Minute-level thunderstorm-gale approaching prediction method and system

Abstract

The application relates to the technical field of weather prediction, and discloses a thunderstorm gale near-forecast method and system for minutes, which comprises the following steps: acquiring multi-source meteorological observation data, pre-processing the multi-source meteorological observation data to obtain pre-processed data, judging time continuity, fusing the pre-processed data to obtain training samples, 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, training the neural network model to obtain a prediction model, processing real-time a-frame observation data based on the prediction model, and obtaining a future a-frame 1-kilometer resolution wind speed prediction image, wherein the wind speed prediction image comprises a wind speed continuous distribution graph and a prediction result of a wind area drop zone of >=8.0 m / s and >=17.2 m / s. The application improves the comprehensive perception ability of strong convective features and realizes minute-level strong wind near-forecasting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather forecasting, in particular to a minute-level thunderstorm gale near-forecast method and system. BACKGROUND

[0002] Thunderstorm gale is a sudden and extremely destructive weather phenomenon caused by strong convective weather systems, often accompanied by heavy rain, lightning, hail, etc. It has the characteristics of short occurrence time, small spatial scale, fast movement speed and strong locality. Especially in the small and medium-sized convective systems that are active frequently in summer in China, the maximum wind speed can form rapidly within a few minutes, leading to serious consequences such as building damage, tree breakage, and maritime transportation accidents, which is one of the important types of meteorological disasters.

[0003] With the development of meteorological technology, multi-source observation means such as radar observation, ground automatic weather stations, lightning positioning systems, etc. have gradually improved, providing more abundant data support for the monitoring and early warning of thunderstorm gale. However, the existing thunderstorm gale near-forecast technology still faces many challenges. On the one hand, traditional extrapolation methods such as TITAN (Tracking Radar Echoes) or SCIT (Storm Identification and Tracking) mainly rely on radar echo patterns for storm tracking, but lack direct modeling of wind speed and its spatial evolution, which cannot 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 have problems of insufficient timeliness and lagging update frequency in terms of minute-level timeliness, rapid rolling update and spatial positioning of extreme wind fall area.

[0004] In addition, multi-source observation data has inconsistencies in spatial resolution, time step, observation accuracy, etc., making data fusion difficult. Traditional forecasting methods often cannot fully exploit the correlation between radar, wind speed, lightning and other factors, resulting in low accuracy in identifying the area of extreme wind speed, especially when the high wind threshold is high (such as ≥17.2m / s), the false negative rate is significantly high.

[0005] Therefore, it is necessary to design a minute-level thunderstorm gale near-forecast method and system to solve the problems existing in the current technology. SUMMARY

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

[0007] In one aspect, the present application proposes a minute-level thunderstorm gale near-forecast method, comprising:

[0008] Obtaining 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 of ground automatic weather station, lightning location data and terrain elevation data, the preprocessing including noise processing and consistency processing;

[0009] Performing time continuity judgment on the preprocessed data and fusing the preprocessed data to obtain training samples;

[0010] 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, training the neural network model based on the training set, the validation set and the test set, and obtaining a prediction model;

[0011] Processing real-time a-frame observation data based on the prediction model to obtain a future a-frame 1-kilometer resolution wind speed prediction image, the wind speed prediction image including a wind speed continuous distribution map and ≥8.0m / s and ≥17.2m / s wind area drop zone prediction results.

[0012] Further, when obtaining multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data, it includes:

[0013] Combining the radar echo data, the maximum wind observation data of the ground automatic weather station and the lightning location data with the terrain elevation data, synchronously registering and uniformly cropping in the time and space dimensions to the same spatial scale and time resolution within the target area; and the radar echo data has a time resolution of 6 minutes and a spatial resolution of 1 kilometer;

[0014] When performing noise processing on the radar echo data, it includes: setting the values of the radar echo data with echo intensity lower than 10dBZ to zero, and eliminating image frames with clutter and abnormalities to generate radar echo images;

[0015] When performing consistency processing on the maximum wind observation data of the ground automatic weather station, it includes: gridding the maximum wind observation data of discrete sites into grid data consistent with the radar echo data by inverse distance weighted interpolation method to generate a maximum wind image;

[0016] When performing consistency processing on the lightning location data, it includes: accumulating in a 10-minute time period in time and expanding the radius to assign values with the observation point as the center in space to generate a lightning density map.

[0017] Further, when performing time continuity judgment on the preprocessed data and fusing the preprocessed data to obtain training samples, it includes:

[0018] checking the proportion of the maximum wind image greater than the strong wind threshold in the pre-processed data based on the check function, and recording the time point information to generate a pkl file;

[0019] Recording data with a time difference of less than or equal to ten minutes between adjacent data in a table file;

[0020] According to the time period existing in the table and the radar echo image and lightning density map at the same time, a training sample is formed.

[0021] Further, based on the sliding window, the training sample is processed, and when the training sample is divided into a training set, a validation set and a test set in time sequence, it includes:

[0022] All the training samples are divided into a training set, a validation set and a test set in time sequence, and the proportion is 8:1:1.

[0023] Further, when constructing a neural network model with an encoder-decoder structure, the encoder includes:

[0024] A multi-scale patch embedding module is used to divide the radar echo image, maximum wind image and lightning density image in the training sample into image blocks, and extract multi-scale features through a convolutional neural network;

[0025] A learnable position encoding module is used to add an encoding vector representing time sequence and spatial position information to the embedded image block, to enhance the model's recognition ability of time variation and spatial distribution;

[0026] A space-time attention module includes a window attention mechanism, a transfer window attention mechanism and a time dimension attention mechanism, which are respectively used to extract local features, cross-region features and time sequence dependence relationships;

[0027] A parallel space-time convolution module is used to enhance the modeling ability of the encoder for local structure information and improve the generalization of the model.

[0028] Further, the decoder is used to receive the space-time feature map output by the encoder, and includes:

[0029] An up-sampling reconstruction module is used to restore the low-resolution feature map to the original spatial resolution through deconvolution or nearest neighbor interpolation method;

[0030] A feature fusion module is used to splice and fuse the encoder output features of multiple scales;

[0031] A patch de-embedding module is used to restore the fused features to image structure, and output a future a-frame wind speed prediction image, with a frame interval of 10 minutes and a spatial resolution of 1 kilometer.

[0032] Further, when training the neural network model based on the training set, the validation set and the test set, the following steps are included:

[0033] Using weighted mean square error as the main loss function, wherein a high loss weight is set for grid points with wind speed ≥ 17.2 m / s;

[0034] Introducing a perception loss function for comparing the feature distribution of the predicted image and the live image, to maintain the spatial structure of the wind area clear;

[0035] Adding a lightning density consistency loss term to strengthen the spatial overlap between the predicted wind area and the actual lightning-prone area;

[0036] Using the Adam optimizer and combining the cosine annealing algorithm to adjust the learning rate, training and validating the model by alternately using the training set and the validation set, and selecting the model parameters with the highest key success index on the validation set as the prediction model.

[0037] Further, based on the sliding window, the training samples are processed, and when dividing the training samples into a training set, a validation set and a test set in chronological order, the following steps are further included:

[0038] The sliding window length is 36 frames, the input sequence is composed of the first 18 frames of multi-source image data, and the prediction sequence is the last 18 frames of maximum wind image.

[0039] Further, a future a-frame 1-kilometer resolution wind speed prediction image is obtained, and the output of the wind speed prediction image includes:

[0040] A continuous value wind speed prediction image for displaying the predicted wind speed value of each grid point at each time;

[0041] A ≥8.0m / s wind speed falling area binary image for indicating the distribution of strong wind falling area;

[0042] A ≥17.2m / s maximum wind falling area binary image for indicating the alarm area of extreme wind area;

[0043] The multi-channel image sequence is stored by frame number, supports dynamic image display and radar live comparison superposition, and is used for Web-GIS platform display and mobile terminal early warning interface push.

[0044] Compared with the prior art, the present application has the beneficial effects that: through the thunderstorm gale approaching prediction method taking the minute-level time effectiveness and the extremely large wind drop area identification as the core target, the limitations of the traditional extrapolation method and numerical prediction in time effectiveness and strong wind drop area identification are broken through. Using radar echo, ground maximum wind, lightning positioning and terrain elevation and other multi-source observation data, after noise processing and consistency registration, high-quality and unified standard training sample construction is realized, and the comprehensive perception ability of strong convective features is improved; the deep neural network model with the structure of encoder-decoder is adopted, combined with multi-scale Patch embedding and space-time attention mechanism, the "birth-movement-change" evolution characteristics of the maximum wind area are captured, and the high spatial-temporal resolution samples generated based on the sliding window and the weighted loss strategy in the training process make the model have higher discrimination ability under the wind speed threshold of 17.2 m / s; through the real-time rolling update every 10 minutes, the future 1-hour frame-by-frame 1-kilometer resolution wind speed distribution map and the strong wind drop area prediction result are output, and the minute-level strong wind approaching prediction is realized.

[0045] On the other hand, the present application also provides a minute-level thunderstorm gale approaching prediction system for applying the above-mentioned minute-level thunderstorm gale approaching prediction method, comprising:

[0046] The acquisition unit is configured to obtain multi-source meteorological observation data and pre-process the multi-source meteorological observation data to obtain processed data, the multi-source meteorological observation data including radar echo data, ground automatic weather station maximum wind observation data, lightning positioning data and terrain elevation data, and the pre-processing including noise processing and consistency processing;

[0047] The judgment unit is configured to perform time continuity judgment on the processed data and fuse the processed data to obtain training samples;

[0048] The processing unit is 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 the structure of encoder-decoder, train the neural network model based on the training set, the validation set and the test set, and obtain a prediction model;

[0049] The prediction unit is configured to process real-time a-frame observation data based on the prediction model to obtain a future a-frame 1-kilometer resolution wind speed prediction image, and the wind speed prediction image includes a wind speed continuous distribution map and ≥8.0 m / s and ≥17.2 m / s wind area drop area prediction results.

[0050] It can be understood that the above-mentioned minute-level thunderstorm gale approaching prediction method and system have the same beneficial effects, which will not be described here. BRIEF DESCRIPTION OF DRAWINGS

[0051] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to limit the present application thereto. Like reference numerals are used to refer to like elements throughout. In the drawings:

[0052] Figure 1 A flow chart of a thunderstorm gale approaching prediction method for minute level provided for an embodiment of the present application;

[0053] Figure 2 A technical roadmap of a thunderstorm gale approaching prediction method for minute level provided for an embodiment of the present application;

[0054] Figure 3 A test score chart of a thunderstorm gale approaching prediction method for minute level provided for an embodiment of the present application;

[0055] Figure 4 A prediction result schematic diagram of a thunderstorm gale approaching prediction method for minute level provided for an embodiment of the present application;

[0056] Figure 5 A functional block diagram of a thunderstorm gale approaching prediction system for minute level provided for an embodiment of the present application. DETAILED DESCRIPTION

[0057] Exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and the scope of the present disclosure can be accurately conveyed to those skilled in the art. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0058] In some embodiments of the present application, referring to Figure 1 A thunderstorm gale approaching prediction method for minute level, comprising:

[0059] S100: Obtain 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 including radar echo data, ground automatic weather station maximum wind observation data, lightning location data and terrain elevation data, the pre-processing including noise processing and consistency processing.

[0060] S200: Perform time continuity judgment on the pre-processed data and fuse the processed data to obtain training samples.

[0061] S300: Based on the sliding window, the training samples are processed, the training samples are divided into training set, validation set and test set in time sequence, a neural network model with encoder-decoder structure is constructed, the neural network model is trained based on the training set, the validation set and the test set, and a prediction model is obtained. S400: Based on the prediction model, real-time a-frame observation data is processed, and a future a-frame 1-kilometer resolution wind speed prediction image is obtained, the wind speed prediction image includes a wind speed continuous distribution map and a ≥8.0m / s and ≥17.2m / s wind area fall area prediction result.

[0067] Specifically, "a-frame" refers to a future continuous number of wind speed prediction frames, and a=6 by default, that is, a wind speed image within 1 hour in the future (with a frame interval of 10 minutes, a total of 6 frames) is predicted; but in actual deployment, it can be flexibly set according to business needs, a∈[1,12]. Specifically, the data of radar echo, ground maximum wind, lightning positioning and terrain elevation are collected through the S100 step, and noise elimination and spatial and temporal consistency processing are performed, so as to construct a high-quality data set with standard structure and continuous time; In S200, time continuity judgment logic is introduced, a data segment with persistent weather evolution characteristics is identified through a sliding time window, and a multi-channel training sample that can be used for model training is formed by fusion; In S300, the training sample is divided into training set, validation set and test set by using sliding window technology, a neural network model containing encoder-decoder structure is constructed, the model encoder integrates multi-scale Patch embedding and spatial-temporal joint attention mechanism, which can fully extract the wind speed and convection evolution characteristics in the strong convection system, and the decoder combines the attention guided time sequence information to reconstruct the future wind field image; Finally, in S400, the model is deployed to the business system, the real-time a-frame observation image is input, and the 10-minute, 1-kilometer resolution wind speed prediction sequence within a future a-frame (such as 60 minutes) is output, and the continuous wind speed distribution and the strong wind fall area prediction result of 8.0m / s and 17.2m / s are directly obtained.

[0068] Specifically, referring to Figure 2 It is shown that the complete technical roadmap of the model of the application is shown. The technical process includes acquisition and standardization preprocessing of multi-source observation data, time continuity screening and sample fusion, sliding window sample generation, deep neural network model training based on encoder-decoder structure, and continuous prediction of future multi-frame wind speed image, which clearly defines the overall technical logic of the system in the minute-level thunderstorm gale approaching prediction task.

[0066] It can be understood that the embodiment improves the identification ability of the extremely large wind area in the thunderstorm gale and the response ability of the minute-level prediction timeliness. By introducing high-frequency observation data and spatial interpolation processing, the difference problem of multi-source data in time and space scales is solved; the time continuity judgment and the sliding window sample construction method enhance the perception ability of the model to the strong wind evolution process with strong locality and high burst; the deep learning model based on the encoder-decoder accurately models the evolution trend of the wind speed at different scales through the space-time attention mechanism, and combines the loss strategy of wind speed intensity weighting, which improves the identification accuracy of the model in the high threshold wind area (≥17.2 m / s); the embodiment has the extremely large wind area warning ability of minute-level update frequency, 1 kilometer accuracy and high hit rate, which improves the accuracy of short-term nowcasting service and the foresight of disaster prevention and control.

[0067] In some embodiments of the present application, the multi-source meteorological observation data is acquired and preprocessed, and when the preprocessed data is obtained, it includes: combining the radar echo data, the maximum wind observation data of the ground automatic weather station and the lightning location data with the terrain elevation data, synchronously registering and uniformly cutting in the time and space dimensions to the same spatial scale and time resolution within the target area. And the time resolution of the radar echo data is 6 minutes, and the spatial resolution is 1 kilometer.

[0068] Specifically, when the radar echo data is processed, it includes: setting the value of the echo intensity lower than 10 dBZ in the radar echo data to zero, and eliminating the image frames with clutter and abnormalities, to generate a radar echo image.

[0069] When the maximum wind observation data of the ground automatic weather station is processed, it includes: gridding the maximum wind observation data of the discrete station through the inverse distance weighted interpolation method to the grid data consistent with the radar echo data, to generate a maximum wind image.

[0070] Specifically, when the lightning location data is processed, it includes: accumulating in 10-minute time periods in time, and expanding and assigning with the observation point as the center in space, to generate a lightning density map.

[0071] It can be understood that multi-source data (such as radar, wind speed, lightning, and terrain) is inconsistent in time resolution, spatial scale, and coordinate system due to coming from different devices and platforms. For this reason, the embodiment adopts time-space synchronization registration, frame alignment of data from different sources at the same time node, and unified clipping to a fixed spatial range (for example, Shandong Province area), and standardizes all data to a uniform grid format with a time step of 6 minutes and a spatial resolution of 1 kilometer, thereby ensuring that each frame of image has a strict space-time correspondence when the model is input. Radar echo images contain a large amount of non-meteorological clutter or weak reflection signals, which are not conducive to the model to extract meaningful convective information. Therefore, 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; at the same time, combined with image visualization means, manually or automatically identify and eliminate image frames with interference clutter or abnormal morphology, and generate radar echo image sequences that can be used for model training. The maximum wind data collected by the ground automatic station is a discrete point observation, and the spatial distribution is uneven, which is difficult to be directly used for the grid input of the deep neural network. The inverse distance weighted interpolation method is introduced, and the wind speed of each grid point is weighted and estimated according to the distance relationship between the observation point and the grid point, so that the wind speed field presents a smooth and continuous distribution in space. For areas where the ground automatic weather station is sparse, a minimum observation point threshold (such as ≥3) is set during interpolation to control the interpolation quality; the nearest neighbor interpolation can be used as a supplement in the boundary area with insufficient points to avoid the absence of extreme values. The interpolated wind speed image has consistent time and spatial resolution with the radar image. Lightning observation data is sparse and unstable in position, and cannot be directly used for model modeling. In order to enhance its spatial expression, the time-space accumulation and neighborhood expansion method is adopted: in the time dimension, all lightning positioning events within every 10 minutes are accumulated to form an activity map for the time period; in the spatial dimension, all grid points within a certain radius (such as 20 kilometers) centered on each lightning point are assigned a value of 1, realizing the spatial expansion of the lightning influence area; the values in the overlapping area of multiple points are added to form a lightning density map, which is used to represent the activity degree of the strong convective area.

[0072] It can be understood that by synchronously registering and standardizing and clipping the multi-source data such as radar, wind speed, lightning, and terrain, the structure of the input data is unified, the physics is consistent, and the resolution is aligned, providing an unambiguous input basis for subsequent model training; at the same time, by using physical threshold + image elimination + spatial interpolation + lightning expansion and other technical means, invalid data is effectively eliminated, the spatial expression ability and physical representativeness of the data are improved. The radar echo image, maximum wind image, and lightning density map constructed are highly consistent in time and space 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 the maximum wind area in thunderstorm gales.

[0073] In some embodiments of the present application, when the time continuity of the pre-processed data is judged and the pre-processed data is fused to obtain a training sample, it includes: checking the spatial proportion of the maximum wind image greater than the strong wind threshold in the pre-processed data based on the check function, and recording the time point information to generate a pkl file.

[0074] Data with adjacent data time difference less than or equal to ten minutes is recorded in a table file.

[0075] According to the time period existing in the table and the radar echo image and lightning density map at the same time, a training sample is formed.

[0076] Specifically, the key to training the neural network for maximum wind prediction is to ensure that the input data contains "meaningful strong wind process". If the wind speed in the image is generally weak and there is no strong wind area, the frame image has a lower effect on model training, and even may interfere with model learning. By presetting a wind speed threshold (such as 8.0 m / s or 17.2 m / s), the check function is used to traverse each frame of maximum wind image, and the ratio of the number of grid points with wind speed greater than the threshold to the total number of grid points in the image, i.e. the spatial proportion, is calculated. If the spatial proportion exceeds the set threshold (such as 0.5%, 1%), the time point of the frame is recorded, indicating that there is a possible thunderstorm gale process at that time. Finally, all time point information that meets the conditions is saved in the form of a pkl file for subsequent sample construction. The deep time sequence model has a high requirement for the continuity of the input time sequence. If there is a time jump or missing frame in the training sample, it will seriously affect the model's learning of the "wind speed evolution trend". After reading all the valid time points in the pkl file, based on the timestamp sorting, it is judged whether the interval between any two adjacent time points is less than or equal to 10 minutes (equal to 1 frame), so as to select the continuous time period. The time period information that meets the continuity condition is recorded into a table file (such as CSV or XLS) for clearly defining the start and end time of each sample. Model training requires multi-channel input (radar, wind speed, lightning, etc.) to be aligned at the same time. Based on each continuous time period in the table, the radar echo image, the interpolated maximum wind image and the generated lightning density map at the same time are extracted. The three types of images are channel spliced to form complete sample input data. Each training sample can contain several frames (such as 18 frames of input + 18 frames of prediction), which are generated in a sliding window manner, and finally used for model training, verification and testing.

[0077] It can be understood that through the automatic data screening and sample generation mechanism of "check function + time difference judgment + data group and", the problems of "high proportion of weak signal samples, serious time frame breakage, and asynchronous data channels" existing in traditional meteorological sample construction are solved. The samples used for training are real, continuous, and representative of the strong wind process, which improves the pertinence and generalization ability of model training; at the same time, through structured output (pkl file + table file), a controllable sample management system is formed, which is convenient for subsequent data updating and model retraining. The fitting ability of the model to the strong convective gale process is improved, and the recognition accuracy and timeliness response ability of the model to the sudden maximum wind area in actual application are enhanced.

[0078] In some embodiments of the present application, the training samples are processed based on a sliding window, and when the training samples are divided into a training set, a validation set and a test set in chronological order, the method comprises: dividing all the training samples into a training set, a validation set and a test set in chronological order, and the ratio is 8:1:1.

[0079] Specifically, the sliding window is a common data enhancement and sequence segmentation method in time series modeling. It takes a fixed number of frames (such as the first 18 frames) as the input window on the original time axis, and generates new samples by sliding a certain number of frames (such as 1 frame) backward, so that all continuous fragments in the time series can be used for training. For the time period data that has passed the time continuity judgment, frame slicing is performed according to the set window length (such as inputting 18 frames and predicting 18 frames), and traversal is performed according to a step size of 1, thereby constructing a large number of continuous samples and improving the learning coverage of the model. The chronological division emphasizes the non-intersection of the training data set, the validation data set and the test data set in the time dimension. It conforms to the logic of "training the model with historical data and predicting the future with future data" in real weather forecasting business, and can effectively prevent model overfitting caused by sample intersection. The generated sliding window samples are arranged in ascending order according to the sample start timestamp; the samples are divided into a training set (for model fitting), a validation set (for hyperparameter tuning), and a test set (for generalization performance evaluation) according to a ratio of 8:1:1; each type of data set does not overlap in time period, maintaining data independence.

[0080] It can be understood that through the sliding window enhanced sample quantity and the introduction of chronological division, the training set, the validation set and the test set each cover a representative thunderstorm gale process and do not interfere with each other in time, effectively improving the stability and generalization ability of model training. The continuous input characteristics of the sliding window ensure that the model fully learns the dynamic law of wind speed evolution, and the chronological division avoids the "information leakage" problem, so that the validation and test results can more truly reflect the prediction effect of the model in actual application. It not only considers the data utilization efficiency, but also ensures the scientificity of evaluation.

[0081] In some embodiments of the present application, when constructing a neural network model in the structure of an encoder-decoder, the encoder comprises:

[0082] 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 patches and extracting multi-scale features through a convolutional neural network.

[0083] a learnable position encoding module for adding an encoding vector representing time sequence and spatial position information to the embedded image patches to enhance the model's recognition ability of time variation and spatial distribution.

[0084] a space-time attention module including a window attention mechanism, a shift window attention mechanism and a time dimension attention mechanism, respectively used for extracting image local features, cross-region features and time sequence dependency.

[0085] a parallel space-time convolution module for enhancing the modeling ability of the encoder on local structure information to improve the generalization of the model.

[0086] Specifically, the multi-scale patch embedding module divides the input image (including the radar reflectivity image, the maximum wind image and the lightning density image) into a plurality of fixed-size image blocks (patches), such as 4x4, 8x8, 16x16 pixels; a lightweight convolutional neural network (CNN) is used to encode each patch to extract low-level and high-level multi-scale spatial features; the spatial position information is preserved so that the model can perceive the geographical distribution of the features in the subsequent processing. The learnable position encoding module constructs a set of trainable position encoding vectors corresponding to the time frame number and spatial position index of each patch in the image; the position encoding vector is added or spliced with the feature vector of each patch, so that the model has the ability to learn "when and where the image occurs"; compared with the traditional sinusoidal position encoding, the learnable position encoding is more adaptive and can be dynamically optimized in actual data. The window attention mechanism in the space-time attention module: a non-overlapping small area is defined in the image, and self-attention calculation is only performed in the local window to reduce the amount of calculation while focusing on the local wind field pattern; the window shift mechanism: the window boundary is staggered and shifted to enable the model to capture long-range spatial dependencies across regions; the time dimension attention mechanism: the values of the same grid point at different time frames are input as a sequence to establish a dynamic dependency relationship of wind speed evolution over time; the outputs of each attention module are fused together to improve the spatio-temporal perception and reasoning ability of the model. Although the Transformer structure has strong modeling ability, it lacks the ability to identify local textures and detailed structures, and is prone to overfitting; the convolution operation has good translation invariance and local pattern capture ability, which can make up for the shortcomings of the attention mechanism. The parallel spatio-temporal convolution module adds a convolution-based spatio-temporal feature extraction path in parallel to the backbone network; 3D convolution or separable spatio-temporal convolution structure is used to extract local details of wind speed changes from spatial and temporal dimensions respectively; the outputs of the attention module are fused to improve the robustness and generalization ability of the model.

[0087] It can be understood that the neural network encoder constructed by fusing multi-scale feature extraction, position perception modeling, space-time decoupled attention mechanism and parallel convolution structure. Not only can accurately capture the multi-dimensional dynamic relationship between radar reflectivity, maximum wind and lightning activity in strong convective systems, but also can effectively balance local precision and overall trend recognition, especially has excellent modeling ability for the "shape, change and motion" process of the maximum wind area. Compared with traditional deep networks or simple Transformer models, the encoder improves the stability, accuracy and generalization ability in the actual thunderstorm gale approaching prediction task while ensuring the depth and expression ability of the model, providing a model basis for minute-level maximum wind accurate early warning.

[0088] In some embodiments of the present application, the decoder is used to receive the spatio-temporal feature map output by the encoder, and includes:

[0089] An up-sampling reconstruction module is configured to restore the low-resolution feature map to the original spatial resolution by deconvolution or nearest neighbor interpolation method.

[0090] A feature fusion module is configured to splice and fuse the encoder output features of multiple scales.

[0091] A Patch de-embedding module is configured to restore the fused features to image structure, and output future a-frame wind speed prediction images, each frame interval being 10 minutes, and the spatial resolution being 1 kilometer.

[0092] Specifically, the Patch embedding adopts non-overlapping division, and the original image is disassembled into small blocks according to a fixed size (such as 8x8) and input into the encoder; in the Patch de-embedding stage, the Patch embedding vectors are rearranged according to the original spatial position, and the image continuity and edge connection are restored through convolution / deconvolution operation. In order to reduce the calculation complexity, the encoder will perform down-sampling processing (such as Patch Embedding) on the input image, resulting in a decrease in the spatial resolution of the feature map. The decoder needs to restore these low-resolution features to the original size, so as to output the wind speed image with spatial integrity. The up-sampling reconstruction module uses deconvolution to up-sample the spatial dimension of the feature map, has learning ability, and can dynamically adjust the restoration method according to the context content; or uses non-parametric methods such as nearest neighbor interpolation for fast up-sampling, which is suitable for deployment scenarios requiring fast inference speed; and ensures that the final output image and the original input image are consistent in spatial size (such as 1 kilometer resolution). The encoder outputs multi-scale and different-level spatio-temporal feature maps, and these features contain semantic information at different granularities. Fusing these information can improve the expression and detail quality of the output image. The feature fusion module splices the feature maps from different branches or different depths of the encoder in the channel direction, and uses convolution fusion operation (such as 1x1 convolution) on the spliced feature maps to reduce dimension and integrate information; the time axis consistency is maintained during the fusion process, so as to ensure the continuity and stability of the prediction image sequence in the time dimension. In the forward process of the model, the input image is subjected to Patch blocking and embedding, and in the decoding stage, these embedding vectors must be restored to complete two-dimensional image structure, so as to output the wind speed field image with real physical space meaning. The Patch de-embedding module rearranges the fused high-dimensional vectors according to the spatial position, and splices them into two-dimensional images; each reconstructed image frame corresponds to the prediction of wind speed distribution at a certain time in the future; the model outputs a sequence of a-frame images (such as 60 minutes in the future, each frame interval being 10 minutes, and a total of 6 frames) at one time, each frame having a spatial resolution of 1 kilometer, which is used to constitute the minute-level nowcast product.

[0093] It can be understood that by introducing the up-sampling reconstruction, feature fusion and patch de-embedding module, the image generation path is constructed. The abstract space-time features can be accurately restored to real wind speed images, which not only maintains the spatial continuity, but also improves the detail expression of the maximum wind drop area; the multi-scale feature fusion mechanism enhances the perception ability of different scale wind field structures, and the patch de-embedding module ensures the consistency and visual compatibility of the output results with the original meteorological observation image.

[0094] In some embodiments of the present application, when training the neural network model based on the training set, the validation set and the test set, the following steps are included:

[0095] The weighted mean square error is used as the main loss function, wherein a high loss weight is set for the grid points with wind speed ≥ 17.2 m / s.

[0096] A perception loss function is introduced to compare the feature distribution of the predicted image and the live image, and to maintain the clear spatial structure of the wind area.

[0097] A lightning density consistency loss term is added to strengthen the spatial overlap between the predicted wind area and the actual high lightning area.

[0098] The Adam optimizer is used and combined with the cosine annealing algorithm to adjust the learning rate, the training set and the validation set are alternately used to train and validate the model, and the model parameters with the highest key success index on the validation set are selected as the prediction model.

[0099] Specifically, the standard mean square error treats all grid points equally, but in the prediction of thunderstorm gales, the maximum wind area (e.g., ≥ 17.2 m / s) is the focus of prediction, usually accounting for a very small proportion. If not weighted, the model tends to optimize the low wind speed area, ignoring the high risk area. By setting a weight for the prediction error of each grid point: if the wind speed ≥ 17.2 m / s, set a larger weight (e.g., 5-10 times); the rest of the area sets the standard weight; in this way, the model pays more attention to the high wind speed area during the training process, thereby improving the identification ability and accuracy of the strong wind area. The conventional pixel-level loss ignores the overall perceptual consistency of the image structure, which is easy to cause the prediction result to be fuzzy and unshaped. The perceptual loss preserves the texture, structure and semantic boundary of the image by comparing the representation of the image in the high-dimensional feature space. The predicted image and the live image are input into the pre-trained image feature extraction network (such as VGG); the feature maps at the intermediate layer are extracted, and the L2 distance between the feature maps is calculated; this loss term serves as an additional constraint to guide the model to keep the edges, morphology and structure of the wind speed area clear. Lightning activity is highly correlated with thunderstorm intensity, and the strong wind area is often accompanied by frequent lightning. If the model can predict a high consistency between the high wind area and the lightning dense area, it means that the model has a good modeling of the convective structure. The actual lightning density map is used as an auxiliary supervision signal; the mask of the area ≥ 17.2 m / s in the model predicted wind speed map is extracted, and the spatial overlap rate (such as IoU or cross entropy) with the lightning density map is calculated; the consistency degree is taken as a loss term added to the total loss, guiding the model to more accurately locate the convective strong center. The lightning consistency loss uses the intersection over union (IoU) method: the ≥ 17.2 m / s area in the predicted wind speed map and the non-zero area in the lightning density map are converted into binary masks, and the IoU value is calculated, and 1-IoU is taken as a loss term added to the total loss function. Use the Adam optimizer combined with the weight decay strategy to improve stability and suppress overfitting; use the cosine annealing (Cosine Annealing) algorithm to dynamically adjust the learning rate, which helps the model to jump out of the local optimum; after each round of training, the performance is evaluated on the validation set; select the critical success index (CSI) as the verification indicator, and take the maximum wind area hit as the main evaluation basis; continuously save the model weight with the highest CSI on the validation set as the final deployment and use of the prediction model.

[0100] Specifically, refer to Figure 3 the average performance score results table on multiple test samples. By introducing multi-dimensional loss strategies such as weighted mean square error, perceptual loss and lightning consistency loss, and combining with indicators such as CSI for evaluation, it is verified that the model's identification accuracy of the ≥ 17.2 m / s wind area is significantly improved, the false negative rate is reduced, and the generalization ability is better than traditional methods. For example, Figure 4As shown, the model prediction effect of a typical thunderstorm gale process is shown. The evolution of the model prediction map and the real-time map (true value map) in the space structure is compared from T=3 to T=18 minutes. It can be seen that the model can not only accurately locate the occurrence area of the strong wind area, but also effectively capture the path and expansion trend of the time evolution, which embodies the short-term nowcasting ability of the model on the minute scale.

[0101] It can be understood that the training scheme improves the prediction ability of the model from three aspects of loss function design, supervised information enhancement and optimization strategy selection. By introducing weighted MSE, the model focuses on the maximum wind area; the addition of perception loss makes the output wind area boundary clear and complete; the perception loss uses VGG-16 network (ImageNet pre-training), and the L2 distance calculation is performed on the activation outputs of conv3_3 and conv4_3 layers; the VGG network weight is kept in a frozen state during the training process and does not participate in gradient update. The lightning consistency loss enhances the spatial matching ability of the model to the deep features of the strong convective system; in combination with the Adam optimizer and the cosine learning rate annealing mechanism, the training stability and convergence efficiency are improved.

[0102] In some embodiments of the present application, the training samples are processed based on a sliding window, and when the training samples are divided into a training set, a validation set and a test set in chronological order, the method further comprises:

[0103] The sliding window length is 36 frames, the input sequence is composed of the first 18 frames of multi-source image data, and the prediction sequence is the last 18 frames of maximum wind image.

[0104] Specifically, the interval of each frame of the sliding window is 10 minutes, so 6 frames represent 60 minutes of observation data as input; the sliding step is 1 frame, i.e. 10 minutes, which ensures the time continuity and sufficiency of the sample generation.

[0105] In some embodiments of the present application, the future a-frame 1-kilometer resolution wind speed prediction image is obtained, and the output of the wind speed prediction image includes:

[0106] The continuous value wind speed prediction map is used to display the predicted wind speed value of each grid point at each time.

[0107] The ≥8.0m / s wind speed area binary map is used to represent the distribution of the strong wind area.

[0108] The ≥17.2m / s maximum wind area binary map is used to represent the alarm area of the extreme wind area.

[0109] The multi-channel image sequence is stored according to the frame number, supports dynamic map display and radar real-time comparison superposition, and is used for Web-GIS platform display and mobile terminal warning interface push.

[0110] Specifically, weather events have temporal evolution characteristics, and neural network models need to capture the "rise, peak, and decline" change process of strong winds through sequence modeling. Using a sliding window approach can maximize the use of continuous observation sequences and improve data utilization. The sliding window length is set to 36 frames, with the first 18 frames as input representing the weather evolution process in the past 180 minutes; the last 18 frames as the prediction target representing the maximum wind evolution trend in the future 180 minutes; the window is slid every 10 minutes (i.e., the sample step is 1 frame), generating a large number of supervised learning samples from all data segments that meet the time continuity requirement, improving sample quantity and spatiotemporal diversity; each frame in the input sample is composed of multiple channel images such as radar echo maps, maximum wind maps, and lightning density maps, and the prediction label is a sequence of 18 consecutive maximum wind images. The model output needs to meet the business needs of multi-level users: not only to provide continuous wind speed values to support quantitative analysis, but also to provide binary fall zone maps to support early warning signal generation and layer overlay display. The output results include a-frame (e.g., 18 frames) wind speed prediction maps, with a time interval of 10 minutes and a spatial resolution of 1 kilometer; each frame contains the following three forms:

[0111] Continuous value wind speed prediction map: provides quantitative wind speed prediction values for each grid point at future time, for quantitative analysis and structure evaluation;

[0112] ≥8.0m / s wind speed fall zone map (binary): used to delineate general strong wind fall zone, supporting low-level wind warning discrimination;

[0113] ≥17.2m / s maximum wind fall zone map (binary): directly serves high-level disaster warning release, marking extreme wind risk area;

[0114] The output is numbered and stored in a multi-channel image sequence structure (e.g., T+10, T+20...), with uniform data structure for each frame, facilitating post-processing; it can be directly connected to Web-GIS platform for layer overlay, dynamic animation playback, and comparison with radar live; it also supports warning information push through interface to mobile terminal or emergency system. Model output images are stored in multi-channel tiff format or NetCDF format, with frame naming method as "YYYYMMDD_HHMM_T+10" etc.; dynamic image sequences generate MP4 or GIF animations by merging a-frame images, providing JSON interface for Web platform layer overlay and playback control.

[0115] It can be understood that by setting the sliding window mechanism, the time series modeling capability of the model for strong wind processes is enhanced, the data utilization efficiency is improved, and the model is more likely to capture the wind speed evolution trend; at the same time, by means of the clear structure and rich content of the output format, the model results are divided into two levels of continuous value prediction graph + binary risk falling area graph, which can be used for quantitative judgment of meteorological analysts, and is also suitable for rapid identification needs of automatic alarm system, and enhances the visualization, usability and business integration capability of model results; combined with the dynamic graph and numbering sequence mechanism, the deployability of the minute-level rolling forecast product in the actual platform is facilitated.

[0116] The thunderstorm gale nowcasting method in the above embodiment takes minute-level timeliness and extreme wind falling area identification as the core target, breaks through the limitations of traditional extrapolation method and numerical prediction in timeliness and extreme wind falling area identification. Using radar echo, ground maximum wind, lightning location and terrain elevation and other multi-source observation data, after noise processing and consistency registration, high-quality, unified standard training sample construction is realized, and the comprehensive perception ability of strong convective features is improved. A deep neural network model with encoder-decoder structure is adopted, combined with multi-scale Patch embedding and space-time attention mechanism, to capture the "birth-movement-change" evolution characteristics of the maximum wind area. During the training process, high spatio-temporal resolution samples generated based on the sliding window and the weighted loss strategy make the model have higher discrimination ability under the threshold of ≥17.2m / s wind speed. Through the way of real-time rolling update every 10 minutes, the future 1-hour frame-by-frame 1-kilometer resolution wind speed distribution graph and strong wind falling area prediction results are output, realizing the minute-level strong wind nowcasting.

[0117] In another preferred mode based on the above embodiment, referring to Figure 5 The present embodiment provides a minute-level thunderstorm gale nowcasting system for applying the above minute-level thunderstorm gale nowcasting method, which comprises:

[0118] 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 including radar echo data, ground automatic weather station maximum wind observation data, lightning location data and terrain elevation data, and the pre-processing including noise processing and consistency processing.

[0119] The judgment unit is configured to perform time continuity judgment on the processed data and fuse the processed data to obtain training samples.

[0120] 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.

[0121] a frame observation data in real time based on a prediction model to obtain a future a frame 1-kilometer resolution wind speed prediction image, the wind speed prediction image including a wind speed continuous distribution map and ≥8.0 m / s and ≥17.2 m / s wind area landing area prediction results.

[0122] It can be understood that, through the thunderstorm gale approaching prediction method taking the minute-level timeliness and great wind landing area identification as the core target, the limitations of the traditional extrapolation method and numerical prediction in timeliness and strong wind landing area identification are broken through. Using radar echo, ground maximum wind, lightning positioning and terrain elevation and other multi-source observation data, after noise processing and consistency registration, high-quality, unified standard training sample construction is realized, and the comprehensive perception ability of strong convective features is improved. A deep neural network model with an encoder-decoder structure is adopted, combined with multi-scale Patch embedding and space-time attention mechanism, to capture the "birth-movement-change" evolution characteristics of the maximum wind area. In the training process, high spatio-temporal resolution samples generated based on sliding window and weighted loss strategy make the model have higher discrimination ability under the ≥17.2 m / s wind speed threshold. Through the real-time rolling update every 10 minutes, the future 1 hour frame-by-frame 1-kilometer resolution wind speed distribution map and strong wind landing area prediction results are output, realizing the minute-level strong wind approaching prediction.

[0123] Those skilled in the art will understand that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0124] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the 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 produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks.

[0125] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0126] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0127] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the same. Even though the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently, and any modification or replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A thunderstorm gale approach forecast method for a minute level, characterized by, The method comprises the following steps: acquiring multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data, wherein the multi-source meteorological observation data comprises radar echo data, maximum wind observation data of a ground automatic weather station, lightning positioning data and terrain elevation data, and the preprocessing comprises noise processing and consistency processing; judging the time continuity of the preprocessed data and fusing the preprocessed 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, training the neural network model based on the training set, the validation set and the test set, and obtaining a prediction model; processing real-time a-frame observation data based on the prediction model to obtain a future a-frame 1-kilometer resolution wind speed prediction image, wherein the wind speed prediction image comprises a wind speed continuous distribution map and a prediction result of a wind area falling area of ≥8.0 m / s and ≥17.2 m / s; when training the neural network model based on the training set, the validation set and the test set, the following steps are included: using a weighted mean square error as a main loss function, wherein a high loss weight is set for a grid point with a wind speed of ≥17.2 m / s; introducing a perception loss function for comparing the feature distribution of the predicted image and the real image to maintain the spatial structure of the wind area clear; adding a lightning density consistency loss term to strengthen the spatial overlap between the predicted wind area and the actual lightning high-occurrence area; using an Adam optimizer and combining a cosine annealing algorithm to adjust the learning rate, training and verifying the model by alternately using the training set and the validation set, and selecting the model parameters with the highest key success index on the validation set as the prediction model; when processing the training samples based on the sliding window and dividing the training samples into the training set, the validation set and the test set in chronological order, the following steps are included: the length of the sliding window is 36 frames, the input sequence is composed of the first 18 frames of multi-source image data, and the prediction sequence is the last 18 frames of maximum wind image.

2. The minute-level thunderstorm-gale approaching forecast method according to claim 1, characterized in that, when acquiring multi-source meteorological observation data and preprocessing the multi-source meteorological observation data to obtain preprocessed data, the following steps are included: combining the radar echo data, the maximum wind observation data of the ground automatic weather station and the lightning positioning data with the terrain elevation data, synchronously registering and uniformly cropping them into the same spatial scale and time resolution within the target area in the time and space dimensions; and the time resolution of the radar echo data is 6 minutes and the spatial resolution is 1 kilometer; when processing the radar echo data, the following steps are included: setting the values of the radar echo data with an echo intensity lower than 10 dBZ to zero, and eliminating image frames with clutter and abnormalities to generate a radar echo image; when processing the maximum wind observation data of the ground automatic weather station, the following steps are included: gridding the maximum wind observation data of discrete stations into grid data consistent with the radar echo data by an inverse distance weighted interpolation method to generate a maximum wind image; The consistency processing of the lightning positioning data comprises: 10-minute time period accumulation in time, and radius expansion assignment taking the observation point as the center in space to generate a lightning density map.

3. The minute-level thunderstorm-gale approaching forecast method according to claim 1, characterized by, The time continuity judgment of the preprocessed data and the fusion of the preprocessed data to obtain a training sample comprise: Based on the check function, the spatial proportion of the maximum wind image in the preprocessed data greater than the strong wind threshold is checked, and the time point information is recorded to generate a pkl file; Data with a time difference of less than or equal to ten minutes are recorded in a table file; According to the time period existing in the table and the radar echo image and the lightning density map at the same time, a training sample is combined.

4. The minute-level thunderstorm-gale approaching forecast method according to claim 1, characterized by, Based on the sliding window, the training sample is processed, and the training sample is divided into a training set, a validation set and a test set in time sequence, comprising: All the training samples are divided into a training set, a validation set and a test set in time sequence, and the proportion is 8:1:

1.

5. The minute-level thunderstorm-gale approaching forecast method according to claim 1, characterized by, When constructing a neural network model with an encoder-decoder structure, the encoder comprises: A multi-scale Patch embedding module is used to divide the radar echo image, the maximum wind image and the lightning density image in the training sample into image blocks, and to extract multi-scale features through a convolutional neural network; A learnable position encoding module is used to add an encoding vector representing time sequence and spatial position information to the embedded image blocks, to enhance the model's recognition ability for time variation and spatial distribution; A space-time attention module includes window attention mechanism, transfer window attention mechanism and time dimension attention mechanism, which are respectively used to extract image local features, cross-region features and time sequence dependence; A parallel space-time convolution module is used to enhance the modeling ability of the encoder for local structure information to improve the generalization of the model.

6. The minute-level thunderstorm gale approaching forecast method according to claim 5, characterized in that, The decoder is used to receive the space-time feature map output by the encoder, and comprises: An up-sampling reconstruction module is used to restore the low-resolution feature map to the original spatial resolution through deconvolution or nearest neighbor interpolation method; A feature fusion module is used to splice and fuse the encoder output features of multiple scales; A Patch de-embedding module is used to restore the fused features to image structure, and output a future a-frame wind speed prediction image, with a frame interval of 10 minutes and a spatial resolution of 1 kilometer.

7. The minute-level thunderstorm-gale approaching forecast method according to claim 1, characterized by, The future a-frame 1-kilometer resolution wind speed prediction image is obtained, and the output of the wind speed prediction image comprises: A continuous value wind speed prediction map is used to display the predicted wind speed value of each grid point at each time; A ≥8.0m / s wind speed falling area binary map is used to represent the strong wind falling area distribution; A ≥17.2m / s maximum wind falling area binary map is used to represent the alarm area of the extreme wind area; A multi-channel image sequence is stored by frame number, supports dynamic image display and radar live comparison superposition, and is used for Web-GIS platform display and mobile terminal warning interface push.

8. A minute-scale-oriented thunderstorm gale nowcast system for applying the minute-scale-oriented thunderstorm gale nowcast method according to any one of claims 1 to 7, characterized in that, It comprises: The collection 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 including radar echo data, ground automatic weather station maximum wind observation data, lightning location data and terrain elevation data, and the pre-processing including noise processing and consistency processing; The judgment unit is configured to perform time continuity judgment on the processed data and fuse the processed data to obtain training samples; The processing unit is 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 prediction unit is configured to process real-time a-frame observation data based on the prediction model to obtain a future a-frame 1-kilometer resolution wind speed prediction image, the wind speed prediction image including a wind speed continuous distribution map and a prediction result of a wind area falling area of ≥8.0 m / s and ≥17.2 m / s.

Images