes: accumulating data in 10-minute time periods, and assigning radius expansion values with the observation point as the center in space to generate a lightning density map.

[0033] It is understandable that multi-source data (such as radar, wind speed, lightning, and terrain) may have inconsistent temporal resolution, spatial scale, and coordinate systems due to their origins from different devices and platforms. Therefore, this embodiment employs spatiotemporal co-registration, aligning data from different sources at the same time point and cropping them to a fixed spatial range (e.g., the Shandong Province region). All data is standardized to a unified grid format with a 6-minute time step and a 1-kilometer spatial resolution, ensuring a strict spatial-temporal correspondence between each frame when used as model input. Radar echo images contain a large amount of non-meteorological clutter or weak reflection signals, which hinders the model from extracting meaningful convective information. To this end, 10 dBZ is set as the physical lower limit of reflectivity intensity, and values below 10 dBZ are set to zero to eliminate weak signal background. Simultaneously, image visualization techniques are used to manually or automatically identify and remove image frames containing interfering clutter or morphological anomalies, generating a radar echo image sequence suitable for model training. Maximum wind data collected by ground-based automatic stations are discrete point observations with uneven spatial distribution, making them difficult to directly use as rasterized input for deep neural networks. The inverse distance weighted interpolation method is introduced. Based on the distance between the observation point and the grid points, a weighted estimate of the wind speed at each grid point is made, resulting in a smooth and continuous distribution of the wind speed field across space. For areas with sparsely distributed ground-based automatic weather stations, a minimum threshold for the number of observation points (e.g., ≥3) is set during interpolation to control interpolation quality. Nearest neighbor interpolation can be used to supplement insufficient observation points in boundary regions to avoid missing extreme values. The interpolated wind speed image has the same temporal and spatial resolution as the radar image. Lightning observation data are sparse and unstable in location, making them unsuitable for direct modeling. To enhance their spatial representation, a combination of spatiotemporal accumulation and neighborhood expansion is employed. In the temporal dimension, all lightning location events within every 10 minutes are accumulated to form an activity map for that time period. In the spatial dimension, all grid points within a certain radius (e.g., 20 kilometers) centered on each lightning point are assigned a value of 1 to spatially expand the lightning-affected area. The values of overlapping points are summed to form a lightning density map, which is used to characterize the activity of severe convective regions.

[0034] It is understandable that by simultaneously registering and standardizing multi-source data, including radar, wind speed, lightning, and terrain, the structural uniformity, physical consistency, and resolution alignment of the input data are achieved, providing an unambiguous input foundation for subsequent model training. Furthermore, the use of techniques such as physical thresholding, image culling, spatial interpolation, and lightning expansion effectively removes invalid data, improving the spatial representation and physical representativeness of the data. The constructed radar echo images, maximum wind images, and lightning density maps are highly consistent in both temporal and spatial dimensions, improving the training efficiency and prediction accuracy of the deep learning model and providing more accurate data support and prior information for the formation of maximum wind drop zones in thunderstorms.

[0035] In some embodiments of the present application, when the temporal continuity of the preprocessed data is judged and the preprocessed data is fused to obtain training samples, it includes: checking the spatial proportion of the maximum wind image in the preprocessed data that is greater than the high wind threshold based on the check function, and recording the time point information to generate a pkl file.

[0036] Record the data with time difference of adjacent data less than or equal to ten minutes in the table file.

[0037] A training sample is formed by combining the time period in the table with the radar echo image and lightning density map at the same time.

[0038] Specifically, the key to training a neural network for extreme wind forecasting is ensuring that the input data contains "meaningful strong wind events." If the wind speed in an image is generally weak, with no areas of strong wind, that frame will have little impact on model training and may even interfere with learning. By presetting a wind speed threshold (such as 8.0 m / s or 17.2 m / s), a check function is used to iterate over each frame of the extreme wind image. The ratio of the number of grid points with wind speeds exceeding the threshold to the total number of grid points in the image, i.e., the spatial fraction, is calculated. If the spatial fraction exceeds the set threshold (such as 0.5% or 1%), the time point of that frame is recorded, indicating the presence of a possible thunderstorm or strong wind event at that moment. Finally, all time points that meet the criteria are saved as a .pkl file for subsequent sample construction. Deep time series models require high continuity in the input time series. Time jumps or missing frames in the training samples can seriously affect the model's learning of wind speed evolution trends. After reading all valid time points from the PKL file, the data is sorted based on timestamps. The interval between any two adjacent time points is determined to be less than or equal to 10 minutes (equal to 1 frame), thereby filtering out continuous time periods. Time periods that meet the continuity criteria are recorded in a table file (such as CSV or XLS) to clearly define the start and end times of each sample. Model training requires the alignment of multiple input channels (radar, wind speed, lightning, etc.) at the same moment. Based on each continuous time segment in the table, the following corresponding images are extracted at the same moment: radar echo image, interpolated maximum wind image, and generated lightning density map. These three image types are then spliced together to form the complete sample input data. Each training sample can contain several frames (e.g., 18 input frames + 18 prediction frames), which are generated using a sliding window approach and ultimately used for model training, validation, and testing.

[0039] It's understandable that the automated data screening and sample generation mechanism of "check function + time difference judgment + data grouping" solves the problems of "high proportion of weak signal samples, severe time frame interruption, and data channel asynchrony" that exist in traditional meteorological sample construction. This ensures that the samples used for training are all real, continuous, and representative fragments of high wind processes, improving the targetedness and generalization capabilities of model training. At the same time, through structured output (pkl file + table file), a controllable sample management system is established, facilitating subsequent data updates and model retraining. This improves the model's ability to fit severe convective wind processes, enhancing the model's accuracy in identifying sudden extreme wind areas and its timely response capabilities in practical applications.

[0040] In some embodiments of the present application, when processing training samples based on a sliding window and dividing the training samples into a training set, a validation set, and a test set in chronological order, it includes: dividing all training samples into a training set, a validation set, and a test set in chronological order in a ratio of 8:1:1.

[0041] Specifically, sliding windows are a common data augmentation and sequence segmentation method in time series modeling. Using a fixed number of frames (e.g., the first 18 frames) as the input window on the original timeline, new samples are generated by sliding backward a number of frames (e.g., 1 frame). This ensures that all continuous segments in the time series are used for training. For time periods that have passed temporal continuity checks, frames are sliced according to a set window length (e.g., 18 frames for input and 18 frames for prediction), and the data is traversed with a step size of 1 to construct a large number of continuous samples, improving the model's learning coverage. Temporal segmentation emphasizes the non-overlapping nature of the training, validation, and test datasets in the temporal dimension. This aligns with the principle of "training models with historical data and forecasting with future data" in real-world weather forecasting, and effectively prevents model overfitting caused by sample overlap. All generated sliding window samples are sorted in ascending order by their starting timestamps and divided into a training set (for model fitting), a validation set (for hyperparameter tuning), and a test set (for generalization performance evaluation) in an 8:1:1 ratio. The time periods of each dataset do not overlap, maintaining data independence.

[0042] It's understandable that increasing the sample size through a sliding window and introducing a time-sequential partitioning scheme ensures that the training, validation, and test sets each cover representative thunderstorm and high wind events without interfering with each other in time, effectively improving the stability and generalization of model training. The continuous input nature of the sliding window ensures that the model fully learns the dynamic patterns of wind speed evolution, while the time-sequential partitioning scheme avoids "information leakage," ensuring that the validation and test results more accurately reflect the model's forecasting performance in real-world applications. This approach balances data efficiency with the scientific nature of the evaluation.

[0043] In some embodiments of the present application, when constructing a neural network model with an encoder-decoder structure, the encoder includes: The multi-scale patch embedding module is used to divide the radar echo images, maximum wind images, and lightning density maps in the training samples into image blocks and extract multi-scale features through convolutional neural networks.

[0044] The learnable position encoding module is used to add encoding vectors representing temporal and spatial position information to the embedded image blocks, enhancing the model's ability to recognize temporal changes and spatial distribution.

[0045] The spatial-temporal attention module includes window attention mechanism, transfer window attention mechanism and time dimension attention mechanism, which are used to extract local image features, cross-region features and time series dependencies respectively.

[0046] A parallel spatiotemporal convolution module is used to enhance the encoder's ability to model local structural information to improve model generalization.

[0047] Specifically, the multi-scale patch embedding module divides the input image (including radar echo images, extreme wind images, and lightning density maps) into several fixed-size image patches (such as 4×4, 8×8, and 16×16 pixels). It encodes each patch using a lightweight convolutional neural network (CNN) to extract low-level and high-level multi-scale spatial features. Spatial location information is retained, allowing the model to perceive the geographic distribution of features during subsequent processing. The learnable position encoding module constructs a set of trainable position encoding vectors, corresponding to the time frame number and spatial location index of each patch in the image. This position encoding vector is added or concatenated with the feature vector of each patch, enabling the model to learn when and where the image occurred. Compared to traditional sinusoidal position encoding, the learnable position encoding is more adaptable and can be dynamically optimized in real-world data. The window attention mechanism in the spatial-temporal attention module demarcates small, non-overlapping regions within the image and performs self-attention calculations only within these local windows, reducing computational effort while focusing on local wind field patterns. The shift window mechanism staggers and shifts the window boundaries, enabling the model to capture long-range spatial dependencies across regions. The temporal attention mechanism uses the values of the same grid point at different time frames as a sequence input to establish dynamic dependencies on wind speed evolution over time. The outputs of each attention module are fused together to enhance the model's spatiotemporal perception and reasoning capabilities. Although the Transformer architecture has strong modeling capabilities, it lacks the ability to recognize local textures and detailed structures and is prone to overfitting. Convolution operations, with their excellent translation invariance and ability to capture local patterns, can compensate for the shortcomings of the attention mechanism. The parallel spatiotemporal convolution module adds a convolution-based spatiotemporal feature extraction pathway to the backbone network in parallel. Using either 3D convolution or a separable spatiotemporal convolutional architecture, it extracts local details of wind speed variations from both the spatial and temporal dimensions. This output is then fused with the attention module output to improve the model's robustness and generalization.

[0048] It is understandable that the neural network encoder, constructed by integrating multi-scale feature extraction, location-aware modeling, spatial-temporal decoupled attention mechanism, and parallel convolutional structure, can not only accurately capture the multi-dimensional dynamic relationship between radar reflectivity, extreme winds, and lightning activity in severe convective systems, but also effectively balance local accuracy with overall trend identification, especially with excellent modeling capabilities for the "shape, change, and movement" processes in extreme wind regions. Compared with traditional deep networks or simple Transformer models, this encoder improves the stability, accuracy, and generalization capabilities in actual thunderstorm and gale forecasting tasks while maintaining model depth and expressiveness, providing a model foundation for accurate warnings of extreme winds at the minute level.

[0049] In some embodiments of the present application, the decoder is used to receive the spatiotemporal feature map output by the encoder, and includes: The upsampling reconstruction module is used to restore the low-resolution feature map to the original spatial resolution through deconvolution or nearest neighbor interpolation methods.

[0050] The feature fusion module is used to concatenate and fuse the output features of encoders at multiple scales.

[0051] The Patch de-embedding module is used to restore the fused features to the image structure and output the wind speed forecast image of the next a frames, with each frame interval of 10 minutes and a spatial resolution of 1 km.

[0052] Specifically, patch embedding uses non-overlapping partitioning, breaking the original image into small patches of a fixed size (e.g., 8×8) and inputting them into the encoder. The patch de-embedding stage rearranges the patch embedding vectors according to their original spatial positions and restores image continuity and edge connectivity through convolution / deconvolution operations. To reduce computational complexity, the encoder downsamples the input image (e.g., patch embedding), resulting in a reduction in the spatial resolution of the feature map. The decoder must restore these low-resolution features to their original size to output a spatially complete wind speed image. The upsampling reconstruction module uses deconvolution to spatially upsample the feature map. It has learning capabilities and can dynamically adjust the restoration method based on context. Alternatively, it uses non-parametric methods such as nearest neighbor interpolation for fast upsampling, making it suitable for deployment scenarios requiring fast inference speed. The final output image maintains the same spatial size as the original input image (e.g., 1 km resolution). The encoder outputs spatiotemporal feature maps at multiple scales and levels, containing semantic information at different granularities. Fusion of this information improves the expressiveness and detail quality of the output image. The feature fusion module concatenates feature maps from different encoder branches or depths channel-wise, using convolutional fusion operations (such as 1×1 convolution) to reduce dimensionality and integrate information. Timeline consistency is maintained during the fusion process to ensure the continuity and stability of the predicted image sequence in the temporal dimension. During the model's forward pass, the input image undergoes patch segmentation and embedding. During the decoding phase, these embedded vectors must be restored to a complete two-dimensional image structure to output a wind speed field image with real physical spatial significance. The patch de-embedding module rearranges the fused high-dimensional vectors by spatial position and reconstructs them into a two-dimensional image. Each reconstructed image frame corresponds to the predicted wind speed distribution at a certain moment in the future. The model outputs a sequence of future a-frame images (e.g., six frames for the next 60 minutes, with a 10-minute interval between each frame), each with a spatial resolution of 1 km, which is used to construct minute-level nowcast products.

[0053] It is understandable that by introducing upsampling reconstruction, feature fusion, and patch de-embedding modules to construct an image generation path, it is possible to accurately restore abstract spatiotemporal features to true wind speed images, maintaining spatial continuity while enhancing the detail of extreme wind drop areas. The multi-scale feature fusion mechanism enhances the perception of wind field structures at different scales, while the patch de-embedding module ensures the consistency and visual compatibility of the output results with the original meteorological observation images.

[0054] In some embodiments of the present application, when training a neural network model based on a training set, a validation set, and a test set, the following steps are included: The weighted mean square error is used as the main loss function, where high loss weights are set for grid points with wind speed ≥ 17.2 m / s.

[0055] A perceptual loss function is introduced to compare the feature distribution of the predicted image with the actual image to keep the spatial structure of the wind area clear.

[0056] A lightning density consistency loss term is added to enhance the spatial overlap between the predicted wind area and the actual high lightning incidence area.

[0057] The Adam optimizer is used in combination with the cosine annealing algorithm for learning rate adjustment. The model is trained and verified by alternating between the training set and the validation set. The model parameters with the highest key success index on the validation set are selected as the prediction model.

[0058] Specifically, the standard mean squared error (SMSE) treats all grid points equally. However, in thunderstorm and high wind forecasts, areas with extremely high wind speeds (e.g., ≥17.2 m / s) are a key focus and typically account for a very small proportion. Without weighting, the model tends to optimize low-wind speed areas and neglect high-risk areas. Weights are assigned to the prediction error at each grid point: if the wind speed is ≥17.2 m / s, a larger weight (e.g., 5-10 times) is assigned; other areas are assigned standard weights. This allows the model to focus more on high-wind speed areas during training, thereby improving the ability and accuracy to identify strong wind areas. Conventional pixel-level loss ignores the overall perceptual consistency of image structure, which can easily lead to blurry and shapeless predictions. Perceptual loss compares the image representation in a high-dimensional feature space to preserve the image's texture, structure, and semantic boundaries. The predicted image and the ground truth image are fed into a pretrained image feature extraction network (e.g., VGG). Feature maps at the intermediate layers are extracted, and the L2 distance between these feature maps is calculated. This loss term acts as an additional constraint, guiding the model to maintain clear edges, morphology, and structure in the wind speed areas. Lightning activity is highly correlated with thunderstorm intensity, and strong wind areas are often accompanied by frequent lightning. If the model can predict high wind areas that are highly consistent with lightning-intensive areas, it means that it has modeled the convective structure well. The actual lightning density map is used as an auxiliary supervisory signal; the mask of the area ≥17.2m / s is extracted from the model's predicted wind speed map, and the spatial overlap rate (such as IoU or cross entropy) is calculated with the lightning density map; the consistency is added as a loss term to the total loss to guide the model to more accurately locate the center of strong convection. The lightning consistency loss uses the intersection over union (IoU) method: the area ≥17.2m / s in the predicted wind speed map and the non-zero area in the lightning density map are converted into binary masks, their IoU values are calculated, and 1-IoU is added as a loss term to the total loss function. The Adam optimizer, combined with a weight decay strategy, improves stability and suppresses overfitting. The cosine annealing algorithm is used to dynamically adjust the learning rate, initially decreasing rapidly and then converging slowly, helping the model escape the local optimum. Performance is evaluated on the validation set after each round of training. The Critical Success Index (CSI) is selected as the validation metric, with hitting the maximum wind drop zone as the primary criterion. The model weight with the highest CSI on the validation set is continuously saved and used as the forecast model for final deployment.

[0059] Specifically, see Figure 3 As shown in the table, the average performance score results on multiple test samples. By introducing multi-dimensional loss strategies such as weighted mean square error, perception loss, and lightning consistency loss, and combining them with indicators such as CSI for evaluation, it is verified that the model's recognition accuracy in wind areas ≥17.2m / s is significantly improved, the false negative rate is reduced, and the generalization ability is better than traditional methods. Figure 4The figure below demonstrates the model's forecast performance for a typical thunderstorm with high winds. The figure compares the spatial evolution of the model's forecast and the actual (ground truth) map from time T = 3 to T = 18 minutes. As can be seen, the model not only accurately locates the strong wind zone's occurrence area but also effectively captures its path and expansion trends over time, demonstrating the proposed model's short-term nowcasting capabilities at the minute-scale.

[0060] It's understandable that the training scheme enhances the model's forecasting capabilities through three aspects: loss function design, supervisory information enhancement, and optimization strategy selection. By introducing weighted mean square error (MSE), the model focuses on areas with extreme winds. The inclusion of perceptual loss ensures clear boundaries and a complete structure for the output wind zones. The perceptual loss uses a VGG-16 network (pre-trained on ImageNet), calculating the L2 distance between the activation outputs of the conv3_3 and conv4_3 layers. The VGG network weights remain frozen during training and are not included in gradient updates. The lightning consistency loss enhances the model's ability to spatially match deep features of severe convective systems. Combined with the Adam optimizer and cosine learning rate annealing mechanism, training stability and convergence efficiency are improved.

[0061] In some embodiments of the present application, when processing training samples based on a sliding window and dividing the training samples into a training set, a validation set, and a test set in chronological order, the process further includes: The sliding window length is 36 frames, the input sequence consists of the first 18 frames of multi-source image data, and the prediction sequence is the next 18 frames of maximum wind images.

[0062] Specifically, the interval between each frame of the sliding window is 10 minutes, so 6 frames represent the observation data of the previous 60 minutes as input; the sliding step is 1 frame, that is, 10 minutes, to ensure the temporal continuity and sufficiency of sample generation.

[0063] In some embodiments of the present application, a wind speed forecast image with a resolution of 1 km in the future a frame is obtained, and the output of the wind speed forecast image includes: The continuous value wind speed forecast map is used to display the wind speed values predicted for each grid point at each time in the future.

[0064] The binary map of wind speed areas ≥8.0m / s is used to represent the distribution of strong wind areas.

[0065] The binary map of the maximum wind drop area ≥17.2m / s is used to indicate the warning area of the extreme wind zone.

[0066] Multi-channel image sequences are stored by frame number, supporting dynamic image display and comparison and overlay with radar real-time conditions, and are used for display on the Web-GIS platform and push to mobile warning interfaces.

[0067] Specifically, meteorological events evolve over time, and neural network models need to capture the "rise, rise, and fall" of high winds through sequence modeling. A sliding window approach maximizes the use of continuous observation sequences and improves data utilization. The sliding window length is set to 36 frames, with the first 18 frames serving as input, representing the weather evolution over the past 180 minutes; the last 18 frames, serving as prediction targets, represent the maximum wind trend over the next 180 minutes. The window is slid every 10 minutes (i.e., a sample step of 1 frame), generating a large number of supervised learning samples from all data segments that meet temporal continuity requirements, increasing both sample quantity and spatiotemporal diversity. Each frame in the input sample consists of multi-channel images, including radar echoes, maximum wind maps, and lightning density maps. The predicted label is a sequence of 18 consecutive maximum wind images. The model output must meet the business needs of users at multiple levels: it must include both continuous wind speed values to support quantitative analysis and binary maps of the impact zone to support warning signal generation and layered presentation. The output includes wind speed forecasts for the next a frames (e.g., 18 frames), with a time interval of 10 minutes and a spatial resolution of 1 km. Each frame contains the following three forms: Continuous wind speed forecast map: provides quantitative wind speed forecast values for each grid point at future moments for quantitative analysis and structural assessment; Wind speed zone map ≥8.0m / s (binarized): used to delineate general strong wind zones and support low-level wind warning identification; Map of the area with the maximum wind speed of ≥17.2m / s (binarized): directly serves the issuance of high-level disaster warnings and marks extreme wind risk areas; The output is stored as a multi-channel image sequence with a numbered structure (e.g., T+10, T+20, etc.). Each frame has a unified data structure, facilitating post-processing. It can be directly connected to a Web-GIS platform for layer overlay, dynamic animation playback, and comparison analysis with live radar data. It also supports interface access to mobile devices or emergency systems for early warning information push. Model output images are stored in multi-channel TIFF or NetCDF format, with frame naming using formats such as "YYYYMMDD_HHMM_T+10." Dynamic image sequences generate MP4 or GIF animations by merging a-frame images. A JSON interface is provided for layer overlay and playback control on the web platform.

[0068] It is understandable that by setting a sliding window mechanism, the model's time-series modeling capability for strong wind processes is enhanced, data utilization efficiency is improved, and the model is easier to capture wind speed evolution trends; at the same time, through a clearly structured and content-rich output format, the model results are divided into two levels: continuous value prediction map + binary risk area map, which can be used for quantitative judgment by meteorological analysts and adapt to the rapid identification needs of automated alarm systems, enhancing the visualization, usability and business integration capabilities of the model results; combined with dynamic graphs and numbering sequence mechanisms, it facilitates the deployability of minute-level rolling forecast products on actual combat platforms.

[0069] The above-mentioned embodiment uses a thunderstorm and gale nowcasting method with minute-level timeliness and identification of extreme wind zones as its core goals, breaking through the limitations of traditional extrapolation methods and numerical forecasts in terms of timeliness and identification of strong wind zones. By utilizing multi-source observation data such as radar echoes, ground extreme winds, lightning locations, and terrain elevation, after noise processing and consistency registration, high-quality, unified standard training samples are constructed, improving the comprehensive perception of strong convective characteristics. A deep neural network model with an encoder-decoder structure is adopted, combined with multi-scale patch embedding and spatial-temporal attention mechanism to capture the "birth-dynamic-change" evolution characteristics of extreme wind zones. During the training process, high-temporal and spatial resolution samples generated based on sliding windows and a weighted loss strategy further enable the model to have higher discrimination capabilities under wind speed thresholds of ≥17.2m / s. Through real-time rolling updates every 10 minutes, a 1km resolution wind speed distribution map and strong wind zone forecast results are output frame by frame for the next hour, achieving minute-level strong wind nowcasting.

[0070] In another preferred embodiment based on the above embodiment, refer to Figure 5 As shown, this embodiment provides a minute-level thunderstorm gale nowcasting system, which is used to apply the above-mentioned minute-level thunderstorm gale nowcasting method, including: The acquisition unit is configured to acquire multi-source meteorological observation data and pre-process the multi-source meteorological observation data to obtain processed data. The multi-source meteorological observation data includes radar echo data, extreme wind observation data from ground automatic weather stations, lightning location data, and terrain altitude data. The pre-processing includes noise processing and consistency processing.

[0071] The judgment unit is configured to judge the temporal continuity of the processed data and fuse the processed data to obtain training samples.

[0072] The processing unit is configured to process the training samples based on the sliding window, divide the training samples into a training set, a validation set and a test set in chronological order, construct a neural network model with an encoder-decoder structure, train the neural network model based on the training set, the validation set and the test set, and obtain a prediction model.

[0073] The forecast unit is configured to process the real-time A-frame observation data based on the forecast model to obtain a wind speed forecast image with a resolution of 1 km for the future A-frame. The wind speed forecast image includes a continuous wind speed distribution map and prediction results for wind zones ≥8.0m / s and ≥17.2m / s.

[0074] It's no surprise that this thunderstorm and gale nowcasting method, with its core objectives of minute-level timeliness and identification of extreme wind zones, overcomes the limitations of traditional extrapolation and numerical forecasting methods in terms of timeliness and strong wind zone identification. By utilizing multi-source observational data, including radar echoes, ground-level extreme winds, lightning locations, and terrain elevation, and through noise removal and consistent registration, high-quality, standardized training samples are constructed, enhancing the comprehensive perception of severe convective features. A deep neural network model with an encoder-decoder architecture, combined with multi-scale patch embedding and a spatial-temporal attention mechanism, captures the "birth-movement-change" evolution of extreme wind zones. The high-temporal-resolution samples generated by a sliding window and a weighted loss strategy during training further enhance the model's discriminative power at wind speed thresholds ≥17.2 m / s. Through a real-time rolling update every 10 minutes, the model outputs a 1 km resolution wind speed distribution map and strong wind zone forecast for the next hour, achieving minute-level strong wind nowcasting.

[0075] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0077] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0079] 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 it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for minute-level thunderstorm and gale nowcasting, characterized in that: include: Acquiring multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data, the multi-source meteorological observation data including radar echo data, maximum wind observation data from a ground automatic weather station, lightning location data, and terrain elevation data, the preprocessing including noise processing and consistency processing; Performing temporal continuity judgment on the pre-processed data and fusing the processed data to obtain training samples; Processing the training samples based on a sliding window, dividing the training samples into a training set, a validation set, and a test set in chronological order, constructing a neural network model with an encoder-decoder structure, and training the neural network model based on the training set, the validation set, and the test set to obtain a prediction model; Based on the forecast model, the real-time a-frame observation data is processed to obtain a wind speed forecast image with a resolution of 1 km for the future a-frame, wherein the wind speed forecast image includes a continuous wind speed distribution map and prediction results of wind zones ≥8.0 m / s and ≥17.2 m / s.

2. The minute-level thunderstorm gale nowcasting method according to claim 1, characterized in that: Acquiring multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data includes: The radar echo data, the maximum wind observation data from the ground automatic weather station, and the lightning location data are combined with the terrain elevation data, synchronously registered in the time and space dimensions, and uniformly cropped to the same spatial scale and time resolution within the target area; the radar echo data has a time resolution of 6 minutes and a spatial resolution of 1 kilometer; When performing noise processing on the radar echo data, the process includes: setting the values of the radar echo data with echo intensity lower than 10 dBZ to zero, and removing image frames with clutter and anomalies to generate a radar echo image; When the maximum wind observation data of the ground automatic weather station is subjected to consistency processing, the method includes: gridding the maximum wind observation data of the discrete stations into grid data consistent with the radar echo data by using an inverse distance weighted interpolation method, and generating a maximum wind image; When the lightning location data is processed for consistency, it includes: accumulating in time periods of 10 minutes, and assigning a radius expansion value with the observation point as the center in space to generate a lightning density map.

3. The minute-level thunderstorm gale nowcasting method according to claim 1, characterized in that: When determining the temporal continuity of the pre-processed data and fusing the pre-processed data to obtain training samples, the method includes: Based on the check function, the spatial proportion of the extremely high wind image in the pre-processed data that is greater than the high wind threshold is checked, and the time point information is recorded to generate a pkl file; Record the data with time difference of adjacent data less than or equal to ten minutes in the table file; The training sample is formed by combining the time period in the table with the radar echo image and the lightning density map at the same time.

4. The minute-level thunderstorm gale nowcasting method according to claim 1, characterized in that: Processing the training samples based on a sliding window and dividing the training samples into a training set, a validation set, and a test set in chronological order includes: All the training samples are divided into a training set, a validation set, and a test set in chronological order, with a ratio of 8:1:

1.

5. The minute-level thunderstorm gale nowcasting method according to claim 1, characterized in that: When constructing a neural network model with an encoder-decoder structure, the encoder includes: a multi-scale patch embedding module for dividing the radar echo image, the maximum wind image, and the lightning density map in the training sample into image blocks and extracting multi-scale features through a convolutional neural network; A learnable position encoding module is used to add encoding vectors representing temporal and spatial position information to the embedded image blocks, enhancing the model's ability to recognize temporal changes and spatial distributions. The spatial-temporal attention module includes the window attention mechanism, the transfer window attention mechanism, and the time dimension attention mechanism, which are used to extract local image features, cross-region features, and time series dependencies respectively; A parallel spatiotemporal convolution module is used to enhance the encoder's ability to model local structural information to improve model generalization.

6. The minute-level thunderstorm gale nowcasting method according to claim 5, characterized in that: The decoder is used to receive the spatiotemporal feature map output by the encoder, and includes: The upsampling reconstruction module is used to restore the low-resolution feature map to the original spatial resolution through deconvolution or nearest neighbor interpolation methods; Feature fusion module, used to splice and fuse encoder output features at multiple scales; The Patch de-embedding module is used to restore the fused features to the image structure and output the wind speed forecast image of the next a frames, with each frame interval of 10 minutes and a spatial resolution of 1 km.

7. The minute-level thunderstorm gale nowcasting method according to claim 6, characterized in that: When training the neural network model based on the training set, validation set and test set, the following steps are included: The weighted mean square error is used as the main loss function, where high loss weights are set for grid points with wind speed ≥ 17.2 m / s; A perceptual loss function is introduced to compare the feature distribution of the predicted image with the actual image to keep the spatial structure of the wind area clear; Add a lightning density consistency loss term to enhance the spatial overlap between the predicted wind area and the actual high-incidence lightning area; The Adam optimizer is used in combination with the cosine annealing algorithm to adjust the learning rate. The model is trained and verified by alternately using the training set and the verification set, and the model parameters with the highest key success index on the verification set are selected as the prediction model.

8. The minute-level thunderstorm gale nowcasting method according to claim 7, characterized in that: When the training samples are processed based on a sliding window and the training samples are divided into a training set, a validation set, and a test set in chronological order, the method further includes: The sliding window length is 36 frames, the input sequence consists of the first 18 frames of multi-source image data, and the prediction sequence is the last 18 frames of maximum wind images.

9. The minute-level thunderstorm gale nowcasting method according to claim 1, characterized in that: Obtain a wind speed forecast image with a resolution of 1 km for the next a frame. The output of the wind speed forecast image includes: Continuous value wind speed forecast graph, used to display the wind speed value predicted for each grid point at each time in the future; The binary map of wind speed areas ≥8.0m / s is used to represent the distribution of strong wind areas; A binary map of the extreme wind drop zone of ≥17.2m / s is used to indicate the warning area of the extreme wind zone; Multi-channel image sequences are stored by frame number, supporting dynamic image display and comparison and overlay with radar real-time conditions, and are used for display on the Web-GIS platform and push to mobile warning interfaces.

10. A minute-level thunderstorm gale nowcasting system, for applying the minute-level thunderstorm gale nowcasting method according to any one of claims 1 to 9, characterized in that: include: an acquisition unit configured to acquire multi-source meteorological observation data and pre-process the multi-source meteorological observation data to obtain processed data, wherein the multi-source meteorological observation data includes radar echo data, maximum wind observation data from a ground automatic weather station, lightning location data, and terrain elevation data, and the pre-processing includes noise processing and consistency processing; a judgment unit configured to perform time continuity judgment on the processed data and fuse the processed data to obtain training samples; a processing unit configured to process the training samples based on a sliding window, divide the training samples into a training set, a validation set, and a test set in chronological order, construct a neural network model with an encoder-decoder structure, train the neural network model based on the training set, the validation set, and the test set, and obtain a prediction model; The forecasting unit is configured to process the real-time a-frame observation data based on the forecasting model to obtain a wind speed forecast image with a resolution of 1 km for the future a-frame, wherein the wind speed forecast image includes a continuous wind speed distribution map and prediction results of wind zones ≥8.0 m / s and ≥17.2 m / s.

Images