Numerical model driven deep learning lightning warning system and method

Through a numerical model-driven deep learning lightning warning system, combined with lightning monitoring data and EC weather forecast data for space-time integration, a deep learning model is built for lightning prediction, which solves the problems of reduced and deviation in the existing technology of lightning prediction accuracy and achieves more Accurate 2-hour forecast of lightning.

CN118114546BActive Publication Date: 2024-12-03WUHAN NARI LIABILITY OF STATE GRID ELECTRIC POWER RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311800415.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-26
Publication Date
2024-12-03
Estimated Expiration
2043-12-26

AI Technical Summary

Technical Problem

The prediction accuracy of the existing lightning prediction methods decreases after 2 hours, and there are spatial and temporal deviations in numerical weather forecasts, making it difficult to accurately predict the time and location of lightning.

Method used

A numerical model-driven deep learning lightning warning system was designed. The lightning monitoring data and EC weather forecast data were fused in space-time through the data fusion processing module, and a deep learning model was constructed for time-by-time prediction, and the Bayesian optimization strategy was used to optimize the super Parameters, the prediction accuracy is evaluated by using the ten-fold hierarchical cross-validation method.

Benefits of technology

A more accurate 2-hour forecast of lightning has been achieved, and EC numerical weather forecast data has no space coverage blind spots, and the meteorological elements are complete and easy to obtain, helping to implement the business-oriented lightning prediction model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118114546B_ABST
    Figure CN118114546B_ABST
Patent Text Reader

Abstract

The present invention discloses a deep learning lightning early warning system and method driven by numerical models, including a data fusion processing module, a sample construction module, a deep learning model construction module, and a parameter optimization verification module. The system uses a spatiotemporal fusion processing method to perform data fusion processing on lightning monitoring data and EC weather forecast data; constructs a lightning hourly prediction deep learning model; trains the lightning hourly prediction deep learning model with lightning monitoring data and EC weather forecast data, optimizes hyperparameters using a Bayesian optimization strategy, and uses a ten-fold stratified cross-validation method to evaluate the prediction accuracy of the lightning hourly prediction deep learning model to obtain an optimized lightning hourly prediction deep learning model. The EC data used in the present invention has no spatial coverage blind spots and is complete in meteorological elements, and the EC data is relatively easy to obtain, which is conducive to the commercial application of lightning prediction models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of lightning warning, and specifically relates to a numerical model driven deep learning lightning warning system and method. Background Art

[0002] Lightning disasters are listed as one of the most serious natural disasters. They pose a huge threat to industries such as electricity and transportation, as well as people's production and life. Predicting the occurrence of lightning can help improve lightning protection capabilities. However, lightning is highly random and its occurrence is inseparable from atmospheric conditions. Whether lightning occurs or not and the characteristics of its activities vary greatly under different conditions. Therefore, it is not easy to accurately predict the time and location of lightning.

[0003] At present, the commonly used lightning prediction methods include extrapolation of real-time lightning monitoring data and forecasting lightning potential using numerical models. However, due to the complex evolution of lightning generation and extinction, the prediction accuracy of the prediction method based on extrapolation will drop rapidly after 2 hours, and numerical weather forecasts have certain deviations in space and time. Therefore, it is necessary to consider integrating historical lightning monitoring data with numerical weather forecasts, using historical lightning monitoring data to reflect the development trend of recent lightning, and combining numerical weather forecasts to provide atmospheric state information corresponding to the target forecast period. In addition, the use of recent historical lightning monitoring data can correct the simulated spatiotemporal deviations of numerical weather forecasts. Summary of the invention

[0004] The purpose of the present invention is to provide a numerical model driven deep learning lightning warning system and method, which effectively integrates lightning monitoring data and EC (European Centre for Medium-Range Weather Forecasts) weather forecast data, and can make hourly rolling predictions of lightning in the next 2 hours based on 3-hour resolution EC weather forecast data. It has high practicality and flexibility, and can provide strong support for lightning prediction and early protection against lightning strikes.

[0005] To achieve this purpose, the numerical model driven deep learning lightning warning system designed by the present invention includes a data fusion processing module, a sample construction module, a deep learning model construction module, and a parameter optimization verification module; the data fusion processing module is used to perform data fusion processing on lightning monitoring data and EC weather forecast data using a spatiotemporal fusion processing method to obtain lightning monitoring rasterized data and EC weather forecast interpolated data with the same spatiotemporal resolution; the sample construction module constructs data samples based on the lightning monitoring rasterized data and EC weather forecast interpolated data obtained by the data fusion processing module as lightning prediction factors and corresponding labels, and divides the data samples into training set sample data and verification set sample data; the deep learning model construction module is used to The invention relates to constructing a deep learning model for hourly lightning prediction; inputting the training set sample data obtained by the sample construction module into the deep learning model for hourly lightning prediction, and using the deep learning model for hourly lightning prediction to mine the mapping relationship between lightning prediction factors and corresponding labels when lightning occurs, so as to obtain a trained deep learning model for hourly lightning prediction; the parameter optimization and verification module uses a Bayesian optimization strategy to input the verification set sample data obtained by the sample construction module into the trained deep learning model for hourly lightning prediction obtained by the deep learning model construction module, optimizes the hyperparameters, and uses a ten-fold stratified cross-validation method to evaluate the prediction accuracy of the deep learning model for hourly lightning prediction, so as to obtain an optimized deep learning model for hourly lightning prediction.

[0006] A numerical model driven deep learning lightning warning method comprises the following steps: using a spatiotemporal fusion processing method to perform data fusion processing on lightning monitoring data and EC weather forecast data to obtain lightning monitoring rasterized data and EC weather forecast interpolation data with the same spatiotemporal resolution;

[0007] Constructing data samples based on the lightning monitoring rasterized data and the EC weather forecast interpolation data as lightning prediction factors and corresponding labels, and dividing the data samples into training set sample data and validation set sample data;

[0008] Constructing a deep learning model for hourly lightning prediction; inputting the training set sample data into the deep learning model for hourly lightning prediction, and using the deep learning model for hourly lightning prediction to mine the mapping relationship between the lightning prediction factor and the corresponding label when lightning occurs, to obtain a trained deep learning model for hourly lightning prediction;

[0009] Using the Bayesian optimization strategy, the validation set sample data is input into the trained deep learning model for hourly lightning prediction, the hyperparameters are optimized, and the prediction accuracy of the deep learning model for hourly lightning prediction is evaluated using the ten-fold stratified cross-validation method to obtain the optimized deep learning model for hourly lightning prediction.

[0010] The beneficial effects of the present invention are as follows: the present invention integrates historical lightning monitoring data with EC numerical weather forecast data, not only digging out the development trend of lightning reflected in the historical lightning monitoring data, but also using the numerical weather forecast to provide atmospheric state information corresponding to the target forecast period, and obtaining a more accurate 2-hour lightning forecast result. The EC numerical weather forecast data used in the present invention has no spatial coverage blind spots, has complete meteorological elements, and the EC data is relatively easy to obtain, which is conducive to the commercial application of lightning prediction models. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a system structure diagram of the present invention;

[0012] Figure 2 is a flow chart of the method of the present invention;

[0013] Figure 3 This is the forecast time chart of EC numerical weather forecast data;

[0014] Figure 4 This is a comparison chart between the model prediction results and the actual lightning monitoring results of a thunderstorm process;

[0015] Among them, 1-data fusion processing module, 2-sample construction module, 3-deep learning model construction module, 4-parameter optimization and verification module, 5-lightning prediction module, and 6-result display and application module. DETAILED DESCRIPTION

[0016] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0017] A numerical model driven deep learning lightning warning system, such as Figure 1 As shown, it includes a data fusion processing module 1, a sample construction module 2, a deep learning model construction module 3, and a parameter optimization verification module 4;

[0018] The data fusion processing module 1 is used to perform data fusion processing on the lightning monitoring data and the EC weather forecast data by using a spatiotemporal fusion processing method to obtain lightning monitoring rasterized data and EC weather forecast interpolation data with the same spatiotemporal resolution;

[0019] The sample construction module 2 constructs data samples based on the lightning monitoring rasterized data and EC weather forecast interpolation data obtained by the data fusion processing module 1 as lightning prediction factors and corresponding labels, and divides the data samples into training set sample data and verification set sample data;

[0020] The deep learning model construction module 3 is used to construct a deep learning model for hourly lightning prediction with time series capture capability and high classification accuracy; the training set sample data obtained by the sample construction module 2 is input into the deep learning model for hourly lightning prediction, and the deep learning model for hourly lightning prediction mines the mapping relationship between the lightning prediction factor (the lightning prediction factor is the input quantity during the training of the deep learning model for hourly lightning prediction, and the prediction factor is input into the trained deep learning model for hourly lightning prediction to obtain the lightning prediction result) and the corresponding label when lightning occurs, so as to obtain the trained deep learning model for hourly lightning prediction; a label refers to a mark used to classify and identify data in deep learning, and in this embodiment, it is a mark of two categories: lightning occurrence and lightning non-occurrence. Each prediction factor requires a corresponding label, and the training model learns how to classify or identify new data samples through the corresponding relationship between the label and the prediction factor;

[0021] The parameter optimization and verification module 4 uses the Bayesian optimization strategy to input the verification set sample data obtained by the sample construction module 2 into the trained lightning hourly prediction deep learning model obtained by the deep learning model construction module 3, optimizes the hyperparameters, and uses the ten-fold stratified cross-validation method to evaluate the prediction accuracy of the lightning hourly prediction deep learning model to obtain an optimized lightning hourly prediction deep learning model.

[0022] In the above technical solution, the lightning monitoring data in the data fusion processing module 1 includes the longitude information and latitude information of the lightning strike area;

[0023] The EC weather forecast data in the data fusion processing module 1 includes atmospheric environmental variables that are closely related to the occurrence of lightning, including convective precipitation, large-scale precipitation, temperature at 500hPa, temperature at 925hPa, zero-degree altitude layer, convective effective potential energy, and vertical relative vorticity at 950hPa; the convective precipitation is recorded as CP; the large-scale precipitation is recorded as LSP; the temperature at 500hPa is recorded as T500; the temperature at 925hPa is recorded as T925; the zero-degree altitude layer is recorded as DEG01; the convective effective potential energy is recorded as CAPE; the vertical relative vorticity at 950hPa is recorded as Vor950. Large-scale precipitation is mainly affected by large-scale circulation (such as low pressure, front, etc.), which can lead to large-scale and continuous precipitation. Large-scale precipitation in EC is the precipitation predicted by the cloud scheme. Convective precipitation is mainly affected by the differences in temperature and humidity in local areas, which can lead to the formation of convective clouds and precipitation. Convective precipitation in EC is the precipitation forecasted by the convective scheme.

[0024] In the above technical solution, the specific implementation method of the data fusion processing module 1 is:

[0025] The discrete lightning monitoring data with a spatial accuracy of hundreds of meters are spatially gridded, and the spatial resolution of the spatially gridded lightning monitoring data is the same as that of the EC weather forecast data; the discrete lightning monitoring data with a temporal accuracy of microseconds are temporally gridded to generate the temporally gridded lightning monitoring data; the temporal resolution of the EC weather forecast data is transformed using the nearest neighbor interpolation method to obtain the interpolated EC weather forecast data, and the temporal resolution of the temporally gridded lightning monitoring data is the same as that of the interpolated EC weather forecast data; the temporal resolution of the temporally gridded lightning monitoring data is 1 hour, the original temporal resolution of the EC weather forecast data is 3 hours, and the temporal resolution of the interpolated EC weather forecast data is 1 hour.

[0026] In the above technical solution, the specific implementation method of the nearest neighbor interpolation method is:

[0027] The method of processing the EC grid data with a time resolution of 3 hours into a time resolution of 1 hour using the nearest neighbor interpolation method is as follows: According to the 3-hour resolution characteristic of EC data, the EC weather forecast data output once every 3 hours is copied backward 2 times to obtain the hourly EC weather forecast data within a 3-hour period.

[0028] In the above technical solution, the specific implementation method of the sample construction module 2 is:

[0029] The data samples constructed by the sample construction module 2 include a historical lightning monitoring rasterized data set, a future lightning monitoring rasterized data set, and a historical EC weather forecast interpolation data set;

[0030] The data samples are divided into training set sample data and validation set sample data according to a set ratio; in this embodiment, the set ratio is preferably 7:3;

[0031] A historical lightning monitoring data set is selected as a lightning prediction factor; the historical lightning monitoring data set is hourly lightning monitoring rasterized data m hours after the model reporting time, where m≥3, and in this embodiment, 3 is preferably selected;

[0032] A future lightning monitoring dataset is selected as a label. The future lightning monitoring dataset is the hourly lightning monitoring rasterized data n hours after the model reporting time. The n is determined according to the time range to be forecasted. In this embodiment, the time range to be forecasted is 2 hours, and n is preferably 2.

[0033] The historical EC weather forecast data set is selected as the lightning prediction factor; the historical EC weather forecast data set is the EC weather forecast interpolation data of the past k times closest to the model start time, and in this embodiment, k is preferably 2. For example, when the start time is 3 o'clock universal time, the EC data times used are the EC data of 0 o'clock and 3 o'clock. According to the characteristic of EC time resolution of 3 hours, three models need to be established for the subsequent hourly rolling prediction of lightning in the next 2 hours.

[0034] In the above technical solution, the deep learning model for hourly lightning prediction trained by the deep learning model construction module 3 includes a deep learning model for hourly lightning prediction A, a deep learning model for hourly lightning prediction B and a deep learning model for hourly lightning prediction C. The specific implementation method is:

[0035] According to the characteristic of EC data with a time resolution of 3 hours, Figure 3 As shown, a strategy of three-round reuse of EC weather forecast data and hourly rolling update of lightning data is proposed;

[0036] The data samples constructed by the sample construction module 2 are divided into the first moment data samples at the time when the EC weather forecast data is updated, the data samples at the corresponding time are divided into the second moment data samples, and the data samples at one time after the corresponding time are divided into the third moment data samples;

[0037] Input the data sample at the first moment into the deep learning model for hourly lightning prediction to obtain deep learning model A for hourly lightning prediction; input the data sample at the second moment into the deep learning model for hourly lightning prediction to obtain deep learning model B for hourly lightning prediction; input the data sample at the third moment into the deep learning model for hourly lightning prediction to obtain deep learning model C for hourly lightning prediction.

[0038] In the above technical solution, the model hyperparameters optimized by the parameter optimization verification module 4 include learning rate, optimizer, batch sample size, training rounds, number of hidden layers, number of hidden layer nodes, and activation function.

[0039] In this embodiment, the number of hidden layer nodes is preferably 4, and the number of nodes in each layer is 256, 512, 512 and 128 respectively. The "early stopping" strategy is used to prevent the model from overfitting. The core idea of ​​the early stopping strategy is to stop training once the verification loss stops decreasing by checking the performance of the deep learning model on the validation set data. This method aims to prevent the model from overfitting by ending the training early.

[0040] A numerical model driven deep learning lightning warning system, which also includes a lightning prediction module 5 and a result display application module 6;

[0041] The lightning prediction module 5 inputs the real-time lightning monitoring rasterized data and the real-time EC weather forecast interpolation data of the target test time period into the optimized lightning hourly prediction deep learning model obtained by the parameter optimization verification module 4, and obtains the lightning warning result for the next n hours, i.e., the probability of lightning occurrence;

[0042] The result display application module 6 is used to visualize the lightning warning results obtained by the lightning prediction module 5 according to the probability of lightning occurrence, store them in a result data table, and synchronize them to the lightning warning data application platform.

[0043] In the above technical solution, the lightning prediction module 5 inputs the lightning monitoring data and EC weather forecast data of the target test time period into the optimized lightning hourly prediction deep learning model obtained by the parameter optimization verification module 4 as follows:

[0044] The lightning monitoring data and EC weather forecast data are divided into the first moment real-time data samples starting from the EC weather forecast data update moment, the lightning monitoring data and EC weather forecast data at the corresponding moment are divided into the second moment real-time data samples, and the lightning monitoring data and EC weather forecast data at 2 moments after the corresponding moment are divided into the third moment real-time data samples;

[0045] Input the real-time data samples at the first moment into the optimized deep learning model A for hourly lightning prediction to obtain the hourly lightning prediction result n hours after the first moment; input the real-time data samples at the second moment into the optimized deep learning model B for hourly lightning prediction to obtain the hourly lightning prediction result n hours after the second moment; input the real-time data samples at the third moment into the optimized deep learning model C for hourly lightning prediction to obtain the hourly lightning prediction result n hours after the third moment.

[0046] In the actual prediction process, the lightning hourly prediction deep learning model A, the lightning hourly prediction deep learning model B and the lightning hourly prediction deep learning model C are used interchangeably, and hourly iterative rolling correction is performed to improve the lightning prediction effect.

[0047] A numerical model driven deep learning lightning warning method, such as Figure 2 As shown, it includes the following steps: using a spatiotemporal fusion processing method to perform data fusion processing on lightning monitoring data and EC weather forecast data to obtain lightning monitoring rasterized data and EC weather forecast interpolation data with the same spatiotemporal resolution;

[0048] Constructing data samples based on the lightning monitoring rasterized data and the EC weather forecast interpolation data as lightning prediction factors and corresponding labels, and dividing the data samples into training set sample data and validation set sample data;

[0049] Constructing a deep learning model for hourly lightning prediction; inputting the training set sample data into the deep learning model for hourly lightning prediction, and using the deep learning model for hourly lightning prediction to mine the mapping relationship between the lightning prediction factor and the corresponding label when lightning occurs, to obtain a trained deep learning model for hourly lightning prediction;

[0050] Using the Bayesian optimization strategy, the validation set sample data is input into the trained deep learning model for hourly lightning prediction, the hyperparameters are optimized, and the prediction accuracy of the deep learning model for hourly lightning prediction is evaluated using the ten-fold stratified cross-validation method to obtain the optimized deep learning model for hourly lightning prediction.

[0051] In this embodiment, in order to achieve a 2-hour lightning prediction for a rectangular space area, the following operations are used:

[0052] Data fusion: lightning monitoring and EC weather forecast data from 2020 to 2021 are used. The discrete lightning monitoring data (as shown in the example data in Table 1) are converted into grid data with a temporal and spatial accuracy of 1 hour and 0.5° respectively. According to the characteristic that the EC data has a temporal resolution of 3 hours (see Figure 2 ), the EC grid data with a time resolution of 3 hours is processed using the nearest neighbor interpolation method, that is, the forecast weather data output by EC is copied back 2 times and processed into a 1-hour time resolution, so as to achieve effective spatiotemporal fusion of lightning monitoring and EC weather forecast data;

[0053] Table 1 Example of lightning monitoring data in Sichuan Province on July 15, 2021 (including the time, latitude and longitude information of lightning occurrence)

[0054] time Microseconds latitude longitude Current 21:20:21 2804791 29.103894 105.059658 10.3 21:20:23 221116 29.432289 91.774612 17.7 21:20:23 221293 29.303452 91.62205 13.2 21:20:25 1430403 29.151595 105.10164 -46.1 21:20:25 3190804 29.166534 105.087603 -34.2 21:20:25 3747653 29.162258 105.085361 -8.7 21:20:25 6138372 29.132312 105.113735 -24 21:20:25 6590998 29.133157 105.112463 -8.3 21:20:26 2964764 30.344877 92.540879 -38.3 21:20:26 2991063 29.199563 105.162618 -12.2 21:20:27 8449831 29.384269 91.606833 18.8 21:20:28 619837 29.39498 91.625346 -8.1 21:20:28 927344 29.395301 91.625251 -7.4 21:20:28 5781953 28.398339 100.852021 28.4 21:20:29 4824370 29.091413 105.164682 13.5 21:20:30 1497839 29.650652 105.263556 -43.3

[0055] Sample construction: lightning in the past three hours and EC in the past two hours are used as lightning prediction factors, and lightning in the next two hours is used as a label. Based on the 3-hour resolution of EC data, lightning data is updated hourly and EC data is reused for two hours, so that lightning and EC sample data are updated hourly.

[0056] Model training: Three deep learning lightning prediction models are constructed using historical sample data sets that are updated hourly. The model framework is a convolutional neural network with four hidden layers. The number of nodes in each layer is 256, 512, 512, and 128, respectively, to ensure effective mining of data information. At the same time, during the training process, the "early stopping" strategy is used to prevent the model from overfitting. The Bayesian optimization strategy is used to optimize the hyperparameters of the deep learning model, and the accuracy of the deep learning model is evaluated using a ten-fold stratified cross-validation. The loss value, accuracy, and ROC curve of the training set and validation set are output and plotted as the training rounds change. Finally, the overall evaluation indicators of the model are obtained. The model hit rate is 85.31%, the false alarm rate is 45.84%, and the risk score is 0.45. The overall performance of the model is relatively good.

[0057] Model prediction: Call the built lightning warning model. Based on the characteristic that the time resolution of EC data is 3 hours, EC data is reused for 2 hours and lightning data is updated hourly to achieve hourly rolling prediction of lightning in the next 2 hours. The three models are used crosswise, and hourly iterative rolling corrections are made to improve the prediction effect.

[0058] Result visualization: Visualize the lightning warning results according to the probability of lightning occurrence (see Figure 4 ), the model reasoning results show that the model prediction results basically reflect the generation and movement trends of lightning. The warning results are stored in the result data table and synchronized to the corresponding lightning warning data application platform.

[0059] The above embodiments illustrate how to use lightning monitoring data and EC numerical weather forecast data to achieve hourly revised lightning forecasts for the next two hours. The present invention proposes using lightning in the past three hours and EC in the past two hours as lightning prediction factors, taking into account the development trend of the atmospheric background field over time, and improving the timeliness and accuracy of warnings. These methods and systems can help users better obtain lightning warning information and provide strong support for lightning protection in advance.

[0060] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0061] The contents not described in detail in this specification belong to the prior art known to those skilled in the art. It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0062] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0063] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0064] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims of the invention.

Claims

1. A numerical model driven deep learning lightning warning system, characterized by: It includes a data fusion processing module (1), a sample construction module (2), a deep learning model construction module (3) and a parameter optimization verification module (4); The data fusion processing module (1) is used to perform data fusion processing on lightning monitoring data and EC weather forecast data using a spatiotemporal fusion processing method to obtain lightning monitoring rasterized data and EC weather forecast interpolation data with the same spatiotemporal resolution; The sample construction module (2) constructs data samples based on the lightning monitoring rasterized data and EC weather forecast interpolation data obtained by the data fusion processing module (1), as lightning prediction factors and corresponding labels, and divides the data samples into training set sample data and verification set sample data; The deep learning model construction module (3) is used to construct a deep learning model for hourly lightning prediction; the training set sample data obtained by the sample construction module (2) is input into the deep learning model for hourly lightning prediction, and the deep learning model for hourly lightning prediction mines the mapping relationship between the lightning prediction factor and the corresponding label when lightning occurs, so as to obtain a trained deep learning model for hourly lightning prediction; The parameter optimization verification module (4) uses a Bayesian optimization strategy to input the verification set sample data obtained by the sample construction module (2) into the trained lightning hourly prediction deep learning model obtained by the deep learning model construction module (3), optimizes the hyperparameters, and obtains an optimized lightning hourly prediction deep learning model; The specific implementation method of the data fusion processing module (1) is: The lightning monitoring data is spatially rasterized, and the spatially rasterized lightning monitoring data has the same spatial resolution as the EC weather forecast data; the lightning monitoring data is temporally rasterized, and the temporally rasterized lightning monitoring data is generated; The time resolution of the EC weather forecast data is transformed by using the nearest neighbor interpolation method to obtain the interpolated EC weather forecast data, and the time resolution of the time-gridded lightning monitoring data is the same as that of the interpolated EC weather forecast data.

2. The deep learning lightning warning system driven by the numerical model according to claim 1 is characterized in that: The prediction accuracy of the deep learning model for hourly lightning prediction was evaluated using the ten-fold stratified cross-validation method.

3. The deep learning lightning warning system driven by the numerical model according to claim 1 is characterized in that: The specific implementation method of the nearest neighbor interpolation method is: The method of processing the EC grid data with a time resolution of 3 hours into a time resolution of 1 hour using the nearest neighbor interpolation method is as follows: the EC weather forecast data output by the EC once every 3 hours is copied backward 2 times to obtain the hourly EC weather forecast data within a 3-hour period.

4. The deep learning lightning warning system driven by the numerical model according to claim 1, characterized in that: The specific implementation method of the sample construction module (2) is: The data samples constructed by the sample construction module (2) include a historical lightning monitoring rasterized data set, a future lightning monitoring rasterized data set, and a historical EC weather forecast interpolation data set; Dividing the data samples into training set sample data and validation set sample data according to a set ratio; A historical lightning monitoring data set is selected as a lightning prediction factor; the historical lightning monitoring data set is hourly lightning monitoring rasterized data m hours past the model reporting time, where m≥3; Select a future lightning monitoring dataset as a label, where the future lightning monitoring dataset is the hourly lightning monitoring rasterized data n hours after the model reporting time, where n is determined according to the time range to be forecasted; A historical EC weather forecast data set is selected as a lightning prediction factor; the historical EC weather forecast data set is the EC weather forecast interpolation data of the past k times closest to the model start time.

5. The deep learning lightning warning system driven by the numerical model according to claim 1, characterized in that: The deep learning model for hourly lightning prediction trained by the deep learning model construction module (3) includes a deep learning model for hourly lightning prediction A, a deep learning model for hourly lightning prediction B, and a deep learning model for hourly lightning prediction C. The specific implementation method is: The data samples constructed by the sample construction module (2) are divided into first-time data samples at the corresponding time, second-time data samples at one time after the corresponding time, and third-time data samples at two time after the corresponding time; starting from the EC weather forecast data update time, the data samples constructed by the sample construction module (2) are divided into first-time data samples at the corresponding time, second-time data samples at one time after the corresponding time, and third-time data samples at two time after the corresponding time; Input the data sample at the first moment into the deep learning model for hourly lightning prediction to obtain deep learning model A for hourly lightning prediction; input the data sample at the second moment into the deep learning model for hourly lightning prediction to obtain deep learning model B for hourly lightning prediction; input the data sample at the third moment into the deep learning model for hourly lightning prediction to obtain deep learning model C for hourly lightning prediction.

6. The deep learning lightning warning system driven by the numerical model according to claim 1, characterized in that: The model hyperparameters optimized by the parameter optimization verification module (4) include learning rate, optimizer, batch sample quantity, training rounds, number of hidden layers, number of hidden layer nodes, and activation function.

7. The deep learning lightning warning system driven by the numerical model according to claim 1, characterized in that: It also includes a lightning prediction module (5) and a result display application module (6); The lightning prediction module (5) inputs the real-time lightning monitoring rasterized data and the real-time EC weather forecast interpolation data of the target test time period into the optimized lightning hourly prediction deep learning model obtained by using the parameter optimization verification module (4), and obtains the lightning warning results for the next n hours; The result display application module (6) is used to visualize the lightning warning results obtained by the lightning prediction module (5), store them in a result data table, and synchronize them to the lightning warning data application platform.

8. The deep learning lightning warning system driven by the numerical model according to claim 7, characterized in that: The lightning prediction module (5) inputs the lightning monitoring data and EC weather forecast data of the target test time period into the optimized lightning hourly prediction deep learning model obtained by using the parameter optimization verification module (4) as follows: The lightning monitoring data and EC weather forecast data are divided into the first moment real-time data samples starting from the EC weather forecast data update moment, the lightning monitoring data and EC weather forecast data at the corresponding moment are divided into the second moment real-time data samples, and the lightning monitoring data and EC weather forecast data at 2 moments after the corresponding moment are divided into the third moment real-time data samples; Input the real-time data samples at the first moment into the optimized deep learning model A for hourly lightning prediction to obtain the hourly lightning prediction result n hours after the first moment; input the real-time data samples at the second moment into the optimized deep learning model B for hourly lightning prediction to obtain the hourly lightning prediction result n hours after the second moment; input the real-time data samples at the third moment into the optimized deep learning model C for hourly lightning prediction to obtain the hourly lightning prediction result n hours after the third moment.

9. A deep learning lightning warning method driven by the numerical model of the system according to claim 1, characterized in that: It includes the following steps: The lightning monitoring data and EC weather forecast data are fused by using the spatiotemporal fusion processing method to obtain lightning monitoring rasterized data and EC weather forecast interpolated data with the same spatiotemporal resolution. Constructing data samples based on the lightning monitoring rasterized data and the EC weather forecast interpolation data as lightning prediction factors and corresponding labels, and dividing the data samples into training set sample data and validation set sample data; Build a deep learning model for hourly lightning prediction; Inputting the training set sample data into the lightning hourly prediction deep learning model, using the lightning hourly prediction deep learning model to mine the mapping relationship between the lightning prediction factor and the corresponding label when lightning occurs, to obtain a trained lightning hourly prediction deep learning model; Using the Bayesian optimization strategy, the validation set sample data is input into the trained deep learning model for hourly lightning prediction, the hyperparameters are optimized, and the prediction accuracy of the deep learning model for hourly lightning prediction is evaluated using the ten-fold stratified cross-validation method to obtain the optimized deep learning model for hourly lightning prediction.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the numerical model-driven deep learning lightning warning method as claimed in claim 9 by executing the computer instructions.

11. A computer-readable medium having a computer program stored thereon, wherein the computer program executes the numerical model driven deep learning lightning warning method according to claim 9 when run.