Strong thunderstorm potential risk and intensity prediction method based on artificial neural network

Abstract

The invention is suitable for the technical field of thunderstorm prediction, and provides a strong thunderstorm potential risk and intensity prediction method based on an artificial neural network, and the method comprises the steps: obtaining the ground-to-ground lightning data of a target region, determining the time distribution characteristics of a thunderstorm event, and recognizing the time valley of the thunderstorm event; collecting multi-source meteorological data in a target area, and screening meteorological prediction factor data; performing space-time alignment processing on the ground-to-ground lightning data and the meteorological prediction factor data, and constructing a grid unit day-by-day sample set based on the aligned data; constructing a probability classification model of a strong thunderstorm event based on the grid unit day-by-day sample set; building a regression neural network model based on the grid units which are judged to be strong thunderstorm high risks by a probability classification model; and a collaborative prediction result is output, a predicted value of a geometric mean value of the ground-to-ground flash frequency and the lightning current amplitude is synchronously output, and a risk assessment basis of classification discrimination and numerical prediction capability is provided for extreme weather early warning and power grid disaster prevention and reduction through double-order modeling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of thunderstorm prediction, and particularly relates to a strong thunderstorm potential risk and intensity prediction method based on an artificial neural network. BACKGROUND

[0002] As a kind of strong convective weather phenomenon, thunderstorm is sudden, short in duration, and often accompanied by lightning, lightning, gale, heavy rain and other extreme weather conditions, which poses a serious threat to the safe and stable operation of power grid. During the process of strong thunderstorm, the local lightning density is high, and a strong thunderstorm process can contribute more than 50% of the lightning density in the lightning center area in a year, and the continuous trip accidents of multi-loop lines caused by it will significantly increase the systematic risk of power grid.

[0003] Currently, thunderstorm potential prediction mainly relies on lightning positioning, atmospheric electric field monitoring and radar remote sensing, and is realized by using Logistic regression method. However, due to the essential contradiction between the nonlinear evolution characteristics of thunderstorm system and the linear assumption of linear regression model, the complex interaction between meteorological elements cannot be effectively characterized by traditional statistical methods, so that the model generalization ability is limited, the probability of strong thunderstorm event is low, and the false negative rate is significantly high. In addition, the current technical solution fails to build a multi-task joint learning framework, and lacks adaptive optimization for the non-equilibrium distribution characteristics of extreme weather samples, so that the cooperative mechanism of potential prediction and intensity evaluation is missing, which limits the single-dimensional prediction output of the differential protection strategy of power grid, and cannot realize dynamic resource scheduling based on the spatial distribution characteristics of lightning intensity.

[0004] In view of this, a strong thunderstorm potential risk and intensity prediction method based on an artificial neural network is proposed. SUMMARY

[0005] The present application provides a strong thunderstorm potential risk and intensity prediction method based on an artificial neural network, which is used to solve the problem of low probability of strong thunderstorm event and high false negative rate.

[0006] The present application provides a strong thunderstorm potential risk and intensity prediction method based on an artificial neural network, which comprises:

[0007] Obtain the ground flash data of the target area, determine the time distribution characteristics of thunderstorm events based on the ground flash data, and identify the time trough of thunderstorm events, which is a persistent period when the occurrence frequency of thunderstorm events is significantly lower than the climate mean value.

[0008] Collect multi-source meteorological data in the target area, screen the meteorological data associated with the formation of thunderstorm events as thunderstorm prediction factors, and obtain meteorological prediction factor data.

[0009] spatially and temporally align the cloud-to-ground lightning data and the meteorological predictor data, construct a grid-cell daily sample set containing cloud-to-ground lightning features and meteorological features based on the aligned data, and divide the grid-cell daily sample set into a training set and a test set in a time sequence;

[0010] construct a probability classification model of strong thunderstorm events based on the grid-cell daily sample set, for predicting the probability of strong thunderstorm occurrence in each grid cell within a preset time;

[0011] determine the grid cells at high risk of strong thunderstorm based on the probability classification model, and construct a regression neural network model for predicting the cloud-to-ground lightning frequency and the geometric mean of lightning current amplitude of each grid cell within a preset time period;

[0012] output the collaborative prediction results, including the potential risk discrimination results of strong thunderstorm in each grid cell within a preset time period; and simultaneously output the predicted values of the cloud-to-ground lightning frequency and the geometric mean of lightning current amplitude for the grid cells determined to have strong thunderstorm.

[0013] Further, the time valley of the thunderstorm event is identified, including:

[0014] divide equal-length time intervals based on the cloud-to-ground lightning data collaboratively monitored by multiple sites, and calculate the cumulative cloud-to-ground lightning frequency in each interval;

[0015] select the continuous time period with the lowest cumulative frequency from the cumulative cloud-to-ground lightning frequency as the time valley, and the duration of the time valley matches the characteristics of the low-occurrence stage of thunderstorm events in a climatological statistical period.

[0016] Further, the multi-source meteorological data in the target area is collected, and the meteorological data associated with the formation of thunderstorm events is selected as thunderstorm predictors to obtain meteorological predictor data, including:

[0017] Based on the physical trigger mechanism of thunderstorm events, meteorological parameters with dynamic response characteristics are selected from multi-dimensional meteorological observation data as thunderstorm predictors; the dynamic change trend of the meteorological parameters is associated with the evolution law of the intensity of thunderstorm events.

[0018] Further, the spatial and temporal alignment of the cloud-to-ground lightning data and the meteorological predictor data includes:

[0019] Define the daily time starting point based on the periodic characteristics of thunderstorm events, and divide the target area into grid cells consistent with the spatial resolution of meteorological parameters;

[0020] Statistically analyze the daily cloud-to-ground lightning frequency and lightning current intensity parameters in each grid cell, and classify the samples according to the intensity standard of thunderstorm events.

[0021] Further, the statistics of the daily lightning frequency and the lightning current intensity parameter in each grid cell, and the sample categories are divided according to the thunderstorm event intensity standard, including:

[0022] Based on the climate characteristic parameters of the target area and the grid cell area, the lightning frequency determination threshold is calculated, the climate characteristic parameters include the annual average thunderstorm day number and the lightning event type ratio parameter, and the calculation of the determination threshold meets the preset intensity level division requirement.

[0023] Further, the grid cell daily sample set includes time identification, lightning frequency statistics, lightning current intensity parameter, strong thunderstorm event discrimination identification and standardized meteorological prediction factor; the division of the training set and the test set meets the time continuity constraint, and the training set covers continuous multi-year observation data.

[0024] Further, the probability classification model of strong thunderstorm event occurrence is constructed based on the grid cell daily sample set, including:

[0025] Based on the training set data of the grid cell daily sample set, the joint distribution law of lightning features and meteorological features is learned by using a neural network architecture;

[0026] The model parameters are optimized by time series cross-validation to generate a strong thunderstorm event probability prediction model.

[0027] Further, the strong thunderstorm event probability prediction model is generated, including:

[0028] The probability value identification of strong thunderstorm event occurrence in each grid cell in a preset period and the corresponding spatio-temporal position coding are output, and the risk level is divided by a probability threshold to generate a grid-based thunderstorm potential discrimination result.

[0029] Further, the regression neural network model is constructed based on the grid cell determined as a strong thunderstorm high-risk by the probability classification model, including:

[0030] Based on the high-risk grid cell sample set screened out by the probability classification model, the dynamic meteorological parameters and lightning historical features in the high-risk grid cell sample set are input, the nonlinear mapping relationship between lightning frequency and lightning current intensity is learned synchronously by using a multi-task neural network architecture, and a quantitative thunderstorm intensity prediction model is generated.

[0031] Further, the output of the collaborative prediction result includes:

[0032] The strong thunderstorm potential discrimination result of each grid cell and the intensity prediction parameter are spatio-temporally associated and mapped to generate a multi-dimensional structured data table, the data table contains time dimension identification, spatial grid coding, event occurrence probability identification and quantitative intensity evaluation parameter, and a standardized thunderstorm risk decision data interface is generated.

[0033] From the above technical scheme can be seen, the present application has the following advantages:

[0034] The present application unifies the cloud-to-ground lightning data and the meteorological data to the same size of the space grid, and implements the sliding time window processing based on the thunderstorm event period, solves the problem of the inaccurate spatio-temporal matching of the meteorological elements and the lightning activity in the traditional method, realizes the accurate alignment of the regional meteorological parameters and the lightning event characteristics, and effectively captures the nonlinear evolution law of the thunderstorm system; based on the artificial neural network, a strong thunderstorm event occurrence probability classification model is constructed, the essential contradiction between the traditional regression method and the nonlinear evolution characteristics of the thunderstorm system is solved, and the probability of measuring the strong thunderstorm event is effectively improved; in the potential prediction stage, after the potential strong thunderstorm grid is screened out through the thunderstorm event occurrence probability classification model, the neural network regression model is further used to quantitatively predict the cloud-to-ground lightning frequency and the geometric mean of the lightning current amplitude, realizes the synchronous prediction of the probability and intensity parameters of the strong thunderstorm event, and provides quantitative decision basis for the differentiated inspection of the power transmission line, the optimization layout of the lightning protection device and the like. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is an embodiment flowchart of the strong thunderstorm potential risk and intensity prediction method based on the artificial neural network in the present application;

[0036] Figure 2 It is a confusion evidence chart output by the strong thunderstorm event occurrence probability classification model in the present application;

[0037] Figure 3 It is a prediction result chart of the cloud-to-ground lightning frequency by the thunderstorm intensity regression prediction model in the present application;

[0038] Figure 4 It is a prediction result chart of the geometric mean of the lightning current amplitude of the strong thunderstorm by the thunderstorm intensity regression prediction model in the present application;

[0039] Figure 5 It is a relative error distribution chart of the cloud-to-ground lightning frequency prediction result by the thunderstorm intensity regression prediction model in the present application;

[0040] Figure 6 It is a relative error distribution of the geometric mean of the lightning current amplitude prediction result by the thunderstorm intensity regression prediction model in the present application. DETAILED DESCRIPTION

[0041] The terms "first", "second", "third", "fourth" etc. (if any) in the description and drawings of the present application are used to distinguish similar objects, and do not necessarily have to be described in a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "correspond to" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or apparatus that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to such processes, methods, products or apparatuses.

[0042] Embodiment one

[0043] The method implemented in this embodiment can be implemented in a system, and can be implemented in a server or a terminal, and the specific implementation is not limited. From the perspective of system implementation, the method for predicting the potential risk and intensity of severe thunderstorms based on artificial neural networks in the present application will be introduced. Please refer to Figure 1 The method provided by the embodiment of the present application includes the following steps:

[0044] S11. Obtain the ground flash data of the target area, determine the time distribution characteristics of the thunderstorm event based on the ground flash data, and identify the time valley bottom of the thunderstorm event, wherein the time valley bottom is a persistent period during which the occurrence frequency of the thunderstorm event is significantly lower than the climatic mean value.

[0045] Based on multi-site collaborative monitoring, ground flash observation data of a certain area is obtained, wherein the ground flash data is obtained from a lightning location system covering a target geographical range, and the geographical range is 20°N-23°N in latitude and 107°E-109°E in longitude. Invalid records are removed through joint positioning of at least three detection sites, and valid ground flash event data containing time stamp, geographical coordinates and lightning current amplitude are retained. The original ground flash records are subjected to time consistency verification, spatial positioning accuracy screening and abnormal amplitude removal to form a standardized ground flash data set.

[0046] In this embodiment, identifying the time valley bottom of the thunderstorm event includes the following:

[0047] 1. Divide the ground flash data based on multi-site collaborative monitoring into equal-length time intervals, and calculate the cumulative ground flash frequency of each interval.

[0048] 2. Select the continuous period with the lowest cumulative frequency from the cumulative ground flash frequency as the time valley bottom, and the duration of the time valley bottom matches the characteristics of the low-occurrence stage of the thunderstorm event in the climatic state statistical period.

[0049] Specifically, the whole day is divided into equal length time intervals based on the starting point of each day in local time, and the cumulative frequency of cloud-to-ground lightning events in each interval is counted; the continuous time period with the lowest cumulative frequency is identified by the sliding window algorithm, and the low-frequency persistence characteristics of the continuous time period are verified by combining the daily variation law of thunderstorm activity in the climatic state statistical period; the time bottom period selected is compared and analyzed with the historical climatic state thunderstorm event low-frequency stage to ensure that the time bottom and the stable period of the regional atmospheric stratification are statistically consistent in the occurrence period and duration.

[0050] S12. Collecting multi-source meteorological data in the target area, selecting meteorological data related to the formation of thunderstorm events as thunderstorm prediction factors to obtain meteorological prediction factor data;

[0051] In this embodiment, obtaining meteorological prediction factor data includes:

[0052] Based on the physical trigger mechanism of thunderstorm formation, meteorological parameters with dynamic response characteristics are selected from multi-dimensional meteorological observation data as thunderstorm prediction factors; the dynamic change trend of the meteorological parameters is associated with the evolution law of the intensity of thunderstorm events.

[0053] Specifically, the meteorological data is derived from global climate reanalysis data sets, covering high spatial resolution gridded data in the target area, including thermodynamic parameters such as convective available potential energy and atmospheric stability index, water vapor condition parameters such as near-surface dew point temperature and relative humidity vertical cumulative amount, dynamic lifting parameters such as low-level wind field components, vertical wind shear, vorticity and divergence field, and thermal vertical gradient parameters.

[0054] By analyzing the nonlinear coupling relationship between the time evolution law of each parameter and the intensity of thunderstorm activity, dynamic characteristic parameters significantly associated with the triggering, development and dissipation process of thunderstorms are selected, and a standardized method is used to eliminate dimensional differences, generating a thunderstorm prediction factor data set matching the spatial and temporal resolution of the cloud-to-ground lightning data. The standardization method includes mean-variance normalization processing of the original meteorological parameters to ensure the consistency of the input features and adapt to the training needs of the neural network model.

[0055] S13. Temporally and spatially align the cloud-to-ground lightning data and meteorological prediction factor data, construct a grid cell daily sample set containing cloud-to-ground lightning features and meteorological features based on the aligned data, and divide it into a training set and a test set in time sequence;

[0056] In this embodiment, the temporally and spatially aligned cloud-to-ground lightning data and meteorological prediction factor data include the following:

[0057] 1. Define the daily time starting point based on the periodic characteristics of thunderstorm events, and divide the target area into grid cells consistent with the spatial resolution of meteorological parameters;

[0058] 2. Statistics of the daily lightning frequency and lightning current intensity parameters in each grid cell, and sample categories are divided according to the thunderstorm event intensity standard.

[0059] The original lightning data is statistically analyzed in a time window sliding with a 24-hour cycle, taking the bottom of the thunderstorm activity time as the starting point of the daily time, to ensure complete coverage of the thunderstorm event. The target area is divided into grid cells consistent with the spatial resolution of meteorological parameters (such as 0.25°x0.25°), and the daily lightning frequency and geometric mean of lightning current amplitude in each grid cell are counted to eliminate spatial scale differences.

[0060] Among them, the sample categories are divided according to the thunderstorm event intensity standard, which is based on the calculation of lightning frequency determination threshold according to the climate characteristic parameters of the target area and the area of the grid cell. The climate characteristic parameters include the annual average thunderstorm day number and the lightning event type ratio parameter, and the calculation of the determination threshold meets the requirements of the pre-set intensity level division.

[0061] According to the climate characteristic parameters of the target area (including the annual average thunderstorm day number and the lightning type ratio parameter) and the area of the grid cell, the lightning frequency determination threshold is dynamically calculated, and the strong thunderstorm event sample category is divided, which meets the requirements of the national standard for the division of thunderstorm intensity level.

[0062] Taking Fangchenggang City in Guangxi as an example, the calculation method of setting grid lightning frequency threshold to determine strong thunderstorm event is as follows:

[0063]

[0064] In the formula, 7.98 times / (km 2 . year) is the minimum standard of the annual average lightning density of the strong lightning area according to the intensity classification standard of GB / T 40621-2021, considering only the first return stroke; 715 is the grid area, unit: km 2 ; 1.8984 is the ratio parameter of total lightning and first return stroke in Fangchenggang City, Guangxi; 85 is the annual average thunderstorm day number in Fangchenggang City, Guangxi.

[0065] In addition, the daily sample set of the grid cell includes time identification, lightning frequency statistics, lightning current intensity parameters, strong thunderstorm event discrimination identification and standardized meteorological prediction factors; the division of the training set and the test set meets the time continuity constraint, and the training set covers continuous multi-year observation data.

[0066] The sample feature database includes: date (date), lightning frequency (N CG ), geometric mean of lightning current amplitude (I CG ), binary classification label of strong thunderstorm event (Q storm ) and 24 standardized thunderstorm prediction factors (X1,..., X 24 ):

[0067] D = {date, N CG ,I CG ,Q storm ,X1,...,X 24}

[0068] For example, in Fangchenggang City, Guangxi, the training set is the daily observation data of 35 geographical grids in Fangchenggang City, Guangxi, China for 5 consecutive years, and the test set is the daily observation data of 35 geographical grids in Fangchenggang City, Guangxi, China for 1 consecutive year. The time series of the training set and the test set have no overlap.

[0069] S14. Based on the grid unit daily sample set, a probability classification model of strong thunderstorm event is constructed, which is used to predict the probability of strong thunderstorm occurrence in each grid unit within a predetermined time;

[0070] 1. Based on the training set data of the grid unit daily sample set, the joint distribution law of the lightning feature and the meteorological feature is learned by using a neural network architecture;

[0071] 2. The model parameters are optimized by time series cross-validation to generate a strong thunderstorm event probability prediction model.

[0072] Wherein, the strong thunderstorm event probability prediction model comprises:

[0073] Output the probability value of each grid unit occurring strong thunderstorm event within a predetermined period and the corresponding spatio-temporal position code, and divide the risk level by a probability threshold to generate a grid-based thunderstorm potential discrimination result.

[0074] Through time series cross-validation strategy to optimize network hyperparameters, ensure the model's ability to capture time-dependent features; use dynamic sample weight adjustment mechanism to alleviate the class imbalance problem caused by low proportion of strong thunderstorm event samples in training data, and improve the model's sensitivity to minority class events. Output the probability value (0-1 interval) of each grid unit occurring strong thunderstorm event within a predetermined period, and bind the spatio-temporal position code; dynamically set the probability threshold according to business needs, map the probability value to multiple risk levels such as low risk, medium risk and high risk, and generate a grid-based thunderstorm potential discrimination map.

[0075] S15. Based on the probability classification model, the grid unit determined as high risk of strong thunderstorm is constructed into a regression neural network model, which is used to predict the lightning frequency and geometric mean of lightning current amplitude of each grid unit within a predetermined period;

[0076] Based on the high-risk grid unit sample set selected by the probability classification model, the dynamic meteorological parameters and lightning history features in the high-risk grid unit sample set are used as input, and a multi-task neural network architecture is used to simultaneously learn the nonlinear mapping relationship between lightning frequency and lightning current intensity, generating a quantitative thunderstorm intensity prediction model.

[0077] The double hidden layer full connection structure is adopted, the first layer realizes high-dimensional feature nonlinear mapping, and the second layer enhances the model expression ability through feature multiplexing; the overfitting risk is inhibited through the random inactivation layer, and multiple base models are fused through the self-adaptive enhancement algorithm; the training set is consistent with the probability classification model, the verification set is based on the strong thunderstorm sample subset of double confirmation, the potential prediction and the actual observation are all determined as strong thunderstorm, and the generalization ability of the model to extreme events is ensured.

[0078] S16. Output the cooperative prediction result, including the strong thunderstorm potential risk judgment result of each grid unit in the preset period; for the grid unit judged to occur strong thunderstorm, the prediction values of the ground flash frequency and the geometric mean of lightning current amplitude are synchronously output.

[0079] Output the cooperative prediction result, including:

[0080] The strong thunderstorm potential judgment result of each grid unit is spatiotemporally associated and mapped with the intensity prediction parameter, a multi-dimensional structured data table is generated, the data table contains time dimension identifier, spatial grid code, event occurrence probability identifier and quantitative intensity evaluation parameter, and a standardized thunderstorm risk decision data interface is generated.

[0081] Taking a coastal area in Guangxi as an example, the hit probability of the probability classification model is significantly higher than the industry benchmark, the hit probability of the strong thunderstorm event of the probability classification model is 94.48%, greater than 90%, and the specific results are shown in Table 1, and the confusion matrix is shown in Figure 2 .

[0082] Table 1

[0083]

[0084] The prediction accuracy of the regression model for lightning current intensity meets the insulation design standard of power grid equipment, and the specific prediction results of the thunderstorm intensity regression prediction model are shown in Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 The sample proportion with a relative prediction error of ground flash frequency less than 50% is 34.42%, greater than 30%; the sample proportion with a relative prediction error of geometric mean of lightning current amplitude less than 50% is 84.31%, greater than 80%. The cooperative prediction result is pushed to the disaster prevention system in real time through a multi-dimensional data table, driving the patrol path planning and lightning protection device priority deployment. Including the strong thunderstorm potential prediction result (occurrence / non-occurrence) of each spatial grid in the future 12-24 hours; for the grid judged to occur strong thunderstorm, the predicted ground flash frequency and the geometric mean of lightning current amplitude are synchronously output; the prediction data is integrated and output in the form of a multi-dimensional structured table, the table contains timestamp, geographic grid code, strong thunderstorm potential prediction and intensity prediction parameter fields. The output result example is shown in Table 2.

[0085]

[0086] The above embodiment proposes a two-stage collaborative prediction framework, which synchronously improves the prediction accuracy of strong storm potential and realizes the quantification of intensity characteristics by optimizing the neural network algorithm while retaining the advantages of nonlinear mapping and adaptive learning.

[0087] It can be understood that those skilled in the art can combine various embodiments in the above embodiments to obtain various embodiments of the technical solutions.

[0088] The above merely describes the preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the potential risk and intensity of severe thunderstorms based on artificial neural networks, characterized in that: include: Obtain ground-to-ground lightning data for the target area, determine the temporal distribution characteristics of thunderstorm events based on the ground-to-ground lightning data, and identify the temporal trough of thunderstorm events, where the temporal trough is a sustained period during which the frequency of thunderstorm events is significantly lower than the climatological mean; Collecting multi-source meteorological data in the target area, screening meteorological data associated with the formation of thunderstorm events as thunderstorm prediction factors, and obtaining meteorological prediction factor data; Performing spatiotemporal alignment processing on the ground-to-ground lightning data and meteorological prediction factor data, constructing a daily grid cell sample set containing ground-to-ground lightning characteristics and meteorological characteristics based on the aligned data, and dividing the set into a training set and a test set according to the time series; Constructing a probability classification model for severe thunderstorm events based on the daily sample set of the grid cells, to predict the probability of severe thunderstorms occurring in each grid cell within a preset time; Based on the grid cells determined by the probability classification model to be at high risk of severe thunderstorms, a regression neural network model is constructed to predict the ground lightning frequency and the geometric mean of the lightning current amplitude in each grid cell within a preset time period; Output collaborative forecast results, including the potential risk identification results of severe thunderstorms in each grid cell within the preset time period; For grid cells identified as experiencing severe thunderstorms, the predicted values ​​of the ground flash frequency and the geometric mean of the lightning current amplitude are output synchronously.

2. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 1, wherein: The identification of the time trough of a thunderstorm event includes: Based on the ground-to-ground flash data monitored by multiple sites, the data is divided into equal time intervals and the cumulative ground-to-ground flash frequency in each interval is calculated. The continuous period with the lowest cumulative frequency is selected from the cumulative ground lightning frequency as the time valley, and the duration of the time valley matches the characteristics of the low-incidence stage of thunderstorm events in the climate statistical period.

3. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 1, wherein: The step of collecting multi-source meteorological data within the target area, screening meteorological data associated with the formation of thunderstorm events as thunderstorm prediction factors, and obtaining meteorological prediction factor data includes: Based on the physical triggering mechanism of thunderstorm events, meteorological parameters with dynamic response characteristics are screened from multi-dimensional meteorological observation data as thunderstorm prediction factors; the dynamic change trend of the meteorological parameters is associated with the evolution law of thunderstorm event intensity.

4. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 1, wherein: The performing spatiotemporal alignment processing on the ground-to-ground lightning data and the meteorological prediction factor data includes: Defining a daily time starting point based on the periodic characteristics of thunderstorm events, and dividing the target area into grid cells consistent with the spatial resolution of meteorological parameters; The daily ground flash frequency and lightning current intensity parameters in each grid unit are counted, and the sample categories are divided according to the intensity standards of thunderstorm events.

5. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 4, wherein: The statistical analysis of the daily ground flash frequency and lightning current intensity parameters in each grid cell and the classification of samples according to the intensity standard of thunderstorm events include: The ground-to-ground lightning frequency determination threshold is calculated based on the climate characteristic parameters of the target area and the grid unit area. The climate characteristic parameters include the annual average number of thunderstorm days and the ground-to-ground lightning event type ratio parameters. The calculation of the determination threshold meets the preset intensity level classification requirements.

6. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 5, wherein: The daily sample set of the grid unit includes time identifiers, ground lightning frequency statistics, lightning current intensity parameters, severe thunderstorm event discrimination identifiers and standardized meteorological prediction factors; the division of the training set and the test set meets the time continuity constraint, and the training set covers observation data for many consecutive years.

7. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 1, wherein: The constructing of a probability classification model for a severe thunderstorm event based on the daily sample set of the grid cells includes: Based on the training set data of the daily sample set of the grid cells, a neural network architecture is used to learn the joint distribution law of ground-to-ground lightning characteristics and meteorological characteristics; The model parameters are optimized through time series cross-validation to generate a probability prediction model for severe thunderstorm events.

8. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 7, wherein: Generating a severe thunderstorm event probability prediction model includes: The probability value identification and corresponding spatiotemporal location code of a severe thunderstorm event occurring in each grid cell within a preset time period are output, and the risk level is divided according to the probability threshold to generate a gridded thunderstorm potential discrimination result.

9. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 1, wherein: The grid cells determined as having a high risk of severe thunderstorms based on the probability classification model are constructed with a regression neural network model, including: Based on the high-risk grid cell sample set screened out by the probabilistic classification model, the dynamic meteorological parameters and historical characteristics of ground-to-ground flashes in the high-risk grid cell sample set are used as input. A multi-task neural network architecture is used to synchronously learn the nonlinear mapping relationship between ground-to-ground flash frequency and lightning current intensity to generate a quantitative thunderstorm intensity prediction model.

10. The method for predicting the potential risk and intensity of severe thunderstorms based on an artificial neural network according to claim 1, wherein: The output collaborative prediction result includes: The severe thunderstorm potential discrimination results of each grid cell are temporally and spatially mapped with the intensity prediction parameters to generate a multi-dimensional structured data table. The data table contains the time dimension identifier, spatial grid code, event probability identifier and quantitative intensity assessment parameters, and generates a standardized thunderstorm risk decision-making data interface.

Images