Intelligent thunderstorm weather forecasting method and system and medium

Abstract

The invention relates to an intelligent thunderstorm weather forecasting method and system and a medium, belongs to the field of thunderstorm forecasting, and provides the intelligent thunderstorm weather forecasting method for solving the problem that an existing mode is insufficient in forecasting capacity, and the intelligent thunderstorm weather forecasting method comprises the following steps: collecting and processing initial data, and obtaining a multi-source fusion initial field of a unified format; constructing a short-time forecasting model based on a Swin Transform backbone network, and outputting a thunderstorm forecasting field in the future 6 hours by inputting a multi-source fusion initial field and underlying surface information; and a short-term forecasting model based on a FuXi large model is constructed, and by inputting the multi-source fusion initial field and underlying surface information obtained in the step 1, possible thunderstorm occurrence areas and intensity distribution data in more than six hours in the future are output. By introducing a Swin Transform backbone network and combining an autoregression prediction strategy and a multi-task loss function, the capturing precision and real-time performance of the thunderstorm time-space evolution process are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of thunderstorm prediction, and particularly relates to an intelligent thunderstorm weather forecasting method, system and medium. BACKGROUND

[0002] As a typical severe convective weather, thunderstorm has a significant impact on aviation, transportation, and agriculture due to its suddenness and destructiveness. Although traditional numerical weather prediction plays an important role in thunderstorm prediction, it is difficult to meet the actual demand due to its resolution, real-time performance, and limitations in capturing small-scale systems. In recent years, with the rapid development of deep learning technology, its advantages in multi-modal data processing, spatio-temporal feature extraction, and efficient prediction have shown great potential in the field of thunderstorm prediction.

[0003] First, efficient fusion of multi-modal data is an important application direction of deep learning in thunderstorm prediction. Traditional thunderstorm prediction relies on a single data source, such as radar echoes or satellite observations, while deep learning models can fuse radar, satellite, ground observations, and numerical prediction data, thereby improving the comprehensiveness and accuracy of prediction. For example, a spatio-temporal sequence prediction model based on ConvLSTM is proposed for radar echo extrapolation, which significantly improves the short-term prediction ability of thunderstorm development. However, these multi-modal data fusion models still face problems such as inconsistent data quality, time synchronization difficulties, and high computational resource requirements, which limit their widespread application in complex weather systems.

[0004] Second, the advantages of deep learning in spatio-temporal feature extraction provide a new approach to thunderstorm prediction. Traditional numerical prediction models are difficult to capture the nonlinear characteristics of thunderstorms, while deep learning models (such as convolutional neural networks CNN, recurrent neural networks RNN, etc.) can automatically learn features and extract key patterns of thunderstorm development. In addition, techniques such as generative adversarial networks (GAN) are used for uncertainty quantification of thunderstorms, providing a new method for risk assessment. However, deep learning models rely on a large amount of labeled data in the process of spatio-temporal feature extraction, and are susceptible to overfitting problems. At the same time, GAN may generate false samples in uncertainty quantification, affecting the reliability of the prediction results.

[0005] In terms of real-time performance, deep learning models can quickly process large-scale meteorological data to achieve short-term nowcasting of thunderstorms. For example, using the U-Net architecture for radar echo extrapolation, high-resolution short-term thunderstorm forecasts have been successfully achieved, and the computational efficiency is superior to traditional methods. In addition, the introduction of transfer learning technology enables deep learning models to quickly adapt in data-scarce areas, further enhancing the model's generalization ability. However, real-time models may still face computational bottlenecks when processing ultra-large-scale data, especially during extreme weather events. In addition, the effectiveness of transfer learning depends on the quality of source domain data and the similarity of the target domain. If the two are significantly different, the model performance may decrease significantly.

[0006] Despite the many advantages of deep learning in thunderstorm prediction, it still faces many challenges, such as the lack of physical interpretation of the model, the limitation of training data quality, and the need to improve the prediction ability of extreme weather events. SUMMARY

[0007] To solve the technical problems existing in the prior art, the present application provides an intelligent thunderstorm weather forecasting method, system and medium, which introduces a Swin Transformer backbone network, combines autoregressive prediction strategy and multi-task loss function, and effectively improves the capture accuracy and real-time performance of the thunderstorm spatiotemporal evolution process.

[0008] An intelligent thunderstorm weather forecasting method, comprising the following processes: Step 1, collect and process initial data to obtain multi-source fusion initial fields in a unified format; Step 2, build a short-term forecasting model based on the Swin Transformer backbone network, input the multi-source fusion initial fields and underlying surface information obtained in step 1, and then output the forecast field of thunderstorms within the next 6 hours; Step 3, build a short-term forecasting model based on the FuXi large model, input the multi-source fusion initial fields and underlying surface information obtained in step 1, and then output the possible occurrence area and intensity distribution data of thunderstorms more than 6 hours in the future.

[0009] The application fuses deep learning methods, uses multi-source data (satellites, radars, ground stations, weather forecast data, underlying surface static data, etc.) and physically refined convective parameters to construct a thunderstorm weather forecasting system with short-term forecasting and short-term diagnosis. By introducing the Swin Transformer backbone network, combining the autoregressive prediction strategy and the multi-task loss function (regression + classification), the accuracy and real-time of capturing the spatio-temporal evolution of thunderstorms is effectively improved. The application uses thunderstorm short-term forecasting (0-6 hours) and short-term forecasting diagnosis based on the FuXi large model (6-72 hours). The short-term forecasting part uses an autoregressive form, which not only inputs the prediction of the previous time step, but also combines the underlying surface information and the FuXi forecast field at the corresponding time; the short-term diagnosis part directly uses the multi-time FuXi forecast as input to predict the thunderstorm result at the corresponding time. Both parts are based on the Encoder-Decoder structure and use stacked Swin Transformer (or 3D Transformer) as the backbone network to efficiently represent high-dimensional and multi-scale features.

[0010] Further, the initial data includes satellite remote sensing information, radar detection information, meteorological station observation information, convective parameters and lightning data.

[0011] Further, the specific process of step 1 includes: S1.1, pre-processing the initial data, the pre-processing including denoising, missing value filling and outlier detection to ensure data quality; S1.2, resampling the pre-processed data with a unified spatial and temporal resolution to obtain resampled data, ensuring the spatio-temporal consistency between different data sources; S1.3, converting the resampled data into a unified format by standardization or normalization method to adapt to the input requirements of subsequent models. For example, satellite remote sensing information and radar detection data can be converted into rasterized form, while meteorological station observation data and convective parameters are feature encoded to ensure that multi-source data can be fused in the same framework.

[0012] The purpose of this step is to generate a multi-source fusion initial field with consistent format, ensuring the uniformity and high quality of input data, and providing a reliable basis for subsequent model input. Convective parameters are key indicators extracted from physical elements of original products, directly representing atmospheric instability and convective potential, which can help the model quickly focus on the key physical mechanism of thunderstorm occurrence.

[0013] Further, the specific process of constructing the short-term forecasting model includes: Step 2.1, quality control, standardization and spatial resampling of the results of step 1 through a feature fusion layer to obtain thunderstorm-related features; Step 2.2, based on the Swin Transformer backbone network, a 2D Encoder and a 2D Decoder module are constructed to obtain a short-term prediction model. The thunderstorm-related features processed in step 2.1 and the underlying surface information are input into the short-term prediction model. Through the processing of the short-term prediction model, the original spatial data is mapped into feature space data to extract the spatio-temporal features of short-term thunderstorm evolution. The role of this step is to capture the short-term dynamic changes of thunderstorms through the short-term prediction model to improve the accuracy of short-term prediction.

[0014] Step 2.3, through the autoregressive iteration method, the results obtained in step 2.2 are iterated multiple times to output continuous lightning intensity prediction values and gradually generate the thunderstorm evolution trend within the next 6 hours. The role of this step is to ensure the coherence and real-time nature of the short-term prediction results through the capture ability of the autoregressive iteration method for thunderstorm spatio-temporal evolution. Step 2.4, set the threshold value of thunderstorm occurrence, use the loss function to calculate the error between the lightning intensity obtained in step 2.3 and the true lightning intensity, and then evaluate the accuracy of the short-term prediction model.

[0015] The training process adopts dual objectives of regression and classification to optimize the model performance more comprehensively. The regression part aims at the prediction of actual lightning times, using the loss function to measure the error between the model output lightning intensity and the true lightning intensity, thereby improving the numerical prediction ability of the model for lightning intensity. The classification part is for the prediction of thunderstorm occurrence or not, through the cross-entropy loss function to model the occurrence of thunderstorm events as a binary classification, ensuring that the model can accurately distinguish between thunderstorm occurrence areas and non-thunderstorm areas.

[0016] The input data for the regression task mainly comes from multi-source lightning observation data and fusion feature fields, and the output is continuous lightning intensity prediction values. The classification task is based on the same input data, and by setting the threshold value of thunderstorm occurrence, the thunderstorm event is converted into a binary classification problem (occurrence / non-occurrence).

[0017] Further, the loss function in step 2.4 includes Charbonnier loss function and BCE loss function, and the loss value obtained by Charbonnier loss function and the loss value obtained by BCE loss function are weighted and added to obtain the final loss value, which is the error between the lightning intensity obtained in step 2.3 and the true lightning intensity.

[0018] wherein the Charbonnier loss function L is defined as:

[0019] wherein, is the Charbonnier loss function, is referred to as a predicted value, is referred to as a true value, is set to 10 -3 ; is defined as follows:

[0020] where y represents the true label (0 or 1), and p represents the probability value (mapped to 0-1 by the sigmoid function) predicted by the model. The advantage of BCE loss is that it can directly output probability prediction, making the prediction result more interpretable and practical. Preferably, the weight of BCEloss is set to 0.5, and the Charbonnier loss is set to 1.

[0021] Considering that in lightning prediction, the model may cause double penalties (Double Penalty) of lightning prediction area omission penalty and area false alarm penalty due to the superposition of spatial position deviation and magnitude deviation, we use a classification branch to focus on whether the occurrence of a precipitation event can be correctly captured, and a regression branch to quantitatively describe the precipitation event that has been determined. Therefore, the loss function for training is also divided into two parts. The first part uses Charbonnier loss, which is a smooth L1 norm approximation loss function commonly used in deep learning tasks such as image restoration, denoising, and super-resolution. Compared with directly using L1 loss, Charbonnier loss is smooth at the origin, facilitating gradient calculation and avoiding the problem of unstable gradient when approaching zero.

[0022] Further, the specific process of constructing the short-term prediction model includes: Step 3.1, based on the FuXi large model, construct 3D Encoder and 3D Decoder modules to obtain a short-term prediction model for capturing the spatiotemporal evolution characteristics of thunderstorms in the future 6 hours or more; Step 3.2, use the data obtained in step 1 and underlying surface information as the input of the short-term prediction model, train the parameters of the 3D Encoder and 3D Decoder modules using multiple time unified training methods, and output the occurrence area and intensity distribution of future thunderstorms in the future 6 hours or more, thereby obtaining a lightning diagnostic field. The role of this step is to model the evolution of thunderstorms on a longer time scale through a deep learning network, improving the spatial resolution and temporal extendibility of short-term prediction.

[0023] The role of this step is to provide complete historical background and short-term trend information for the short-term prediction model, ensuring the continuity and accuracy of the prediction.

[0024] The parameters of the 3D Encoder and 3D Decoder modules are trained by unified training methods at multiple time points to ensure that the model can accurately predict the occurrence area and intensity distribution of future thunderstorms. The final output is a lightning diagnostic field for the next 7 to 72 hours, providing specific results for short-term thunderstorm forecasting. The purpose of this step is to enhance the generalization ability of the model through long-term training, providing high-quality short-term thunderstorm prediction information for weather forecasting.

[0025] An intelligent thunderstorm weather forecasting system comprises: A data processing module collects and processes initial data to obtain multi-source fusion initial fields in a unified format. A short-term forecasting module constructs a short-term forecasting model based on a Swin Transformer backbone network, inputs the multi-source fusion initial fields and underlying surface information obtained by the data processing module, and then outputs a thunderstorm forecasting field for the next 6 hours. A short-term forecasting module constructs a short-term forecasting model based on a FuXi large model, inputs the multi-source fusion initial fields and underlying surface information obtained by the data processing module, and then outputs the possible occurrence area and intensity distribution data of thunderstorms more than 6 hours in the future.

[0026] A computer-readable medium comprising a memory and one or more processors, the memory storing executable code, and the one or more processors executing the executable code to implement the intelligent thunderstorm weather forecasting method.

[0027] Compared with the prior art, the beneficial effects of the present application are as follows: (1) Traditional thunderstorm forecasting mainly relies on numerical models and human experience, with limited spatial resolution and real-time performance. The present scheme integrates multi-source satellite, radar, ground observation, FuXi large model prediction, underlying surface static information, and convective parameters, making the thunderstorm forecasting more comprehensive and delicate in space and time, so as to achieve more accurate prediction results. The integration of short-term and short-term forecasting breaks through the limitations of traditional thunderstorm forecasting in a single time scale, and through short-term autoregressive prediction within 6 hours and diagnostic prediction based on the FuXi large model, it realizes more detailed forecasting levels and time span coverage.

[0028] (2) In model design, traditional CNN or basic regression methods are difficult to adapt to thunderstorms, which are strong convective weather with variable space-time scales. The present application uses a Swin Transformer backbone network, which uses local attention mechanisms and sliding window strategies, and deeply integrates satellite, radar, ground observation, and underlying surface static data to enhance the ability to capture and represent complex thunderstorm space-time features, and better capture multi-scale dynamic features.

[0029] (3) In terms of prediction range and form, the scheme distinguishes between short-term prediction within 6 hours and short-term diagnosis for more than 6 hours, realizes rapid refreshing and medium-term trend control, and significantly enhances the performance of refinement and real-time. During the training process, a dual loss function of regression and classification is adopted, and the POD, FAR, TS and other indicators are used for comprehensive evaluation. Not only does it focus on whether the thunderstorm occurs (classification), but also quantifies the number of lightning strikes using the regression branch, supplemented by the combination of Charbonnier and cross-entropy loss in multi-task training, and the use of POD, FAR, TS and other meteorological indicators for comprehensive measurement, which can effectively balance the hit rate and false alarm rate, and further improve the practicality and accuracy of thunderstorm prediction. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a flowchart of Example 1; Figure 2 is a structural diagram of a short-term prediction model; Figure 3 is a result diagram of an application example. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0032] Example 1 An intelligent thunderstorm weather prediction method, the flowchart is as shown in Figure 1 , including the following processes: Step 1, collect and process initial data to obtain a unified format of multi-source fusion initial field; the initial data includes satellite remote sensing information, radar detection information, weather station observation information, convective parameters and lightning data, and the specific process includes: S1.1, pre-processing the initial data, the pre-processing includes denoising, missing value filling and outlier detection to ensure data quality; remove outliers by data cleaning, and match data from different sources in space and time; S1.2, resample the pre-processed data with a unified spatial and temporal resolution to obtain resampled data, and ensure the spatio-temporal consistency between different data sources; S1.3, convert the resampled data into a unified format by standardization or normalization method to adapt to the input requirements of subsequent models. For example, satellite remote sensing information and radar detection data can be converted into a grid form, and weather station observation data and convective parameters can be feature encoded to ensure that multi-source data can be fused and processed in the same framework.

[0033] Step 2, a short-term prediction model based on the Swin Transformer backbone network is constructed. The Swin Transformer contains multiple stacked Swin Transformer blocks. After each block, the image is down-sampled by a factor of two. Each block adopts a sliding window attention mechanism. By inputting the multi-source fusion initial field and underlying surface information obtained in step 1, the prediction field of thunderstorms within the next 6 hours is output. The specific process includes: using a multi-scale feature extraction enhancement model to enhance the expression ability of complex meteorological data, combining an attention mechanism to optimize the prediction accuracy of key areas, and capturing the dynamic characteristics of thunderstorm evolution through a time convolution module to achieve efficient short-term prediction.

[0034] The structure of the short-term prediction model is shown in Figure 2 Figure 2 (a) is the overall architecture of the SwinTransformer, which is composed of multiple layers of stacked SwinTransformer modules, and Figure 2 (b) is the structure of two SwinTransformer modules.

[0035] Step 2.1, the results of step 1 are quality controlled, standardized, and spatially resampled through a feature fusion layer to obtain thunderstorm-related features. Step 2.2, based on the Swin Transformer backbone network, a 2D Encoder and a 2D Decoder module are constructed to obtain a short-term prediction model. The thunderstorm-related features processed in step 2.1 and the underlying surface information are input into the short-term prediction model to extract the spatio-temporal features of short-term thunderstorm evolution. The role of this step is to capture the short-term dynamic changes of thunderstorms through a deep learning model to improve the accuracy of short-term prediction.

[0036] Step 2.3, the results obtained in step 2.2 are iterated multiple times through autoregressive iteration to output continuous lightning intensity prediction values and gradually generate the evolution trend of thunderstorms within the next 6 hours. The role of this step is to enhance the model's ability to capture the spatio-temporal evolution of thunderstorms through an autoregressive strategy, ensuring the coherence and real-time nature of the short-term prediction results. Step 2.4, the error between the lightning intensity obtained in step 2.3 and the true lightning intensity is calculated using a loss function, and a threshold for thunderstorm occurrence is set to evaluate the accuracy of the short-term prediction model. The loss function includes a Charbonnier loss function and a BCE loss function. The Charbonnier loss function is defined as:

[0037] where, is the Charbonnier loss function, ​is referred to as a predicted value, is referred to as a true value, is set to 10 -3 ; is defined as follows:

[0038] where y represents the true label (0 or 1), and p represents the probability value predicted by the model (mapped to 0-1 by the sigmoid function). The advantage of BCE loss is that it can directly output probability prediction, making the prediction result more interpretable and practical. The loss function in step 2.4 includes Charbonnier loss function and BCE loss function, and the loss value obtained by Charbonnier loss function and the loss value obtained by BCE loss function are weighted and added to obtain the final loss value, which is the error between the lightning intensity obtained in step 2.3 and the true lightning intensity. The final loss is the weighted sum of the two losses, the weight of BCEloss is set to 0.5, and the weight of Charbonnier loss is set to 1.

[0039] The prediction effect of the model in step 2 is evaluated by meteorological indicators such as POD (hit rate), FAR (false alarm rate), and TS (Threat Score). These indicators can comprehensively measure the performance of the model in the thunderstorm prediction scenario, ensuring that the model hits the real thunderstorm event while minimizing false alarms, thereby achieving a balance between accuracy and reliability.

[0040] Step 3, build a short-term prediction model based on FuXi large model, input the multi-source fusion initial field and underlying surface information obtained in step 1, and then output the possible occurrence area and intensity distribution data of thunderstorms in 7-72 hours. The specific process includes: Step 3.1, based on FuXi large model, build 3D Encoder and 3D Decoder modules to obtain a short-term prediction model for capturing the spatiotemporal evolution characteristics of thunderstorms in the next 7-72 hours; Step 3.2, input the data obtained in step 1 and underlying surface information into the short-term prediction model, use multiple time unified training method, optimize the loss function by Adam, and then train the parameters of 3D Encoder and 3D Decoder modules to output the occurrence area and intensity distribution of future thunderstorms in more than 6 hours, and then obtain the lightning diagnostic field. The role of this step is to model the evolution of thunderstorms in a longer time scale through a deep learning network, improving the spatial resolution and time extension of short-term prediction.

[0041] The application combines thunderstorm short-term prediction and short-term prediction diagnosis based on FuXi large model by deep learning fusion of multi-source observation and prediction data, and constructs a nationwide thunderstorm weather prediction model with high precision and high timeliness by using Swin Transformer. At the data level, satellite, radar, ground station and convection parameter information are integrated, and the following static data of underlying surface are used to more comprehensively depict the thunderstorm generation mechanism; at the model level, an autoregressive strategy is introduced to realize 0-6 hour short-term prediction, and multi-time FuXi prediction is directly input to complete 7-72 hour diagnosis; at the training and evaluation level, through the multi-task mechanism of Charbonnier loss and cross-entropy loss, combined with POD, FAR and TS indexes, the balance between the sensitivity of capturing thunderstorm events and reducing false alarms is ensured, so as to provide a more reliable reference basis for thunderstorm prediction.

[0042] Since lightning belongs to an element with large spatio-temporal variability, the influence time range of the initial field is short, so the combination scheme of lightning short-term prediction (0-6 hours) + mode short-term prediction result diagnosis (6-72 hours) is adopted.

[0043] The lightning short-term prediction model uses t-1 and t0 time multi-modal data and grid lightning data as the initial information, and then passes through a deep learning model in the form of autoregression. In each autoregressive iteration process, in addition to the output of the last time, the underlying surface information and the FuXi prediction field corresponding to the next time output are also added to the input end, and the lightning prediction field of 0-6 hours is obtained after 6 iterations.

[0044] The lightning diagnosis model based on FuXi short-term prediction directly uses the subsequent 66 times as the time channel, constructs a 3D input, and directly predicts the lightning field corresponding to the 66 times by adding the underlying surface information.

[0045] Both of the above two models adopt the Encoder-Decoder architecture, and the backbone adopts a stacked SwinTransformer. The lightning short-term prediction model and the lightning diagnosis model based on FuXi short-term prediction both adopt a stacked SwinTransformer (the diagnosis model can be improved to a 3D Transformer). Swin Transformer is a high-efficiency visual Transformer model that effectively reduces the computational complexity of high-resolution image processing while maintaining strong feature extraction capability by introducing local attention mechanism and sliding window mechanism. It adopts a hierarchical pyramid structure, gradually reduces the spatial resolution and increases the channel number, and adapts to the demand of multi-scale visual tasks. Compared with traditional CNN, Swin Transformer performs excellently in image classification, target detection and semantic segmentation tasks, and has good migration ability, and has become an important basic model in the field of computer vision.

[0046] Embodiment 2 An intelligent thunderstorm weather forecasting system comprises: A data processing module collects and processes initial data to obtain a multi-source fusion initial field in a unified format; A short-term prediction module constructs a short-term prediction model based on a Swin Transformer backbone network, inputs the multi-source fusion initial field and underlying surface information obtained by the data processing module, and then outputs a prediction field of thunderstorms within the next 6 hours; A short-term prediction module constructs a short-term prediction model based on a FuXi large model, inputs the multi-source fusion initial field and underlying surface information obtained by the data processing module, and then outputs the possible occurrence area and intensity distribution data of thunderstorms within the next 7-72 hours.

[0047] Embodiment 3 A computer-readable medium comprises a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the intelligent thunderstorm weather forecasting method of embodiment 1.

[0048] Application Example Select the South China region as the test area, which has typical thunderstorm characteristics with frequent thunderstorms in summer.

[0049] The intelligent thunderstorm weather forecasting method of embodiment 1 is processed as follows: Step 1, collect and process initial data to obtain a multi-source fusion initial field in a unified format; The initial data includes: 1. Satellite Remote Sensing Data: Cloud top brightness temperature (TBB) and water vapor distribution data obtained from FY-4A geostationary meteorological satellite.

[0050] 2. Radar Detection Data: Reflectivity factor, radial velocity, and vertical liquid water content (VIL) obtained from local CINRAD radar network.

[0051] 3. Ground Observation Data: Temperature, humidity, barometric pressure, wind speed, and other routine observation information provided by weather stations.

[0052] 4. Historical Lightning Data: Lightning occurrence time, location, and intensity obtained from Lightning Location System (LLS).

[0053] 5. Numerical Weather Prediction Background Field: Temperature, humidity, wind field, and convective parameter field from GRAPES model.

[0054]

[0055] Interpolate the above data in space and time to ensure that all input data have uniform resolution (e.g., 1 km x 1 km spatial resolution, 10-minute time interval).

[0056] Step 2, Construct a short-term prediction model based on Swin Transformer backbone network, input the multi-source fusion initial field and underlying surface information obtained in step 1, and then output the prediction field of thunderstorms in the next 0-6 hours; use Perform quality control, standardization processing, and spatial resampling on the results of step 1 through the feature fusion layer to obtain thunderstorm-related features such as convective available potential energy (CAPE), lifting condensation level (LCL), and radar echo intensity.

[0057] Short-term prediction results: Through the short-term prediction model, generate a thunderstorm evolution trend chart within the next 6 hours. Test results show that the model accurately captures the occurrence of thunderstorms in a certain area within the next 2 hours and predicts the thunderstorm center position and lightning intensity to be consistent with actual observations.

[0058] Evaluation indicators: Through the evaluation of test results, the following performance indicators are obtained: POD (Hit Rate): 0.85, indicating that the model better captures the area where thunderstorms occur.

[0059]

[0060] FAR (False Alarm Rate): 0.15, indicating that the model has a low false alarm rate.

[0061]

[0062] TS (Threat Score): 0.74, indicating that the model has achieved a good balance between accuracy and reliability.

[0063]

[0064] Where H represents the number of hits, F represents the number of false alarms, and M represents the number of missed reports.

[0065] Based on the test results, the model issued a thunderstorm yellow warning in the Pearl River Delta region within the next 2 hours, reminding relevant departments to take preventive measures. Actual observations show that the warning accurately captured the occurrence time, area, and intensity of the thunderstorm, helping to reduce the damage caused by thunderstorms in the local area.

[0066] Step 3: Build a short-term forecasting model based on the FuXi large model, input the multi-source fusion initial field and underlying surface information obtained in step 1, and then output the possible occurrence area and intensity distribution data of thunderstorms in the next 7-72 hours.

[0067] Short-term forecasting results: The short-term forecasting model outputs a thunderstorm distribution map for the next 24 hours, successfully predicting the high-risk areas of thunderstorm activity from the afternoon to the evening, especially in the Pearl River Delta region. The intensity and range of the thunderstorm center are very close to the actual observations.

[0068] The results are shown in Figure 3 The legend represents the lightning density, representing the number of lightning occurrences per unit time at each grid point. Figure 3 From the figure, we can see the location of lightning occurrence and the corresponding lightning density.

[0069] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the above examples, those skilled in the art can still modify or equivalently replace the specific embodiments of the present application without departing from the spirit and scope of the present application. Any modification or equivalent replacement that does not deviate from the spirit and scope of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A method of intelligent thunderstorm weather forecasting, characterized by, The method comprises the following steps: Step 1, collecting and processing initial data to obtain a multi-source fusion initial field in a unified format; Step 2, constructing a short-term prediction model based on a Swin Transformer backbone network, inputting the multi-source fusion initial field and underlying surface information obtained in step 1, and then outputting a prediction field of thunderstorms within 6 hours in the future; Step 3, constructing a short-term prediction model based on a FuXi large model, inputting the multi-source fusion initial field and underlying surface information obtained in step 1, and then outputting possible occurrence area and intensity distribution data within 6 hours in the future, i.e. a diagnostic field of thunderstorms within 6 hours in the future.

2. The intelligent thunderstorm weather forecasting method of claim 1, wherein, The initial data includes satellite remote sensing information, radar detection information, meteorological station observation information, convective parameters and lightning data.

3. The intelligent thunderstorm weather forecasting method of claim 1, wherein, The specific process of step 1 comprises: S1.1, preprocessing the initial data, wherein the preprocessing comprises denoising, missing value filling and abnormal value detection; S1.2, resampling the preprocessed data at a unified spatial and temporal resolution to obtain resampled data, and ensuring the spatio-temporal consistency between different data sources; S1.3, converting the resampled data into a unified format by standardization or normalization method.

4. The intelligent thunderstorm weather forecasting method according to claim 1 or 2 or 3, characterized in that, The specific process of constructing the short-term prediction model comprises: Step 2.1, quality control, standardization processing and spatial resampling of the results of step 1 through a feature fusion layer to obtain thunderstorm-related features; Step 2.2, based on the Swin Transformer backbone network, constructing a 2D Encoder and a 2D Decoder module to obtain a short-term prediction model, inputting the thunderstorm-related features processed in step 2.1 and underlying surface information into the short-term prediction model, and mapping the original spatial data into feature space data through the processing of the short-term prediction model to extract the spatio-temporal features of short-term thunderstorm evolution. Step 2.3, through a self-recursive iteration method, iteratively processing the results obtained in step 2.2 to output continuous lightning intensity prediction values and gradually generate the evolution trend of thunderstorms within 6 hours in the future; the function of this step is to capture the spatio-temporal evolution of thunderstorms through the self-recursive iteration method to ensure the continuity and real-time performance of the short-term prediction results; Step 2.4, setting a threshold value of thunderstorm occurrence, calculating the error between the lightning intensity obtained in step 2.3 and the real lightning intensity by using a loss function, and then evaluating the accuracy of the short-term prediction model.

5. The intelligent thunderstorm weather forecasting method of claim 4, wherein, The thunderstorm-related features in step 2.1 include convective available potential energy, lifting condensation level and radar echo intensity.

6. The intelligent thunderstorm weather forecasting method of claim 4, wherein, The loss function in step 2.4 includes a Charbonnier loss function and a BCE loss function, and the final loss value is obtained by weighting and adding the loss value obtained by the Charbonnier loss function and the loss value obtained by the BCE loss function.

7. The intelligent thunderstorm weather forecasting method according to claim 1 or 2 or 3, characterized in that, The specific process of constructing the short-term prediction model comprises: Step 3.1, based on the FuXi large model, constructing a 3D Encoder and a 3D Decoder module to obtain a short-term prediction model for capturing the spatio-temporal evolution characteristics of thunderstorms within 6 hours in the future; Step 3.2, taking the data obtained in step 1 and underlying surface information as inputs of the short-term prediction model, training the parameters of the 3D Encoder and 3D Decoder modules using multiple time unified training method, outputting the occurrence area and intensity distribution of future thunderstorms more than 6 hours in the future, and then obtaining the lightning diagnosis field.

8. An intelligent thunderstorm weather forecasting system characterized in that, Comprise: a data processing module for collecting and processing initial data to obtain multi-source fusion initial fields in a unified format; a short-term prediction module for constructing a short-term prediction model based on a Swin Transformer backbone network, inputting the multi-source fusion initial fields and underlying surface information obtained by the data processing module, and then outputting a thunderstorm prediction field within 6 hours in the future; a short-term prediction module for constructing a short-term prediction model based on a FuXi large model, inputting the multi-source fusion initial fields and underlying surface information obtained by the data processing module, and then outputting the possible occurrence area and intensity distribution data of thunderstorms more than 6 hours in the future.

9. A computer readable medium characterized by A computer device comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the intelligent thunderstorm weather prediction method of any one of claims 1 to 7.

Images