An intelligent thunderstorm weather forecasting method, system, and medium

Abstract

The application is a kind of intelligent thunderstorm weather forecasting method, system and medium, belongs to thunderstorm prediction field, in view of the problem of insufficient prediction ability of existing mode, provide a kind of intelligent thunderstorm weather forecasting method, including the following process: collect and process initial data, obtain uniform format multi-source fusion initial field;Construct a short-term forecasting model based on Swin Transformer backbone network, by inputting multi-source fusion initial field and underlying surface information, and then output the forecast field of thunderstorm within 6 hours in the future;Construct a short-term forecasting model based on FuXi large model, by inputting 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 thunderstorm more than 6 hours in the future.Through the introduction of Swin Transformer backbone network, combined with autoregressive prediction strategy and multi-task loss function, the capture accuracy and real-time performance of thunderstorm spatio-temporal evolution process are effectively improved.

Description

Technical Field

[0001] This application belongs to the field of thunderstorm prediction, and specifically relates to an intelligent thunderstorm weather forecasting method, system and medium. Background Technology

[0002] Thunderstorms, a typical type of severe convective weather, significantly impact aviation, transportation, and agriculture due to their suddenness and destructiveness. While traditional numerical weather prediction plays a crucial role in thunderstorm forecasting, its limitations in resolution, real-time performance, and ability to capture small- to medium-scale systems make it insufficient for practical needs. In recent years, the rapid development of deep learning technology, with its advantages in multimodal data processing, spatiotemporal feature extraction, and efficient prediction, has demonstrated its immense potential in the field of thunderstorm forecasting.

[0003] First, efficient fusion of multimodal data is an important application direction of deep learning in thunderstorm forecasting. Traditional thunderstorm forecasting relies on a single data source, such as radar echoes or satellite observations, while deep learning models can fuse multi-source data such as radar, satellite, ground observations, and numerical forecasts, thereby improving the comprehensiveness and accuracy of forecasts. For example, a spatiotemporal sequence forecasting model based on Convolutional Long Short-Term Memory (ConvLSTM) networks has been proposed for extrapolating radar echoes, significantly improving the short-term forecasting capability for thunderstorm development. However, these multimodal data fusion models still face challenges in practical applications, such as inconsistent data quality, difficulties in time synchronization, and high computational resource requirements, limiting their widespread application in complex weather systems.

[0004] Secondly, the advantages of deep learning in spatiotemporal feature extraction offer new insights for thunderstorm prediction. Traditional numerical weather prediction models struggle to capture the nonlinear characteristics of thunderstorms, while deep learning models (such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) can extract key patterns in thunderstorm occurrence and development through automatic feature learning. Furthermore, techniques like Generative Adversarial Networks (GANs) have been used to quantify the uncertainty of thunderstorms, providing new methods for risk assessment. However, deep learning models rely heavily on labeled data in spatiotemporal feature extraction and are susceptible to overfitting. Additionally, GANs may generate false samples during uncertainty quantification, thus affecting the reliability of prediction results.

[0005] In terms of real-time performance, deep learning models can quickly process large-scale meteorological data, enabling short-term and nowcasting of thunderstorms. For example, using the U-Net architecture for radar echo extrapolation successfully achieved high-resolution short-term thunderstorm forecasts, outperforming traditional methods in computational efficiency. Furthermore, the introduction of transfer learning techniques allows deep learning models to quickly adapt to regions with scarce data, further enhancing their generalization ability. However, real-time models may still face computational bottlenecks when processing extremely large datasets, especially during extreme weather events. Moreover, the effectiveness of transfer learning depends on the quality of the source domain data and the similarity to the target domain; if the differences are significant, model performance may degrade considerably.

[0006] Despite the many advantages that deep learning has shown in thunderstorm prediction, it still faces many challenges, such as insufficient physical interpretability of models, limitations in the quality of training data, and the need to improve its predictive ability for extreme weather events. Summary of the Invention

[0007] To address the technical problems existing in the prior art, this application provides an intelligent thunderstorm weather forecasting method, system, and medium. By introducing the Swin Transformer backbone network and combining an autoregressive prediction strategy with a multi-task loss function, the accuracy and real-time performance of capturing the spatiotemporal evolution of thunderstorms are effectively improved.

[0008] A smart thunderstorm weather forecasting method includes the following process: Step 1: Collect and process initial data to obtain a multi-source fusion initial field in a unified format; Step 2: Construct a short-term forecast model based on the Swin Transformer backbone network. By inputting the multi-source fusion initial field and underlying surface information obtained in Step 1, the forecast field of thunderstorms in the next 6 hours is output. Step 3: Construct a short-term forecast model based on the FuXi large model. By inputting the multi-source fusion initial field and underlying surface information obtained in Step 1, the model outputs the possible areas and intensity distribution data of thunderstorms in the next 6 hours or more.

[0009] This application integrates deep learning methods, utilizing multi-source data (satellite, radar, ground station, weather forecast data, underlying surface static data, etc.) and physically extracted convection parameters to construct a thunderstorm weather forecasting system that combines short-term forecasting and short-term diagnosis. By introducing a Swing Transformer backbone network and combining an autoregressive prediction strategy with a multi-task loss function (regression + classification), the accuracy and real-time performance of capturing the spatiotemporal evolution of thunderstorms are effectively improved. This application utilizes short-term thunderstorm forecasts (0-6 hours) and short-term forecast diagnosis based on a FuXi large model (6-72 hours). The short-term forecast part adopts an autoregressive form, where each time step inputs not only the forecast from the previous time step but also underlying surface information and the corresponding FuXi forecast field. The short-term diagnosis part directly uses multi-time FuXi forecasts as input to predict the thunderstorm results for the corresponding time step. Both parts are based on an Encoder-Decoder structure and use stacked Swing Transformers (or 3D Transformers) as the backbone network to achieve efficient representation of high-dimensional, multi-scale features.

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

[0011] Furthermore, the specific process of step 1 includes: S1.1, Preprocess the initial data, including noise reduction, missing value imputation and outlier detection, to ensure data quality; S1.2, resample the preprocessed data using a uniform spatial and temporal resolution to obtain resampled data, ensuring spatiotemporal consistency between different data sources; S1.3, the resampled data is converted into a unified format through standardization or normalization methods to meet the input requirements of subsequent models. For example, satellite remote sensing information and radar detection data can be converted into raster form, while meteorological station observation data and convection parameters are feature-encoded to ensure that multi-source data can be fused and processed within the same framework.

[0012] The purpose of this step is to generate a multi-source fused initial field with a consistent format, ensuring the uniformity and high quality of the input data and providing a reliable foundation for subsequent model input. Convection parameters are key indicators extracted from the raw element products from a physical perspective. They directly characterize atmospheric instability and convection potential, and can help the model quickly focus on the key physical mechanisms of thunderstorm occurrence.

[0013] Furthermore, the specific process of constructing the short-term forecast model includes: Step 2.1: The results of Step 1 are subjected to quality control, standardization, and spatial resampling through a feature fusion layer to obtain thunderstorm-related features; Step 2.2: Based on the Swin Transformer backbone network, construct 2D Encoder and 2D Decoder modules to obtain the short-term forecast model. Input the thunderstorm-related features and underlying surface information processed in Step 2.1 into the short-term forecast model. Through the processing of the short-term forecast model, the original spatial data is mapped into feature spatial data to extract the spatiotemporal features of short-term thunderstorm evolution. The purpose of this step is to capture the short-term dynamic changes of thunderstorms through the short-term forecast model and improve the accuracy of short-term forecasts.

[0014] Step 2.3: Using the autoregressive iterative method, the results obtained in Step 2.2 are iterated multiple times to output continuous lightning intensity prediction values, gradually generating the thunderstorm evolution trend for the next 6 hours. The purpose of this step is to ensure the consistency and real-time nature of short-term forecast results by using the autoregressive iterative method to capture the spatiotemporal evolution of thunderstorms. Step 2.4: Set the threshold for thunderstorm occurrence, use the loss function to calculate the error between the lightning intensity obtained in Step 2.3 and the actual lightning intensity, and then evaluate the accuracy of the short-term forecast model.

[0015] The training process employs a dual objective of regression and classification to comprehensively optimize model performance. The regression part predicts the actual number of lightning strikes, using a loss function to measure the error between the model's output lightning intensity and the actual lightning intensity, thereby improving the model's numerical prediction capability for lightning intensity. The classification part predicts whether thunderstorms will occur, using a cross-entropy loss function to perform binary classification modeling of thunderstorm events, ensuring the model can accurately distinguish between thunderstorm-affected areas and non-thunderstorm areas.

[0016] The input data for the regression task mainly comes from multi-source lightning observation data and fused feature fields, and the output is a continuous lightning intensity prediction value; the classification task, on the other hand, is based on the same input data and transforms the lightning event into a binary classification problem (occurred / not occurred) by setting a threshold for the occurrence of lightning.

[0017] Furthermore, the loss function in step 2.4 includes the Charbonnier loss function and the BCE loss function. Weights are assigned to the loss values ​​obtained through the Charbonnier loss function and the loss values ​​obtained through the BCE loss function, and then they are summed to obtain the final loss value. This final loss value is the error between the lightning intensity obtained in step 2.3 and the actual lightning intensity.

[0018] The Charbonnier loss function L is defined as follows:

[0019] in, The Charbonnier loss function is used. This refers to the predicted value. Refers to the truth value. Set to 10 -3 ; The BCE loss function is defined as follows:

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

[0021] Considering that in lightning forecasting, the model may suffer a double penalty due to the superposition of spatial location and magnitude biases, resulting in penalties for both missed and false alarms in the predicted lightning area. Therefore, we use a classification branch to focus on correctly capturing the occurrence of precipitation events, and a regression branch to quantify the confirmed precipitation events. Thus, the training loss function is also divided into two parts. The first part uses Charbonnier loss, a smooth L1 norm approximation loss function commonly used in deep learning tasks such as image restoration, denoising, and super-resolution. Compared to directly using L1 loss, Charbonnier loss is smooth at the origin, facilitating gradient calculation and avoiding the instability of gradients near zero.

[0022] Furthermore, the specific process of constructing the short-term forecast model includes: Step 3.1: Based on the FuXi large model, construct 3D Encoder and 3D Decoder modules to obtain a short-term forecast model, which is used to capture the spatiotemporal evolution characteristics of thunderstorms in the next 6 hours or more. Step 3.2 uses the data obtained in Step 1 and the underlying surface information as input to the short-term forecast model. A unified training method across multiple time periods is used to train the parameters of the 3D Encoder and 3D Decoder modules, outputting the occurrence areas and intensity distributions of future thunderstorms over the next 6 hours or more, thus obtaining the lightning diagnostic field. The purpose of this step is to model the evolution of thunderstorms over longer time scales using deep learning networks, thereby improving the spatial resolution and temporal extension of short-term forecasts.

[0023] The purpose of this step is to provide short-term forecasting models with complete historical background and short-term trend information, ensuring the continuity and accuracy of forecasts.

[0024] The parameters of the 3D Encoder and 3D Decoder modules are trained using a unified training method across multiple time intervals to ensure 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 forecasts. This step enhances the model's generalization ability through long-term training, providing high-quality short-term thunderstorm prediction information for weather forecasting.

[0025] A smart thunderstorm weather forecasting system includes: The data processing module collects and processes the initial data to obtain a multi-source fusion initial field in a unified format; The short-term forecast module constructs a short-term forecast model based on the Swing Transformer backbone network. It inputs the multi-source fusion initial field and underlying surface information obtained from the input data processing module, and then outputs the forecast field of thunderstorms for the next 6 hours. The short-term forecast module constructs a short-term forecast model based on the FuXi large model. By inputting the multi-source fusion initial field and underlying surface information obtained from the data processing module, it outputs the possible areas and intensity distribution data of thunderstorms in the next 6 hours or more.

[0026] A computer-readable medium includes a memory and one or more processors, wherein executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the intelligent thunderstorm weather forecasting method.

[0027] Compared with the prior art, the beneficial effects of this application are as follows: (1) Traditional thunderstorm forecasting mainly relies on numerical models and human experience, which has limited spatial resolution and real-time performance. However, this scheme integrates multi-source satellite, radar, ground observation, FuXi large model forecast, underlying static information and convection parameters, making the spatiotemporal characterization of thunderstorm forecasts more comprehensive and detailed, so as to make the prediction results more accurate. The integrated design of short-term and short-term forecasts breaks through the limitations of the single time scale of traditional thunderstorm forecasts. Through short-term autoregressive forecasts within 6 hours and diagnostic forecasts based on FuXi large models, more detailed forecast levels and time span coverage are achieved.

[0028] (2) In terms of model design, traditional CNN or basic regression methods are difficult to adapt to strong convective weather with varying spatiotemporal scales, such as thunderstorms. This application adopts the Swin Transformer backbone network, which uses local attention mechanism and sliding window strategy, and deeply integrates with satellite, radar, ground observation and underlying static data to enhance the ability to capture and represent the complex spatiotemporal characteristics of thunderstorms, and can better capture multi-scale dynamic features.

[0029] (3) In terms of forecast scope and form, the scheme distinguishes between short-term forecasts within 6 hours and short-term diagnoses beyond 6 hours, enabling rapid updates and control of medium-term trends, significantly enhancing the precision and real-time performance. During training, a dual loss function of regression and classification is adopted, combined with indicators such as POD, FAR, and TS for comprehensive evaluation. It not only focuses on whether thunderstorms occur (classified), but also uses regression branches to quantify the number of thunderstorms. It is supplemented by multi-task training combining Charbonnier and cross-entropy loss, and comprehensively measured by meteorological indicators such as POD, FAR, and TS, which can effectively balance the hit rate and false alarm rate, further improving the practicality and accuracy of thunderstorm forecasts. Attached Figure Description

[0030] Figure 1 This is a flowchart of Example 1; Figure 2 This is a schematic diagram of the structure of a short-term forecasting model; Figure 3 The result is shown in the example diagram. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] Example 1 A smart thunderstorm weather forecasting method, flowchart as follows: Figure 1 As shown, the process includes the following: Step 1: Collect and process initial data to obtain a multi-source fused initial field in a unified format; the initial data includes satellite remote sensing information, radar detection information, meteorological station observation information, convection parameters, and lightning data. The specific process includes: S1.1, Preprocess the initial data, including noise reduction, missing value imputation, and outlier detection to ensure data quality; remove outliers through data cleaning and perform spatial and temporal matching on data from different sources; S1.2, resample the preprocessed data using a uniform spatial and temporal resolution to obtain resampled data, ensuring spatiotemporal consistency between different data sources; S1.3, the resampled data is converted into a unified format through standardization or normalization methods to meet the input requirements of subsequent models. For example, satellite remote sensing information and radar detection data can be converted into raster form, while meteorological station observation data and convection parameters are feature-encoded to ensure that multi-source data can be fused and processed within the same framework.

[0033] Step 2 involves constructing a short-term forecast model based on the Swin Transformer backbone network. The Swin Transformer consists of stacked Swin Transformer blocks; each time a block is passed, the image is downsampled by a factor of two. Each block employs a sliding window attention mechanism. By inputting the multi-source fusion initial field and underlying surface information obtained in Step 1, the model outputs the forecast field for thunderstorms over the next 6 hours. Specifically, this process includes: enhancing the model's ability to represent complex meteorological data through multi-scale feature extraction; optimizing the prediction accuracy of key areas by combining an attention mechanism; and capturing the dynamic features of thunderstorm evolution through a temporal convolution module to achieve efficient short-term forecasting.

[0034] The structure for constructing a short-term forecast model is as follows: Figure 2 As shown, Figure 2 (a) shows the overall architecture of the SwinTransformer, which consists of multiple stacked SwinTransformer modules. Figure 2 (b) shows the structure of the two SwinTransformer modules.

[0035] Step 2.1: The results of Step 1 are subjected to quality control, standardization, and spatial resampling through a feature fusion layer to obtain thunderstorm-related features; Step 2.2: Based on the Swin Transformer backbone network, construct 2D Encoder and 2D Decoder modules to obtain the short-term forecast model. Input the thunderstorm-related features and underlying surface information processed in Step 2.1 into the short-term forecast model to extract the spatiotemporal features of short-term thunderstorm evolution. The purpose of this step is to capture the short-term dynamic changes of thunderstorms through deep learning models and improve the accuracy of short-term forecasts.

[0036] Step 2.3 involves iterating the results obtained in Step 2.2 multiple times through autoregression to output continuous lightning intensity prediction values ​​and gradually generate the thunderstorm evolution trend for the next 6 hours. The purpose of this step is to enhance the model's ability to capture the spatiotemporal evolution of thunderstorms through the autoregression strategy, ensuring the consistency and real-time nature of short-term forecast results. Step 2.4: Using the loss function, calculate the error between the lightning intensity obtained in Step 2.3 and the actual lightning intensity, set the threshold for thunderstorm occurrence, and then evaluate the accuracy of the short-term forecast model. The loss function includes the Charbonnier loss function and the BCE loss function, wherein the Charbonnier loss function is defined as:

[0037] in, The Charbonnier loss function is used. This refers to the predicted value. Refers to the truth value. Set to 10 -3 ; The BCE loss function 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 between 0 and 1 by the sigmoid function). The advantage of BCE loss is that it can directly output probability predictions, making the forecast results more interpretable and practical. The loss functions in step 2.4 include the Charbonnier loss function and the BCE loss function. Weights are assigned to the loss values ​​obtained by the Charbonnier loss function and the loss values ​​obtained by the BCE loss function, and then summed to obtain the final loss value. This final loss value is the error between the lightning intensity obtained in step 2.3 and the actual lightning intensity. The final loss is the weighted sum of the two losses, with the weight of BCE loss set to 0.5 and the weight of Charbonnier loss set to 1.

[0039] The model prediction performance in step 2 is evaluated using commonly used meteorological indicators such as POD (Pick-up Rate), FAR (False Alarm Rate), and TS (Threat Score). These indicators comprehensively measure the model's performance in thunderstorm forecasting scenarios, ensuring that the model accurately predicts real thunderstorm events while minimizing false alarms, thus achieving a balance between accuracy and reliability.

[0040] Step 3: Construct a short-term forecast model based on the FuXi large-scale model. By inputting the multi-source fused initial field and underlying surface information obtained in Step 1, the model outputs data on the possible areas and intensity distribution of thunderstorms over the next 7-72 ​​hours. The specific process includes: Step 3.1: Based on the FuXi large model, construct 3D Encoder and 3D Decoder modules to obtain a short-term forecast model, which is used to capture the spatiotemporal evolution characteristics of thunderstorms in the next 7-72 ​​hours; Step 3.2 uses the data obtained in Step 1 and the underlying surface information as input to the short-term forecast model. A unified training method across multiple time periods is employed, with the Adam algorithm used to optimize the loss function. This process trains the parameters of the 3D Encoder and 3D Decoder modules, outputting the occurrence areas and intensity distributions of future thunderstorms over a period of 6 hours or more, thus obtaining the lightning diagnostic field. The purpose of this step is to model the evolution of thunderstorms over longer time scales using deep learning networks, thereby improving the spatial resolution and temporal extension of short-term forecasts.

[0041] This invention integrates multi-source observation and forecast data through deep learning, combining short-term thunderstorm forecasting with short-term forecast diagnosis based on the FuXi large model. It utilizes the Swing Transformer to construct a national thunderstorm weather forecasting model that combines high accuracy and timeliness. At the data level, it integrates information from satellites, radar, ground stations, and convection parameters, supplemented by static underlying surface data, to more comprehensively characterize the thunderstorm occurrence mechanism. At the model level, it introduces an autoregressive strategy to achieve short-term forecasts of 0-6 hours, and directly inputs multi-time FuXi forecasts to complete 7-72 ​​hour diagnosis. At the training and evaluation level, through a multi-task mechanism using Charbonnier loss and cross-entropy loss, combined with indicators such as POD, FAR, and TS, it ensures a balance between the model's sensitivity to capturing thunderstorm events and reducing false alarms, thus providing a more reliable reference for thunderstorm forecasting.

[0042] Since lightning is a highly spatiotemporally variable element and the influence of the initial field is relatively short-term, a combination of short-term lightning forecast (0-6 hours) and model short-term forecast result diagnosis (6-72 hours) is adopted.

[0043] The short-term lightning forecast model uses multimodal data and gridded lightning data at times t-1 and t0 as initial information, and then employs an autoregressive deep learning model. In each autoregressive iteration, in addition to the output of the previous time step, the input includes underlying surface information and the FuXi forecast field corresponding to the output of the next time step. After 6 iterations, the lightning forecast field for 0-6 hours is obtained.

[0044] The lightning diagnostic model based on FuXi short-term forecast directly uses the subsequent 66 time periods as time channels to construct a 3D input, plus underlying surface information, to directly predict the lightning field corresponding to the 66 time periods.

[0045] Both models above employ an Encoder-Decoder architecture, with a stacked SwinTransformer backbone. The short-term lightning forecast model and the lightning diagnostic model based on FuXi short-term forecasts both use stacked SwinTransformers (the diagnostic model can be improved to a 3D Transformer). The Swin Transformer is an efficient visual Transformer model that effectively reduces the computational complexity of high-resolution image processing by introducing local attention and sliding window mechanisms, while maintaining strong feature extraction capabilities. It adopts a hierarchical pyramid structure, gradually reducing spatial resolution and increasing the number of channels to adapt to multi-scale visual tasks. Compared to traditional CNNs, the Swin Transformer performs superiorly in tasks such as image classification, object detection, and semantic segmentation, and possesses good transferability, making it an important foundational model in the field of computer vision.

[0046] Example 2 A smart thunderstorm weather forecasting system includes: The data processing module collects and processes the initial data to obtain a multi-source fusion initial field in a unified format; The short-term forecast module constructs a short-term forecast model based on the Swing Transformer backbone network. It inputs the multi-source fusion initial field and underlying surface information obtained from the input data processing module, and then outputs the forecast field of thunderstorms for the next 6 hours. The short-term forecast module constructs a short-term forecast model based on the FuXi large model. By inputting the multi-source fusion initial field and underlying surface information obtained from the data processing module, it outputs the possible areas and intensity distribution data of thunderstorms in the next 7-72 ​​hours or more.

[0047] Example 3 A computer-readable medium includes a memory and one or more processors, wherein executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the intelligent thunderstorm weather forecasting method described in Embodiment 1.

[0048] Application examples South China was selected as the test area, as the region experiences frequent thunderstorms in summer and exhibits typical thunderstorm characteristics.

[0049] The processing procedure of the intelligent thunderstorm weather forecasting method in Example 1 is 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 acquired from the FY-4A geostationary meteorological satellite.

[0050] 2. Radar detection data: reflectivity factor, radial velocity, and vertical liquid water content (VIL) obtained from the local CINRAD radar network.

[0051] 3. Ground observation data: Routine observation information such as temperature, humidity, air pressure, and wind speed provided by meteorological stations.

[0052] 4. Historical lightning data: The time, location, and intensity of lightning strikes obtained from the Lightning Location System (LLS).

[0053] 5. Numerical weather forecast background field: Temperature, humidity, wind field and convection parameter field from the GRAPES model.

[0054]

[0055] The above data is interpolated spatially and temporally to ensure that all input data have a uniform resolution (e.g., 1km×1km spatial resolution, 10-minute time interval).

[0056] Step 2: Construct a short-term forecast model based on the Swin Transformer backbone network. By inputting the multi-source fused initial field and underlying surface information obtained in Step 1, the model outputs the forecast field for thunderstorms in the next 0-6 hours. The results of step 1 are subjected to quality control, standardization, and spatial resampling through a feature fusion layer to obtain thunderstorm-related features, such as convective effective potential energy (CAPE), condensation height (LCL), and radar echo intensity.

[0057] Short-term forecast results: A short-term forecast model was used to generate a trend map of thunderstorm evolution over the next 6 hours. Test results show that the model accurately captured the occurrence of a thunderstorm in a certain area within the next 2 hours, and the predicted location of the thunderstorm center and lightning intensity are largely consistent with actual observations.

[0058] Evaluation indicators: The following performance metrics were obtained through evaluation of the test results: POD (hit rate): 0.85, indicating that the model captured the area where the thunderstorm occurred relatively well.

[0059]

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

[0061]

[0062] Threat Score (TS): 0.74, indicating that the model achieves 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 alarms.

[0065] Based on the test results, the model issued a yellow thunderstorm warning for the Pearl River Delta region within the next two hours, reminding relevant departments to take preventative measures. Actual observations showed that the warning accurately captured the time, location, and intensity of the thunderstorms, helping to reduce damage caused by them.

[0066] Step 3: Construct a short-term forecast model based on the FuXi large model. By inputting the multi-source fusion initial field and underlying surface information obtained in Step 1, the model outputs the possible areas and intensity distribution data of thunderstorms in the next 7-72 ​​hours.

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

[0068] The results are as follows Figure 3 As shown in the figure, the legend represents lightning density, which represents the number of lightning strikes per unit time at each grid point. Figure 3 The location of the lightning strike and the corresponding lightning density can be seen from the data.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still modify or make equivalent substitutions to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A smart thunderstorm weather forecasting method, characterized in that, The process includes the following: Step 1: Collect and process initial data to obtain a multi-source fusion initial field in a unified format; Step 2: Construct a short-term forecast model based on the Swin Transformer backbone network. By inputting the multi-source fused initial field and underlying surface information obtained in Step 1, the model outputs the forecast field for thunderstorms over the next 6 hours. The specific process includes: Step 2.1: The results of Step 1 are subjected to quality control, standardization, and spatial resampling through a feature fusion layer to obtain thunderstorm-related features; Step 2.2: Based on the Swin Transformer backbone network, construct 2D Encoder and 2D Decoder modules to obtain the short-term forecast model. Input the thunderstorm-related features and underlying surface information processed in Step 2.1 into the short-term forecast model. Through the processing of the short-term forecast model, the original spatial data is mapped into feature spatial data to extract the spatiotemporal features of short-term thunderstorm evolution. Step 2.3: Using the autoregressive iterative method, the results obtained in Step 2.2 are iterated multiple times to output continuous lightning intensity prediction values, gradually generating the thunderstorm evolution trend for the next 6 hours. The purpose of this step is to ensure the consistency and real-time nature of short-term forecast results by using the autoregressive iterative method to capture the spatiotemporal evolution of thunderstorms. Step 2.4: Set a threshold for thunderstorm occurrence, and use a loss function to calculate the error between the lightning intensity obtained in Step 2.3 and the actual lightning intensity, thereby evaluating the accuracy of the short-term forecast model; the loss function in Step 2.4 includes the Charbonnier loss function and the BCE loss function. Weights are assigned to the loss values ​​obtained through the Charbonnier loss function and the loss values ​​obtained through the BCE loss function, and they are summed to obtain the final loss value; Step 3: Construct a short-term forecast model based on the FuXi large model. By inputting the multi-source fusion initial field and underlying surface information obtained in Step 1, the model outputs data on the possible areas and intensity distributions of thunderstorms in the next 6 hours or more, i.e., the diagnostic field for thunderstorms in the next 6 hours or more. The specific process includes: Step 3.1: Based on the FuXi large model, construct 3D Encoder and 3D Decoder modules to obtain a short-term forecast model, which is used to capture the spatiotemporal evolution characteristics of thunderstorms in the next 6 hours or more. Step 3.2: Using the data obtained in Step 1 and the underlying surface information as input to the short-term forecast model, the parameters of the 3D Encoder and 3D Decoder modules are trained using a unified training method with multiple time intervals. The model outputs the occurrence area and intensity distribution of future thunderstorms for more than 6 hours in the future, thereby obtaining the lightning diagnostic field.

2. The intelligent thunderstorm weather forecasting method according to claim 1, characterized in that, The initial data includes satellite remote sensing information, radar detection information, meteorological station observation information, convection parameters, and lightning data.

3. The intelligent thunderstorm weather forecasting method according to claim 1, characterized in that, Step 1 includes the following specific steps: S1.1, Preprocess the initial data, the preprocessing including denoising, missing value imputation and outlier detection; S1.2, resample the preprocessed data using a uniform spatial and temporal resolution to obtain resampled data, ensuring spatiotemporal consistency between different data sources; S1.3, convert the resampled data into a uniform format through standardization or normalization methods.

4. The intelligent thunderstorm weather forecasting method according to claim 1, characterized in that, The thunderstorm-related features in step 2.1 include convective effective potential energy, lifting condensation height, and radar echo intensity.

5. An intelligent thunderstorm weather forecasting system, characterized in that, include: The data processing module collects and processes the initial data to obtain a multi-source fusion initial field in a unified format; The short-term forecast module constructs a short-term forecast model based on the Swing Transformer backbone network. It takes the multi-source fused initial field and underlying surface information obtained from the input data processing module as input, and then outputs the forecast field for thunderstorms over the next 6 hours. Specifically, it includes: The feature fusion layer performs quality control, standardization, and spatial resampling on the results of the data processing module to obtain thunderstorm-related features. Based on the Swin Transformer backbone network, 2D Encoder and 2D Decoder modules are constructed to obtain a short-term forecast model. The processed thunderstorm-related features and underlying surface information are input into the short-term forecast model. Through the processing of the short-term forecast model, the original spatial data is mapped into feature spatial data to extract the spatiotemporal features of short-term thunderstorm evolution. The results are iterated multiple times using the autoregressive iterative method to output continuous lightning intensity prediction values ​​and gradually generate the thunderstorm evolution trend for the next 6 hours. The purpose of this step is to ensure the consistency and real-time nature of the short-term forecast results by using the autoregressive iterative method to capture the spatiotemporal evolution of thunderstorms. A threshold for thunderstorm occurrence is set, and a loss function is used to calculate the error between the obtained lightning intensity and the actual lightning intensity, thereby evaluating the accuracy of the short-term forecast model. The loss function includes the Charbonnier loss function and the BCE loss function. Weights are assigned to the loss values ​​obtained through the Charbonnier loss function and the loss values ​​obtained through the BCE loss function, and then they are summed to obtain the final loss value. The short-term forecast module constructs a short-term forecast model based on the FuXi large model. It inputs multi-source fused initial field and underlying surface information obtained from the data processing module, and then outputs data on the possible areas and intensity distribution of thunderstorms in the next 6 hours or more. The specific process includes: Based on the FuXi large model, 3D Encoder and 3D Decoder modules are constructed to obtain a short-term forecast model, which is used to capture the spatiotemporal evolution characteristics of thunderstorms in the next 6 hours or more. The obtained data and underlying surface information are used as input to the short-term forecast model. The parameters of the 3D Encoder and 3D Decoder modules are trained using a unified training method at multiple time intervals. The model outputs the occurrence area and intensity distribution of future thunderstorms more than 6 hours in the future, thereby obtaining the lightning diagnostic field.

6. A computer-readable medium, characterized in that, The system includes 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 forecasting method according to any one of claims 1 to 4.

Images