A lightning proximity warning method based on Swin-Unet

Abstract

This invention relates to the field of lightning monitoring and early warning technology, and in particular to a lightning imminent warning method based on Swin-Unet. The method includes: acquiring and processing multi-source observation data of the forecast area to construct a dataset; inputting the dataset into the deep learning model Swin-Unet for training, selecting a model that meets a preset accuracy as the lightning imminent warning model; acquiring current multi-source observation data of the forecast area and inputting it into the lightning imminent warning model, outputting the future lightning distribution, and completing the lightning imminent warning for the forecast area. This invention can achieve lightning imminent warning by fusing multi-source observation data and considering physical environmental constraints.

Description

Technical Field

[0001] This invention relates to the field of lightning monitoring and early warning technology, and in particular to a lightning proximity early warning method based on Swin-Unet. Background Technology

[0002] Lightning is a powerful electrical discharge phenomenon in thunderstorm systems, often occurring alongside severe weather events such as hail, strong winds, and heavy rainfall. It poses a serious threat to people's lives and property, making timely and accurate lightning warnings and forecasts of significant practical and social value. However, lightning activity is a small-to-medium scale weather phenomenon, characterized by strong locality and rapid evolution. Accurate lightning prediction remains a challenge for current meteorological operations and a prominent problem that urgently needs to be addressed.

[0003] Currently, lightning nowcasting is mainly based on multi-source observational data from weather radar, meteorological satellites, lightning locators, and ground electric field meters, both domestically and internationally. The main methods can be broadly categorized into two types: one is to extract characteristic quantities from the observational data and set discrimination thresholds based on statistical results to achieve lightning nowcasting warnings; the other is to perform linear or nonlinear extrapolation based on radar or satellite data, combined with specific algorithms, such as fuzzy logic, to predict the timing and location of lightning strikes. While these methods can achieve good results, their operational applications suffer from high false alarm and false alarm rates, indicating significant limitations. First, the location and intensity of lightning change rapidly during the evolution of thunderstorm systems. Second, the conditions for the occurrence and development of thunderstorms and the range of characteristic thresholds vary considerably across different seasons and terrains, and many valuable characteristic physical quantities remain to be discovered. Furthermore, the initiation mechanism of lightning is complex, and existing extrapolation algorithms lack the ability to predict the evolution of convection generation and dissipation, thus limiting their applicability.

[0004] In recent years, the field of artificial intelligence has developed rapidly. AI technologies, represented by deep learning models, have become an effective way to improve lightning warning effectiveness due to their outstanding ability to solve complex problems and perform nonlinear modeling. Some scholars have already used deep learning to fuse numerical models and multi-source observation data to achieve short-term lightning forecasting. Zhou et al. used a deep semantic segmentation model to extract the spatiotemporal occurrence and development characteristics of lightning from multi-source observation data, achieving effective 0-1 hour lightning forecasting for the first time, with good forecasting results in the initial stage of thunderstorms in South China. Li et al., based on the State Grid's wide-area lightning and ground flash monitoring data and the Himawari 8 / 9 satellite cloud imagery, used a convolutional-gated recurrent unit network with an attention mechanism to conduct nowcasting of the lightning strike area and frequency in Central China during the warm season. Their established deep learning lightning forecasting model can effectively predict the development trend of lightning strike area and frequency in organized thunderstorms.

[0005] However, current research and applications are still in their early stages. To further improve lightning early warning services and lightning disaster prevention and mitigation capabilities, this invention provides a lightning proximity early warning method based on Swin-Unet. Summary of the Invention

[0006] The purpose of this invention is to provide a lightning proximity warning method based on Swin-Unet. By leveraging the powerful feature extraction capabilities of deep learning, it deeply mines effective thunderstorm information from multi-source observation data such as weather radar, lightning locators, and atmospheric electric field meters. Based on a deep learning semantic segmentation model, it introduces geographical information such as terrain and underlying surface that affect the evolution of thunderstorm systems and lightning activity, and constructs a lightning proximity warning model with higher accuracy, thereby improving the accuracy of lightning proximity warnings.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A lightning proximity warning method based on Swin-Unet includes:

[0009] Acquire and process multi-source observation data and physical environment data of the forecast area to construct a dataset;

[0010] The dataset is input into the deep learning model Swin-Unet for training, and the model that meets the preset accuracy is selected as the lightning imminent warning model.

[0011] The current multi-source observation data and physical environment data of the forecast area are obtained and input into the lightning imminent warning model, which outputs the future lightning distribution and completes the lightning imminent warning for the forecast area.

[0012] Optionally, multi-source observation data and physical environment data of the forecast area are acquired and processed to construct a dataset including:

[0013] The radar three-dimensional network mosaic data of the forecast area is obtained, and radar feature parameters are extracted from the radar three-dimensional network mosaic data to obtain radar feature grid data.

[0014] Acquire three-dimensional lightning detection data of the forecast area, and extract lightning activity feature parameters from the three-dimensional lightning detection data to obtain lightning feature grid data;

[0015] Acquire atmospheric electric field instrument detection data for the forecast area, extract atmospheric electric field characteristic parameters from the atmospheric electric field instrument detection data, and obtain atmospheric electric field characteristic grid data;

[0016] The terrain height and underlying surface data of the forecast area are obtained and normalized to obtain environmental feature grid data.

[0017] The radar feature grid data, lightning feature grid data, atmospheric electric field feature grid data, and environmental feature grid data are constructed into a dataset according to the observation time.

[0018] Optionally, acquiring radar 3D network mosaic data of the forecast area, and extracting radar feature parameters from the radar 3D network mosaic data, including acquiring radar feature grid data:

[0019] Obtain radar 3D network mosaic data for the forecast area;

[0020] The reflectivity factor REF, combined reflectivity factor CR, vertical integral liquid water content VIL, and echo top height TOP of several height layers within the forecast area are extracted from the radar three-dimensional network mosaic data to form radar feature grid data with a preset resolution.

[0021] Optionally, acquiring three-dimensional lightning detection data of the forecast area and extracting lightning activity feature parameters from the three-dimensional lightning detection data, obtaining lightning feature grid data includes:

[0022] Acquire three-dimensional lightning detection data for the forecast area;

[0023] The frequency information of cloud-to-cloud lightning, ground-to-ground lightning, and total lightning occurring within the forecast area in time period A before the radar time is extracted from the three-dimensional lightning detection data. By setting several distance thresholds, the presence or absence of lightning activity within different distance ranges is determined and marked, forming lightning feature grid data with a preset resolution.

[0024] Optionally, atmospheric electric field meter data for the forecast area is acquired, and atmospheric electric field characteristic parameters are extracted from the atmospheric electric field meter data. Acquiring atmospheric electric field characteristic grid data includes:

[0025] Acquire atmospheric electric field data for the forecast area;

[0026] The atmospheric electric field intensity data within the forecast area during the B-period prior to the radar time is extracted from the atmospheric electric field instrument detection data, and statistical analysis is performed to obtain the maximum value, median value, difference between the maximum and minimum values, and jump value of the atmospheric electric field intensity, forming atmospheric electric field feature grid data with a preset resolution.

[0027] Optionally, the topographic height and underlying surface data of the forecast area are obtained and normalized to obtain environmental feature grid data, including:

[0028] Obtain digital elevation model data of the forecast area, extract terrain height data of the forecast area from the digital elevation model data, and perform minimum-maximum normalization processing on the terrain height data to form terrain height grid data with a preset resolution.

[0029] Land use data of the forecast area is obtained, underlying surface data of the forecast area is extracted from the land use data, and the underlying surface data is normalized according to several land use types to form underlying surface grid data with a preset resolution.

[0030] Optionally, the dataset is input into the deep learning model Swin-Unet for training, and a model that meets a preset accuracy is selected as the lightning proximity warning model, including:

[0031] The dataset is divided into a training set and a validation set;

[0032] Based on the training set, the deep learning model Swin-Unet is trained by taking radar feature grid data, lightning feature grid data, atmospheric electric field feature grid data and environmental feature grid data from the past period a as input and lightning activity prediction results from the future period b as output.

[0033] Based on the validation set, the trained deep learning model Swin-Unet is validated and evaluated, and the trained model whose evaluation results reach the preset value is selected as the lightning proximity warning model.

[0034] Optionally, selecting the trained model whose evaluation results reach a preset value as the lightning proximity warning model includes:

[0035] The critical success index was used as the evaluation metric for the trained deep learning model Swin-Unet, and the model with the highest critical success index value was selected as the lightning proximity warning model.

[0036] The beneficial effects of this invention are as follows:

[0037] This invention, based on multi-source observation data from radar, lightning, and atmospheric electric fields, and combined with physical environmental characteristics affecting the formation and evolution of thunderstorm systems such as terrain height and underlying surface, employs the deep learning model Swin-Unet to train and optimize a lightning nowcasting warning model, thereby achieving professional lightning nowcasting warning services. This method effectively solves the problems existing in current technologies, improves the accuracy of lightning nowcasting warnings, and can be applied to existing operational radar, lightning, and atmospheric electric field detection networks, effectively enhancing the professional service level of lightning warnings. This invention improves the accuracy of lightning nowcasting warnings and can be applied to existing professional meteorological service systems. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of a lightning proximity warning method based on Swin-Unet according to an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the lightning proximity warning model structure according to an embodiment of the present invention;

[0041] Figure 3 This is a comparison chart of the lightning proximity warning model forecast results and the actual lightning situation for a large-scale lightning event according to an embodiment of the present invention;

[0042] Figure 4 This is a comparison chart of the lightning proximity warning model forecast results and the actual lightning situation for a local lightning event according to an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Example 1:

[0046] This embodiment provides a lightning proximity warning method based on Swin-Unet, including:

[0047] Acquire and process multi-source observation data and physical environment data of the forecast area to construct a dataset;

[0048] The dataset is input into the deep learning model Swin-Unet for training, and the model that meets the preset accuracy is selected as the lightning imminent warning model.

[0049] The current multi-source observation data and physical environment data of the forecast area are obtained and input into the lightning imminent warning model, which outputs the future lightning distribution and completes the lightning imminent warning for the forecast area.

[0050] Furthermore, multi-source observation data and physical environment data of the forecast area are acquired and processed to construct a dataset including:

[0051] The radar three-dimensional network mosaic data of the forecast area is obtained, and radar feature parameters are extracted from the radar three-dimensional network mosaic data to obtain radar feature grid data.

[0052] Acquire three-dimensional lightning detection data of the forecast area, and extract lightning activity feature parameters from the three-dimensional lightning detection data to obtain lightning feature grid data;

[0053] Acquire atmospheric electric field instrument detection data for the forecast area, extract atmospheric electric field characteristic parameters from the atmospheric electric field instrument detection data, and obtain atmospheric electric field characteristic grid data;

[0054] The terrain height and underlying surface data of the forecast area are obtained and normalized to obtain environmental feature grid data.

[0055] The radar feature grid data, lightning feature grid data, atmospheric electric field feature grid data, and environmental feature grid data are constructed into a dataset according to the observation time.

[0056] Further, radar 3D network mosaic data of the forecast area is acquired, and radar feature parameters are extracted from the radar 3D network mosaic data. The acquisition of radar feature grid data includes:

[0057] Obtain radar 3D network mosaic data for the forecast area;

[0058] The reflectivity factor REF, combined reflectivity factor CR, vertical integral liquid water content VIL, and echo top height TOP of several height layers within the forecast area are extracted from the radar three-dimensional network mosaic data to form radar feature grid data with a preset resolution.

[0059] Further, three-dimensional lightning detection data of the forecast area is acquired, and lightning activity feature parameters are extracted from the three-dimensional lightning detection data. The acquisition of lightning feature grid data includes:

[0060] Acquire three-dimensional lightning detection data for the forecast area;

[0061] The frequency information of cloud-to-cloud lightning, ground-to-ground lightning, and total lightning occurring within the forecast area in time period A before the radar time is extracted from the three-dimensional lightning detection data. By setting several distance thresholds, the presence or absence of lightning activity within different distance ranges is determined and marked, forming lightning feature grid data with a preset resolution.

[0062] Further, atmospheric electric field meter data for the forecast area is acquired, and atmospheric electric field characteristic parameters are extracted from the atmospheric electric field meter data. Acquiring atmospheric electric field characteristic grid data includes:

[0063] Acquire atmospheric electric field data for the forecast area;

[0064] The atmospheric electric field intensity data within the forecast area during the B-period prior to the radar time is extracted from the atmospheric electric field instrument detection data, and statistical analysis is performed to obtain the maximum value, median value, difference between the maximum and minimum values, and jump value of the atmospheric electric field intensity, forming atmospheric electric field feature grid data with a preset resolution.

[0065] Further, the topographic height and underlying surface data of the forecast area are obtained and normalized to obtain environmental feature grid data, including:

[0066] Obtain digital elevation model data of the forecast area, extract terrain height data of the forecast area from the digital elevation model data, and perform minimum-maximum normalization processing on the terrain height data to form terrain height grid data with a preset resolution.

[0067] Land use data of the forecast area is obtained, underlying surface data of the forecast area is extracted from the land use data, and the underlying surface data is normalized according to several land use types to form underlying surface grid data with a preset resolution.

[0068] Furthermore, the dataset is input into the deep learning model Swin-Unet for training, and models that meet the preset accuracy are selected as lightning proximity warning models, including:

[0069] The dataset is divided into a training set and a validation set;

[0070] Based on the training set, the deep learning model Swin-Unet is trained by taking radar feature grid data, lightning feature grid data, atmospheric electric field feature grid data and environmental feature grid data from the past period a as input and lightning activity prediction results from the future period b as output.

[0071] Based on the validation set, the trained deep learning model Swin-Unet is validated and evaluated, and the trained model whose evaluation results reach the preset value is selected as the lightning proximity warning model.

[0072] Furthermore, selecting the trained model whose evaluation results reach a preset value as the lightning proximity warning model includes:

[0073] The critical success index was used as the evaluation metric for the trained deep learning model Swin-Unet, and the model with the highest critical success index value was selected as the lightning proximity warning model.

[0074] Specifically, this embodiment, based on multi-source observation data such as radar, lightning, and atmospheric electric fields, and combined with physical environmental characteristics affecting the formation and evolution of thunderstorm systems, such as terrain height and underlying surface, employs the deep learning model Swin-Unet to train and optimize a lightning proximity warning model, thereby achieving professional lightning proximity warning services. This embodiment effectively solves the problems existing in the prior art, improves the accuracy of lightning proximity warnings, and can be applied to existing operational radar, lightning, and atmospheric electric field detection networks, effectively enhancing the professional service level of lightning warnings. This embodiment improves the accuracy of lightning proximity warnings and can be applied to existing professional meteorological service systems. Figure 1 As shown, the specific steps include:

[0075] S1: Obtain radar 3D network mosaic data, extract radar feature parameters of the forecast area from it, and obtain radar feature grid data;

[0076] The radar 3D network mosaic data is acquired, and radar characteristic parameters such as reflectivity factor REF, combined reflectivity factor CR, vertical integral liquid water content VIL, and echo top height TOP are extracted from different altitude layers h (h=1,3,5,6,7,8,9km) within the forecast area (specified latitude and longitude range) to form radar characteristic grid data with a resolution of 1km.

[0077] S2: Acquire three-dimensional lightning detection data, extract lightning activity characteristic parameters of the forecast area from it, and perform gridding processing to obtain lightning characteristic grid data;

[0078] Acquire three-dimensional lightning detection data, extract frequency information of cloud-to-cloud lightning, ground-to-ground lightning, and total lightning occurring within 6 minutes prior to the radar time in the forecast area (specified latitude and longitude range), and set different distance thresholds R (R=5, 10, 20km) to determine whether there is lightning activity within different distance ranges. If there is lightning, it is 1; otherwise, it is 0, forming lightning feature grid data with a resolution of 1km.

[0079] S3: Acquire atmospheric electric field instrument detection data, extract atmospheric electric field characteristic parameters of the forecast area from it, and obtain atmospheric electric field characteristic grid data by adopting the nearest principle;

[0080] S31: Acquire atmospheric electric field instrument detection data, and extract atmospheric electric field intensity data within the 6 minutes prior to the radar time, with a time resolution of 1 second;

[0081] S32: Based on the atmospheric electric field intensity data obtained in step S31, the maximum value of the atmospheric electric field intensity EF_Max is statistically obtained;

[0082] S33: Based on the atmospheric electric field intensity data obtained in step S31, the median value of the atmospheric electric field intensity EF_Men is statistically obtained;

[0083] S34: Based on the atmospheric electric field intensity data obtained in step S31, the difference between the maximum and minimum atmospheric electric field intensity EF_Del is statistically obtained.

[0084] S35: Based on the atmospheric electric field intensity data obtained in step S31, determine whether there is a jump in the electric field intensity, that is, whether the electric field intensity is greater than 3σ, where σ is the standard deviation of the atmospheric electric field intensity in the 6-minute period, and statistically obtain the atmospheric electric field intensity jump value EF_Jmp.

[0085] S36: Based on the atmospheric electric field intensity characteristic data obtained in steps S31 to S35, the nearest atmospheric electric field instrument data is used as the standard to form atmospheric electric field characteristic grid data with a resolution of 1km.

[0086] S4: Obtain the topographic height and underlying surface data of the forecast area, and perform normalization processing to obtain environmental feature grid data;

[0087] S41: Read the China Digital Elevation Model (DEM) data (spatial resolution of 1km), extract the terrain height data within the forecast area (specified latitude and longitude range), and perform min-max normalization on the terrain height data to form 1km resolution terrain height grid data Ter_Hgt;

[0088] S42: Read China Land Use (LUCC) data (spatial resolution of 1km), extract underlying surface data within the forecast area (specified latitude and longitude range), and normalize the underlying surface data according to 6 categories, including water area, cultivated land, forest land, grassland, residential land and unused land (0.0, 0.2, 0.4, 0.6, 0.8, 1.0), to form underlying surface grid data Land_Use with a resolution of 1km.

[0089] S5: Based on the various grid data obtained in steps S1 to S4, construct a dataset and divide it into a training dataset and a validation dataset as needed;

[0090] Based on the various grid point data obtained in steps S1 to S4, a dataset is constructed; the temporal continuity of the observation data is maintained, and the dataset is divided into a training dataset and a validation dataset in a 3:1 ratio.

[0091] S6: Based on the deep learning model Swin-Unet, a lightning near-term warning model is constructed, which takes radar, lightning, atmospheric electric field observation data and environmental characteristic parameters of the past 1 hour as input and the probability of lightning activity in the next 2 hours as output.

[0092] S61: Constructing a lightning proximity warning model based on the deep learning model Swin-Unet, such as Figure 2 As shown, the model takes as input radar feature grid data from the past hour (including reflectivity factors REF, combined reflectivity factors CR, vertical integrated liquid water content VIL, and echo top height TOP at different altitudes), lightning feature grid data (including the frequency of cloud-to-cloud lightning, ground-to-ground lightning, and total lightning), atmospheric electric field feature grid data (including the maximum, median, and difference values ​​of electric field intensity, and whether jumps exist), and environmental feature grid data (including terrain height and underlying surface). The model outputs the lightning activity prediction results for the next 2 hours, specifically the probability distribution of whether lightning will occur, with a spatial resolution of 1 km and a temporal resolution of 6 minutes. By leveraging the feature extraction and forecasting capabilities of the Swin-Unet deep learning model, multi-source data and environmental features are combined to achieve high-precision nowcasting and early warning of lightning activity.

[0093] S62: Train the lightning nowcasting model using the training dataset partitioned in step S5. After training, evaluate the model using the validation dataset partitioned in step S5, primarily using the Critical Success Index (CSI). CSI is an indicator used to evaluate forecast accuracy, considering both missed and false alarms. Its value ranges from 0 to 1, with higher values ​​indicating higher forecast accuracy. The model's performance is evaluated by calculating the CSI values ​​between the model's predictions and actual observations.

[0094] S7: Based on the dataset obtained in step S5, train and evaluate the lightning proximity warning model, and obtain the best model through hyperparameter optimization;

[0095] By adjusting hyperparameters such as learning rate, batch size, and weights, step S6 involves batch training and evaluation of the model. Ultimately, the model with the highest CSI value is selected as the optimal lightning proximity warning model to ensure its accuracy and reliability in practical applications.

[0096] S8: Based on the optimal lightning proximity warning model obtained in step S7, conduct professional lightning proximity warning services and output lightning warning results.

[0097] Example 2:

[0098] The following section uses a city as an example to apply and verify a lightning proximity warning method based on Swin-Unet, specifically including:

[0099] Step S1: Obtain radar 3D network mosaic data of a certain province from April to September 2020 to 2022, and extract radar characteristic parameters (reflectivity factor REF, combined reflectivity factor CR, vertical integral liquid water content VIL and echo top height TOP) for a certain city area to obtain radar characteristic grid data;

[0100] Step S2: Obtain observation data from a three-dimensional lightning detection network in a certain province from April to September 2020 to 2022, extract lightning activity characteristic parameters (frequency of cloud lightning, ground lightning and total lightning) for a certain city area, and perform gridding processing to obtain lightning characteristic gridded data.

[0101] Step S3: Obtain the observation data of the atmospheric electric field meter detection network of a certain province from April to September 2020 to 2022, extract the atmospheric electric field characteristic parameters (maximum value, median value, difference value of electric field intensity and whether there is a jump) of a certain city area, and obtain the atmospheric electric field characteristic grid data by adopting the principle of proximity (taking the atmospheric electric field meter data closest to the grid point as the standard).

[0102] Step S4: Obtain the topographic elevation and underlying surface data of a certain city area, and perform normalization processing to obtain environmental feature grid data;

[0103] Step S5: Based on the various grid data obtained in steps S1 to S4, construct a dataset and divide it into a training dataset and a validation dataset as needed.

[0104] Step S6: Based on the deep learning model Swin-Unet, construct a lightning near-term warning model that takes radar, lightning, atmospheric electric field observation data and environmental characteristic parameters from the past 1 hour as input and the probability of lightning activity in the next 2 hours as output.

[0105] Step S7: Based on the dataset obtained in Step S5, train and evaluate the lightning proximity warning model, and obtain the best model through hyperparameter optimization;

[0106] Step S8: Based on the optimal lightning proximity warning model obtained in Step S7, conduct professional lightning proximity warning services and output lightning warning results.

[0107] Furthermore, in step S1, the specific steps for obtaining radar feature grid data are as follows:

[0108] Read the radar 3D network mosaic data of a certain province, and extract radar characteristic parameters such as reflectivity factor REF, combined reflectivity factor CR, vertical integral liquid water content VIL, and echo top height TOP at different height layers h (h=1,3,5,6,7,8,9km) within a certain city area (224km×224km) to form radar characteristic grid data with a resolution of 1km and 224×224 grid points.

[0109] Furthermore, in step S2, the specific steps for obtaining lightning feature grid data are as follows:

[0110] Data from a three-dimensional lightning detection network in a province is read, and the frequency information of cloud-to-cloud lightning, ground-to-ground lightning, and total lightning (total lightning = cloud-to-cloud lightning + ground-to-ground lightning) occurring within a 224km×224km area of ​​a city is extracted within the 6 minutes prior to the radar time. Different distance thresholds R (R=5, 10, 20km) are set to determine whether there is lightning activity within different distance ranges. The value is 1 if there is lightning and 0 otherwise, resulting in lightning feature grid data with a resolution of 1km and 224×224 grid points.

[0111] Furthermore, in step S3, the specific steps for obtaining the atmospheric electric field characteristic grid data are as follows:

[0112] S31: Read the data from the atmospheric electric field meter detection network of a certain province, and extract the atmospheric electric field intensity data within the 6 minutes before the radar time, with a time resolution of 1 second;

[0113] S32: Based on the atmospheric electric field intensity data obtained in step S31, the maximum value of the atmospheric electric field intensity EF_Max is statistically obtained;

[0114] S33: Based on the atmospheric electric field intensity data obtained in step S31, the median value of the atmospheric electric field intensity EF_Men is statistically obtained;

[0115] S34: Based on the atmospheric electric field intensity data obtained in step S31, the difference between the maximum and minimum atmospheric electric field intensity EF_Del is statistically obtained.

[0116] S35: Based on the atmospheric electric field intensity data obtained in step S31, determine whether there is a jump in the electric field intensity, that is, whether the electric field intensity is greater than 3σ, where σ is the standard deviation of the atmospheric electric field intensity within a 6-minute period, and statistically obtain the atmospheric electric field intensity jump value EF_Jmp.

[0117] S36: Based on the atmospheric electric field intensity characteristic data obtained in steps S31 to S35, the nearest principle is adopted, that is, the data of the atmospheric electric field instrument closest to the grid point is used to form atmospheric electric field characteristic grid data with a resolution of 1km and a grid number of 224×224.

[0118] Furthermore, in step S4, the specific steps for obtaining environmental feature grid data are as follows:

[0119] S41: Read the China Digital Elevation Model (DEM) data (spatial resolution of 1km), extract the terrain height data within a certain city area (224km×224km), and perform min-max normalization on the terrain height data to form terrain height grid data Ter_Hgt with a resolution of 1km and 224×224 grid points;

[0120] S42: Read China Land Use (LUCC) data (spatial resolution of 1km), extract underlying surface data within a certain city area (224km×224km), and normalize the underlying surface data according to 6 categories (water area, cultivated land, forest land, grassland, residential land and unused land) (0.0, 0.2, 0.4, 0.6, 0.8, 1.0) to form underlying surface grid data Land_Use with a resolution of 1km and 224×224 grid points.

[0121] Furthermore, in step S5, the specific steps for obtaining the training dataset and the validation dataset are as follows:

[0122] S51: Based on the gridded data of multi-source observations (radar, lightning, atmospheric electric field) and physical environment (topographic height and underlying surface) of a certain city area (224km×224km) obtained in step S4 from April to September 2020 to 2022, construct a dataset;

[0123] S52: Maintain the temporal continuity of the observation data and divide the dataset obtained in step S51 into a training dataset and a validation dataset in a 3:1 ratio.

[0124] Furthermore, in step S6, the specific steps for constructing the lightning proximity warning model based on the deep learning model Swin-Unet are as follows:

[0125] S61: A lightning proximity warning model is constructed based on the deep learning model Swin-Unet. The model takes radar feature grid data (reflectivity factor REF, combined reflectivity factor CR, vertical integral liquid water content VIL, and echo top height TOP) from the past 1 hour as input data, lightning feature grid data (frequency of cloud lightning, ground lightning, and total lightning), atmospheric electric field feature grid data (maximum electric field strength EF_Max, median electric field strength EF_Men, difference electric field strength EF_Del, and whether there is a jump electric field EF_Jmp) and environmental feature grid data (terrain height Ter_Hgt and underlying surface area Land_Use) as output data; and grid data of the probability of lightning activity in the next 2 hours as output data (spatial resolution of 1 km, temporal resolution of 6 min).

[0126] Specifically, the input data for the lightning imminent warning model consists of multi-channel gridded data, with each channel corresponding to a specific feature parameter. The model uses the Swin-Unet architecture to fuse and process this multi-source data, extracting feature information related to lightning activity. The Swin-Unet model combines the efficient feature extraction capabilities of the Swin Transformer with the encoder-decoder structure of U-Net, effectively processing high-resolution gridded data and generating predictions of lightning activity for the next two hours.

[0127] S62: Train the lightning nowcasting model using the training dataset partitioned in step S5. The training dataset includes multi-source observation data and environmental feature data from the past hour, as well as actual observations of lightning activity for the next two hours. The goal of model training is to minimize the error between the predicted results and the actual observations. After training, evaluate the performance of the trained model using the validation dataset partitioned in step S5. The main evaluation metric is the Critical Success Index (CSI), calculated as follows:

[0128] CSI = TP / (TP + FA + MS);

[0129] Where TP represents the number of grid points where both prediction and observation of lightning occurred; FA represents the number of grid points where lightning was predicted but not observed (missed reports); and MS represents the number of grid points where lightning was observed but not predicted (false alarms). The higher the CSI value (ranging from 0 to 1), the higher the prediction accuracy of the model.

[0130] Furthermore, in step S7, the specific steps for obtaining the optimal model through hyperparameter optimization are as follows:

[0131] By adjusting model hyperparameters such as learning rate, batch size, and weight, the model training and evaluation in step S6 are carried out in batches to continuously optimize the model's performance. Finally, the model with the highest CSI value is selected as the best lightning proximity warning model.

[0132] The optimal model achieved a high CSI value, indicating that the model can effectively predict lightning activity and has high practicality and reliability.

[0133] In this embodiment, the deep learning model selected is Swin-Unet, a novel segmentation network based on the Swin Transformer and incorporating features of the U-Net network. Based on multi-source observation data such as radar, lightning, and atmospheric electric fields, and incorporating physical features that potentially influence lightning activity, including terrain height and underlying surface information, a lightning proximity warning model based on Swin-Unet was constructed through continuous hyperparameter adjustment during model training and evaluation. This resulted in a lightning proximity warning method based on deep learning technology, thereby improving the accuracy of lightning proximity warnings and enhancing the professional service level of lightning warnings.

[0134] To verify the practical effectiveness of this method, the lightning forecasting performance of the trained lightning nowcasting model was evaluated using two thunderstorm events with different lightning activity characteristics. The lightning nowcasting model used multi-source observation data from the previous hour and fixed environmental parameters as input to predict the lightning distribution for the next two hours, with a temporal resolution of 6 minutes and a spatial resolution of 1 km. Figure 3 , Figure 4 This is a comparison chart of the forecast results (LFM) and actual lightning situation (OBS) of the two thunderstorm processes mentioned above (the forecast results and actual lightning situation in the chart are at 12-minute intervals).

[0135] like Figure 3 As shown, this was a thunderstorm process accompanied by significant lightning, with widespread lightning activity. The forecast results of the lightning nowcasting model were basically consistent with the actual lightning situation, especially the lightning strike areas in the first hour, indicating a high degree of agreement. The overall forecast performance was good. Meanwhile, for thunderstorm processes with only localized lightning activity, such as... Figure 4 As shown, the lightning proximity warning model also demonstrated good performance. The model successfully predicted two lightning strike areas in the first hour, and the lightning activity ended in the second hour. Although there were some false alarms in the northern region, the false alarm area was not large and decreased rapidly as the forecast duration increased. Overall, the forecast performance was still satisfactory.

[0136] This embodiment focuses on two thunderstorm events with different lightning activity characteristics. Using a dataset comprised of multi-source observations (radar, lightning, atmospheric electric field) and environmental feature (topographic height and underlying surface) gridded data from a city area between April and September 2020 and 2022, a lightning nowcasting model based on the deep learning model Swin-Unet was trained and optimized. The actual effectiveness of the lightning nowcasting model was evaluated using real-world lightning data as a comparison. The following conclusions were drawn:

[0137] (1) Compared with the traditional observation-based lightning proximity warning method, this method fully leverages the outstanding ability of deep learning technology to solve complex problems and nonlinear modeling, and has the characteristics of fast computation, flexible adjustment and high accuracy.

[0138] (2) This method introduces multi-source observation data that characterizes lightning activity, such as radar, lightning, and atmospheric electric field. At the same time, it considers the physical environment characteristics such as terrain and underlying surface that have a potential impact on the generation and dissipation of thunderstorm systems. This effectively overcomes the problem of incomplete data utilization in conventional lightning warning methods, improves the problem of too many false alarms when relying solely on radar observation for lightning warning, further improves the accuracy of lightning imminent warning, and enhances the professional service level of lightning warning.

[0139] The above conclusions demonstrate that this method can interpolate multi-source observation data from weather radar, lightning locators, and atmospheric electric field meters into gridded data. Combined with gridded terrain height, underlying surface, and other environmental characteristic parameters, a lightning proximity warning model is constructed using the deep learning model Swin-Unet. This achieves lightning proximity warning based on deep learning technology, integrating multi-source observation data, and considering physical environmental constraints.

[0140] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for lightning nowcasting based on Swin-Unet, characterized in that, The method comprises the following steps: acquiring and processing multi-source observation data and physical environment data of a forecast area to construct a data set; inputting the data set into a deep learning model Swin-Unet for training, and selecting a model meeting a preset accuracy as a lightning near-warnings model; acquiring current multi-source observation data and physical environment data of the forecast area and inputting the data into the lightning near-warnings model to output a future lightning distribution, thereby completing lightning near-warnings of the forecast area; acquiring and processing multi-source observation data and physical environment data of a forecast area to construct a data set comprises: acquiring radar three-dimensional networking mosaic data of the forecast area, and extracting radar characteristic parameters from the radar three-dimensional networking mosaic data to acquire radar characteristic grid data; acquiring three-dimensional lightning detection data of the forecast area, and extracting lightning activity characteristic parameters from the three-dimensional lightning detection data to acquire lightning characteristic grid data; acquiring atmospheric electric field instrument detection data of the forecast area, and extracting atmospheric electric field characteristic parameters from the atmospheric electric field instrument detection data to acquire atmospheric electric field characteristic grid data; acquiring terrain height and underlying surface data of the forecast area, and performing normalization processing to acquire environmental characteristic grid data; constructing the radar characteristic grid data, the lightning characteristic grid data, the atmospheric electric field characteristic grid data and the environmental characteristic grid data into a data set according to observation time; acquiring radar three-dimensional networking mosaic data of the forecast area, and extracting radar characteristic parameters from the radar three-dimensional networking mosaic data to acquire radar characteristic grid data comprises: acquiring radar three-dimensional networking mosaic data of the forecast area; extracting reflectivity factor REF, combined reflectivity factor CR, vertical integrated liquid water content VIL and echo top height TOP of a plurality of height layers in the forecast area from the radar three-dimensional networking mosaic data to form radar characteristic grid data of a preset resolution; acquiring three-dimensional lightning detection data of the forecast area, and extracting lightning activity characteristic parameters from the three-dimensional lightning detection data to acquire lightning characteristic grid data comprises: acquiring three-dimensional lightning detection data of the forecast area; extracting cloud flash, ground flash and total flash frequency information occurring in an A time period before a radar time of the forecast area from the three-dimensional lightning detection data, marking whether there is lightning activity in different distance ranges by setting a plurality of distance thresholds, and forming lightning characteristic grid data of a preset resolution; acquiring atmospheric electric field instrument detection data of the forecast area, and extracting atmospheric electric field characteristic parameters from the atmospheric electric field instrument detection data to acquire atmospheric electric field characteristic grid data comprises: acquiring atmospheric electric field instrument detection data of the forecast area; extracting atmospheric electric field intensity data of the forecast area in a B time period before the radar time of the forecast area from the atmospheric electric field instrument detection data, and performing statistics to acquire a maximum value, a median value, a difference between the maximum value and a minimum value, and a jump value of the atmospheric electric field intensity, thereby forming atmospheric electric field characteristic grid data of a preset resolution.

2. The Swin-Unet-based lightning nowcasting method according to claim 1, wherein, acquiring terrain height and underlying surface data of the forecast area, and performing normalization processing to acquire environmental characteristic grid data comprises: Obtaining digital elevation model data of the forecast area, extracting terrain height data in the forecast area from the digital elevation model data, and performing minimum-maximum normalization processing on the terrain height data to form terrain height grid data of a preset resolution; Obtaining land use data of the forecast area, extracting underlying surface data in the forecast area from the land use data, and performing normalization processing on the underlying surface data according to a plurality of land use types to form underlying surface grid data of a preset resolution.

3. The Swin-Unet based lightning nowcasting method of claim 1, wherein, Inputting the data set into a deep learning model Swin-Unet for training, and selecting a model meeting a preset accuracy as the lightning nowcasting model includes: Dividing the data set into a training set and a validation set; Based on the training set, using radar feature grid data, lightning feature grid data, atmospheric electric field feature grid data, and environmental feature grid data in a past a period as input, and using lightning activity prediction results in a future b period as output, training a deep learning model Swin-Unet; Based on the validation set, verifying and evaluating the trained deep learning model Swin-Unet, and selecting a trained model with an evaluation result reaching a preset value as the lightning nowcasting model.

4. The Swin-Unet-based lightning nowcasting method of claim 3, wherein, Selecting a trained model with an evaluation result reaching a preset value as the lightning nowcasting model includes: Using a critical success index as an evaluation index of the trained deep learning model Swin-Unet, and selecting a model with the highest critical success index value as the lightning nowcasting model.

Images