CN117148360B

CN117148360B - Lightning approach prediction method and device, electronic equipment and computer storage medium

Info

Publication number: CN117148360B
Application number: CN202311422852.5A
Authority: CN
Inventors: 姜睿娇; 张国平; 于廷照; 王曙东; 薛冰; 丁劲; 王阔音
Original assignee: Public Meteorological Service Center Of China Meteorological Administration National Early Warning Information Release Center
Current assignee: Public Meteorological Service Center Of China Meteorological Administration National Early Warning Information Release Center
Priority date: 2023-10-31
Filing date: 2023-10-31
Publication date: 2024-01-19
Anticipated expiration: 2043-10-31
Also published as: CN117148360A

Abstract

The invention provides a lightning approach forecasting method, a device, electronic equipment and a computer storage medium, and relates to the technical field of weather forecasting, wherein the method comprises the following steps: acquiring current multi-source heterogeneous acquisition data of a target research area; adjusting the preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data; inputting the current multi-source weather grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data; the prediction network model is a Visual Transformer model introduced with AFNO; and inputting the multi-source heterogeneous prediction data into a trained lightning prediction model to obtain a lightning prediction result of the target research area. Therefore, the Visual Transformer model introduced into the AFNO is used as a prediction network model, and the multi-source heterogeneous acquisition data are combined, so that the thunderstorm raw and elimination evolution characteristics can be effectively extracted, and the prediction resolution and accuracy are improved.

Description

Lightning approach prediction method and device, electronic equipment and computer storage medium

Technical Field

The invention relates to the technical field of weather forecast, in particular to a lightning approach forecast method, a lightning approach forecast device, electronic equipment and a computer storage medium.

Background

Traditional lightning prediction algorithms are mainly divided into two main categories: and forecasting the lightning potential and forecasting the lightning value. The principle of lightning potential prediction is to obtain the potential trend of lightning activity according to various convection parameters which can indicate the intensity of the unstable atmospheric state. Because the internal junction of thunderstorm cloud is complex and changeable when thunderstorm activity is developed, the observation of data in the cloud has limitation, so that the thunderstorm cloud is simulated and predicted to become an important supplement for forecasting the thunderstorm potential through a numerical mode. The machine learning is continuously superior to the performance level of the traditional forecasting method in the lightning forecasting, especially the application of the deep neural network, can effectively extract the change characteristics of the observed data in a period of time, and greatly improves the forecasting effect.

However, the general machine learning algorithm cannot capture the spatial correlation and time continuity of thunderstorm development, while the CNN (Convolutional Neural Networks, convolutional neural network) structure in deep learning does not perform well in capturing small scale variations in many time steps of auto regression reasoning, and it is difficult to extract the correlation between global spatial features and different variable factors. The above limitations have faced the bottleneck for higher spatial-temporal resolution lightning forecasting. However, industries which are seriously affected by lightning disaster such as aviation, railway, petrochemical industry, electric power and the like have urgent demands for refined lightning forecast.

Disclosure of Invention

The invention aims to provide a lightning approach forecasting method, a device, electronic equipment and a computer storage medium, so as to improve forecasting resolution and accuracy.

In a first aspect, an embodiment of the present invention provides a lightning approach prediction method, including:

acquiring current multi-source heterogeneous acquisition data of a target research area; the current multi-source heterogeneous acquisition data comprise Doppler radar combined reflectivity jigsaw in a current sampling period, satellite radiation imaging data of a plurality of preset channels, earth surface type data and altitude data, and the current sampling period comprises a plurality of moments before a current forecast starting moment;

adjusting the preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data;

inputting the current multi-source meteorological grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data in a current prediction period; wherein the current prediction period includes the current prediction starting time and a plurality of times after the current prediction starting time; the prediction network model is a Visual Transformer model introduced with an adaptive Fourier neural operator AFNO;

Inputting the multi-source heterogeneous prediction data into a trained lightning prediction model to obtain a lightning prediction result of the target research area in the current prediction period; the lightning prediction model is trained based on historical multi-source heterogeneous acquisition data and lightning observation data thereof.

Further, the preset space-time resolution is the space-time resolution of the satellite radiation imaging data; the adjusting the preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data comprises the following steps:

performing space-time resolution conversion on Doppler radar combined reflectivity puzzles in the current multi-source heterogeneous acquisition data through a preset interpolation algorithm to obtain radar combined reflectivity grid data consistent with the space-time resolution of the satellite radiation imaging data;

respectively converting the earth surface type data and the altitude data in the current multi-source heterogeneous acquisition data to obtain earth surface type grid data and altitude grid data consistent with the space-time resolution of the satellite radiation imaging data; the conversion processing comprises scale conversion, spatial interpolation and time dimension expansion;

Determining the satellite radiation imaging data, the radar combined reflectivity grid data, the earth surface type grid data and the altitude grid data as the current multi-source weather grid data.

Further, the lightning approach prediction method further includes:

acquiring a plurality of first multi-source heterogeneous acquisition data of a first research area and second multi-source heterogeneous acquisition data in a prediction period of the first multi-source heterogeneous acquisition data;

respectively adjusting the preset space-time resolution of each first multi-source heterogeneous acquisition data and each second multi-source heterogeneous acquisition data to obtain first multi-source meteorological grid data and second multi-source meteorological grid data;

and training the initial prediction network model by taking each first multi-source weather grid data as the input of the model and the corresponding second multi-source weather grid data as the label to obtain a trained prediction network model.

Further, the current multi-source weather grid data is composed of a plurality of two-dimensional images; the processing procedure of the prediction network model on the current multi-source meteorological grid data comprises the following steps:

performing image segmentation, vector mapping and position coding addition on each two-dimensional image in the current multi-source meteorological grid data to obtain a current middle layer characteristic corresponding to each two-dimensional image;

Performing space mixing operation and channel mixing operation on each current middle layer characteristic for multiple times to obtain multiple predicted middle layer characteristics at the next moment;

converting each prediction intermediate layer characteristic into a plurality of image block characteristics through a linear decoding layer, and performing image splicing on each image block characteristic through position decoding to obtain a prediction image corresponding to each prediction intermediate layer characteristic;

obtaining the multi-source heterogeneous prediction data through multiple iterations; wherein the multi-source heterogeneous prediction data includes a plurality of predicted images at each time instant within the current prediction period.

Further, the spatial mixing operation includes discrete fourier transform, dynamic filtering of the first multi-layer perceptron, superposition of soft threshold layers, and inverse discrete fourier transform; the channel mixing operation includes dynamic filtering of the second multi-layer perceptron.

Further, the lightning approach prediction method further includes:

acquiring a plurality of historical multi-source heterogeneous acquisition data and lightning observation data of a second research area;

adjusting the preset space-time resolution of each historical multi-source heterogeneous acquisition data to obtain historical multi-source meteorological grid data;

And training the initial lightning prediction model by taking the data of each grid point in each historical multi-source meteorological grid data at each moment as one sample and the corresponding lightning observation data as a label to obtain a trained lightning prediction model.

Further, the acquiring the plurality of historical multi-source heterogeneous acquisition data of the second research area and the lightning observation data thereof includes:

for each historical multi-source heterogeneous acquisition data, acquiring lightning imager data and CNLDN lightning data in a sampling period corresponding to the historical multi-source heterogeneous acquisition data;

and carrying out data fusion on the lightning imager data and the CNLDN lightning data to obtain lightning observation data corresponding to the historical multi-source heterogeneous acquisition data.

In a second aspect, an embodiment of the present invention further provides a lightning proximity prediction apparatus, including:

the acquisition module is used for acquiring current multi-source heterogeneous acquisition data of the target research area; the current multi-source heterogeneous acquisition data comprise Doppler radar combined reflectivity jigsaw in a current sampling period, satellite radiation imaging data of a plurality of preset channels, earth surface type data and altitude data, and the current sampling period comprises a plurality of moments before a current forecast starting moment;

The adjusting module is used for adjusting the preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data;

the prediction module is used for inputting the current multi-source meteorological grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data in a current prediction period; wherein the current prediction period includes the current prediction starting time and a plurality of times after the current prediction starting time; the prediction network model is a Visual Transformer model introduced with an adaptive Fourier neural operator AFNO;

the determining module is used for inputting the multi-source heterogeneous prediction data into a trained lightning prediction model to obtain a lightning prediction result of the target research area in the current prediction period; the lightning prediction model is trained based on historical multi-source heterogeneous acquisition data and lightning observation data thereof.

In a third aspect, an embodiment of the present invention further provides an electronic device, including a memory, and a processor, where the memory stores a computer program that can run on the processor, and when the processor executes the computer program, the processor implements the lightning proximity prediction method in the first aspect.

In a fourth aspect, an embodiment of the present invention further provides a computer storage medium, where a computer program is stored, where the computer program when executed by a processor performs the lightning approach prediction method according to the first aspect.

According to the lightning approach prediction method, the device, the electronic equipment and the computer storage medium provided by the embodiment of the invention, when the lightning approach prediction is carried out, the current multi-source heterogeneous acquisition data of the target research area are firstly obtained; the current multi-source heterogeneous acquisition data comprise Doppler radar combined reflectivity jigsaw in a current sampling period, satellite radiation imaging data of a plurality of preset channels, earth surface type data and altitude data, and the current sampling period comprises a plurality of moments before a current forecast starting moment; then, adjusting preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data; inputting the current multi-source meteorological grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data in a current prediction period; the current prediction period comprises a current prediction starting time and a plurality of times after the current prediction starting time; the prediction network model is a Visual Transformer model introduced with an adaptive Fourier neural operator AFNO; inputting the multi-source heterogeneous prediction data into a trained lightning prediction model to obtain a lightning prediction result of the target research area in the current prediction period; the lightning prediction model is trained based on historical multi-source heterogeneous acquisition data and lightning observation data thereof. Therefore, the Visual Transformer model of the adaptive Fourier neural operator AFNO is introduced as a prediction network model, and the multi-source heterogeneous acquisition data are combined, so that the thunderstorm raw and extinction evolution characteristics can be effectively extracted, and the prediction resolution and accuracy are improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the present invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of a lightning approach prediction method according to an embodiment of the present invention;

FIG. 2 is a technical roadmap of a lightning approach prediction method according to an embodiment of the invention;

FIG. 3 is a data processing flow chart of a lightning approach prediction method according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of a lightning proximity prediction apparatus according to an embodiment of the present invention;

fig. 5 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

The technical solutions of the present invention will be clearly and completely described in connection with the embodiments, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The observation data of radars, satellites and the like and the lattice data of numerical modes can be regarded as the category of images, and deep learning has been very successful in the fields of image recognition and prediction in recent years, and some prediction algorithms of thunderstorm and lightning activities are developed based on image extrapolation technologies.

The powerful performance of the transducer can bring a new revolution to the meteorological field with abundant space-time data, but the development speed of the current weather forecast technology based on machine learning is difficult to catch up with the daily and monthly updating steps of the neural network model. Currently, the ViT (Visual Transformer) network, which is the "peak at the start" in the CV (Computer Vision) field, has not yet played its strong role in lightning forecasting. Based on the above, according to the lightning proximity forecasting method, the device, the electronic equipment and the computer storage medium provided by the embodiment of the invention, the Visual Transformer model is applied to lightning forecasting, and the AFNO (Adaptive Fourier Neural Operator ) is introduced, so that the thunderstorm life-elimination evolution characteristics can be extracted, and the forecasting resolution and accuracy are improved.

For the convenience of understanding the present embodiment, a lightning approach prediction method disclosed in the present embodiment is first described in detail.

The embodiment of the invention provides a lightning approach prediction method which can be executed by electronic equipment with image processing capability. Referring to a schematic flow chart of a lightning approach prediction method shown in fig. 1, the method mainly includes the following steps S102 to S108:

step S102, current multi-source heterogeneous acquisition data of a target research area are acquired; the current multi-source heterogeneous acquisition data comprise Doppler radar combined reflectivity jigsaw in a current sampling period, satellite radiation imaging data of a plurality of preset channels, earth surface type data and altitude data, and the current sampling period comprises a plurality of moments before a current forecast starting moment.

The target research area can be selected according to actual requirements, and is not limited herein.

For example, the Doppler radar combined reflectivity jigsaw data can adopt business products pushed by a weather detection center of the China weather department, and the used variable is the combined reflectivity.

The satellite radiation imaging data may be data detected using an FY-4A meteorological satellite-loaded AGRI (Advanced Geostationary Radiation Imager, orbital radiation imager). AGRI carried by FY-4A has 14 channels in visible light, near infrared, short wave infrared, medium wave infrared, water vapor and long infrared wave bands, and the number of the AGRI is nearly three times that of 5 channels of weather satellite No. two of the wind cloud. Because the visible light channel and the near-short wave infrared channel can only observe effective data in daytime, 7 channels in water vapor, middle infrared and far infrared wave bands can be selected for training a model available in 24 h all the day, different physical characteristics can be observed by different wavelengths, and the thermal condition and the water vapor condition of thunderstorm occurrence can be well indicated. Based on this, the preset channels include 7 channels of water vapor, mid-infrared and far-infrared bands.

The surface type data may employ 30 m high resolution surface coverage type data made using Landsat satellite images, ten types including water, wetland, artificial surface, cultivated land, forest, shrubs, grasslands, bare land, moss and permanent snow/ice.

The altitude data may be land surface altitude data prepared from observations of the new generation of earth-facing observation satellite Terra of NASA (National Aeronautics and Space Administration, american aerospace agency), with a vertical resolution of 20 m and a horizontal resolution of 30 m.

The multi-source heterogeneous acquisition data of the past n time (i.e. instant) can be selected, the data of the future m time can be predicted, namely the sampling period comprises the past n time, and the prediction period comprises the future m time; wherein, m and n can be set according to practical requirements, and are not limited herein.

And step S104, adjusting the preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data.

Considering that the space-time resolution of the Doppler radar combined reflectivity jigsaw is higher than that of the satellite radiation imaging data, the preset space-time resolution can be set to be the space-time resolution of the satellite radiation imaging data, and the Doppler radar combined reflectivity jigsaw is adjusted to be the space-time resolution, so that consistency of different types of data in the space-time resolution is realized, and subsequent processing is facilitated.

Based on this, the above step S104 may be implemented by the following procedure: performing space-time resolution conversion on Doppler radar combined reflectivity puzzles in current multi-source heterogeneous acquisition data through a preset interpolation algorithm to obtain radar combined reflectivity grid data consistent with the space-time resolution of satellite radiation imaging data; respectively converting the earth surface type data and the altitude data in the current multi-source heterogeneous acquisition data to obtain earth surface type grid data and altitude grid data which are consistent with the space-time resolution of satellite radiation imaging data; the conversion processing comprises scale conversion, spatial interpolation and time dimension expansion; the satellite radiation imaging data, the radar combined reflectivity grid data, the earth surface type grid data and the altitude grid data are determined to be current multi-source meteorological grid data.

The interpolation algorithm described above may be, but is not limited to, a bilinear interpolation algorithm. The satellite radiation imaging data, the radar combined reflectivity grid data, the earth surface type grid data and the altitude grid data in the current multi-source meteorological grid data are two-dimensional images.

Step S106, inputting the current multi-source meteorological grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data in a current prediction period; the current prediction period comprises a current prediction starting time and a plurality of times after the current prediction starting time; the predictive network model is a Visual Transformer model incorporating an adaptive fourier neural operator AFNO.

The embodiment of the invention takes ViT as a backbone network, because the spatial dependence of each part of the image can be well simulated. The advantage of combining ViT backbone network with AFNO, AFNO-ViT for short, AFNO-ViT is that the mixing operation of token is set to a continuous global convolution, which can be effectively implemented with FFT (Fast Fourier Transform ) in the fourier domain. The complexity of the spatial mixing is thus reduced toThis allows the model to be flexibly extended in terms of spatial resolution and number of channels, and speed up in parallel when computed.

In some possible embodiments, the current multisource weather grid data is made up of a plurality of two-dimensional images; the processing procedure of the prediction network model for the current multi-source meteorological grid data comprises the following steps: performing image segmentation, vector mapping and position coding addition on each two-dimensional image in the current multi-source meteorological grid data to obtain a current middle layer characteristic corresponding to each two-dimensional image; performing space mixing operation and channel mixing operation on each current middle layer characteristic for multiple times to obtain multiple predicted middle layer characteristics at the next moment; converting each prediction intermediate layer characteristic into a plurality of image block characteristics through a linear decoding layer, and performing image splicing on each image block characteristic through position decoding to obtain a prediction image corresponding to each prediction intermediate layer characteristic; multiple iterations are carried out to obtain multi-source heterogeneous prediction data; wherein the multi-source heterogeneous prediction data includes a plurality of predicted images at each time instant within the current prediction period.

In the specific implementation, each two-dimensional picture in the current multi-source weather grid data can be divided into a plurality of image blocks, each image block is projected into a one-dimensional vector, and position codes are added to obtain the current middle layer characteristics corresponding to each two-dimensional image; the spatial blending operation may include discrete fourier transform, dynamic filtering of the first multi-layer perceptron, superposition of soft threshold layers, and inverse discrete fourier transform; the channel mixing operation may include dynamic filtering of the second multi-layer perceptron. The first multi-layer perceptron is a three-layer multi-layer perceptron, and the second multi-layer perceptron is a double-layer multi-layer perceptron. The number of mixing times may be determined according to actual requirements, and is not limited herein.

Step S108, inputting the multi-source heterogeneous prediction data into a trained lightning prediction model to obtain a lightning prediction result of the target research area in the current prediction period; the lightning prediction model is trained based on historical multi-source heterogeneous acquisition data and lightning observation data thereof.

The lightning prediction model can output lightning prediction results corresponding to the multi-source heterogeneous prediction data, and the lightning prediction results can comprise lightning falling areas and probability distribution in the falling areas at each moment. The lightning prediction model may employ a LightGBM model. The LightGBM model is an integrated tree model whose main idea is to iteratively train with weak classifiers (decision trees) based on a strategy of single-sided sampling of gradients and exclusive feature binding to get the optimal effect. The method supports high-efficiency parallel training, and has the advantages of higher training speed, lower memory consumption, better accuracy, support of distributed rapid processing of mass data and the like.

In the embodiment of the invention, when lightning approaching prediction is performed, current multi-source heterogeneous acquisition data of a target research area are firstly acquired; the current multi-source heterogeneous acquisition data comprise Doppler radar combined reflectivity jigsaw in a current sampling period, satellite radiation imaging data of a plurality of preset channels, earth surface type data and altitude data, and the current sampling period comprises a plurality of moments before a current forecast starting moment; then, adjusting preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data; inputting the current multi-source meteorological grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data in a current prediction period; the current prediction period comprises a current prediction starting time and a plurality of times after the current prediction starting time; the prediction network model is a Visual Transformer model introduced with an adaptive Fourier neural operator AFNO; inputting the multi-source heterogeneous prediction data into a trained lightning prediction model to obtain a lightning prediction result of the target research area in the current prediction period; the lightning prediction model is trained based on historical multi-source heterogeneous acquisition data and lightning observation data thereof. Therefore, the Visual Transformer model of the adaptive Fourier neural operator AFNO is introduced as a prediction network model, and the multi-source heterogeneous acquisition data are combined, so that the thunderstorm raw and extinction evolution characteristics can be effectively extracted, and the prediction resolution and accuracy are improved.

The embodiment of the invention also provides a training process of the prediction network model, which comprises the following steps:

step a1, a plurality of first multi-source heterogeneous acquisition data of a first research area and second multi-source heterogeneous acquisition data in a prediction period are acquired.

The first study area may be the same as the target study area or may be different from the target study area. The first multi-source heterogeneous acquisition data may include doppler radar combined reflectivity tiles, satellite radiation imaging data for a plurality of preset channels, earth surface type data, and altitude data for a sampling period thereof.

And a2, respectively adjusting preset space-time resolution of each first multi-source heterogeneous acquisition data and each second multi-source heterogeneous acquisition data to obtain first multi-source meteorological grid data and second multi-source meteorological grid data.

The specific process may refer to the corresponding content of step S104, and will not be described herein.

And a3, training the initial prediction network model by taking each first multi-source weather grid data as the input of the model and taking the corresponding second multi-source weather grid data as the label to obtain a trained prediction network model.

The embodiment of the invention also provides a training process of the lightning prediction model, which comprises the following steps:

And b1, acquiring a plurality of historical multi-source heterogeneous acquisition data and lightning observation data of a second research area.

The second study area may be different from both the first study area and the target study area, or may be the same as either the first study area or the target study area. The historical multi-source heterogeneous acquisition data may include doppler radar combined reflectivity tiles over its sampling period, satellite radiation imaging data for a plurality of preset channels, earth surface type data, and altitude data. The historical multi-source heterogeneous acquisition data is the same as the sampling period of the lightning observation data.

In order to improve the accuracy of the lightning observation data, the lightning observation data may be obtained by: for each historical multi-source heterogeneous acquisition data, acquiring lightning imager data and CNLDN lightning data in a sampling period corresponding to the historical multi-source heterogeneous acquisition data; and carrying out data fusion on the lightning imager data and the CNLDN lightning data to obtain lightning observation data corresponding to the historical multi-source heterogeneous acquisition data.

The lightning imager data can be acquired through LMI (Lightning Mapping Imager, lightning imager) mounted on FY-4A meteorological satellite. The CNLDN lightning data can be national lightning positioning data of a meteorological detection center of the China meteorological bureau, the CNLDN lightning data are acquired through CNLDN (China National Lightning Detection Network), the CNLDN comprises a plurality of sub-stations, and the sub-stations can adopt ADTD type equipment or DDW1 type equipment. ADTD type equipment is also called an ADTD type lightning locator, works in a VLF/LF frequency band, adopts a foundation multi-station direction finding technology and a time difference method to carry out positioning, and the system mainly detects high-power ground flashback, and the positioning result mainly comprises the information of ground flashback time, longitude and latitude, intensity and the like and is updated every minute. The DDW1 type equipment is also called a DDW1 type lightning positioning instrument, so that detection of partial cloud flash activity is realized, and the long-term cloud flash monitoring blank is filled.

And b2, adjusting the preset space-time resolution of each historical multi-source heterogeneous acquisition data to obtain historical multi-source meteorological grid data.

And b3, training the initial lightning prediction model by taking the data of each grid point in each historical multi-source meteorological grid data at each moment as one sample and the corresponding lightning observation data as a label, so as to obtain a trained lightning prediction model.

For ease of understanding, the lightning approach prediction method described above is described below as an example.

The lightning approach forecasting method provided by the embodiment of the invention is a high-space-time resolution lightning approach forecasting method based on an AFNO-ViT neural network. As shown in FIG. 2, the overall technology route is sorted, scaled and fused according to multi-source heterogeneous data collectionAFNO-ViT neural network construction ∈>Iterative extrapolation of past n time-dependent multisource meteorological grid data to future m time-dependent data +.>Super-parametric tuning and training of the LightGBM model +.>Thought expansion based on lightning landing areas and probability forecast of multiple weather and geographic factors.

1. Collecting and sorting multi-source heterogeneous data

The distribution of lightning in space is random, discrete and not easy to track. The observation and extraction of the occurrence and development characteristics of thunderstorm cloud is an effective means for identifying and forecasting thunder and lightning activities. The weather radar can observe the internal structure of the convection cloud, different channels of the static satellite can observe the occurrence and development characteristics of the convection cloud top, the weather radar has the recognition capability of primary convection, and can observe the water vapor and the temperature at different heights, so that the two types of data can form good advantage complementation, and the generation, development and dissipation processes of the thunderstorm cloud are reflected together. In addition, thunderstorms are susceptible to the formation of elevated terrain, and thus ground surface height and ground surface type are also key factors that affect the occurrence of thunderstorms. In the embodiment, 7 channels of radiation imaging data (namely satellite radiation imaging data), earth surface coverage type (namely earth surface type data) and height (namely altitude data) which can reflect the characteristics of layers at different heights of thunderstorms in Doppler radar combined reflectivity jigsaw and FY-4A are selected as lightning forecasting factors.

(1) Doppler radar combined reflectivity jigsaw

The Doppler weather radar jigsaw data adopts business products pushed by a weather detection center of the Chinese weather bureau, the used variables are combined reflectivity, the spatial coverage range of the data is 73.67-135.03 DEG E, 3.87-53.55 DEG N, the spatial resolution is 0.01 DEG, and the time resolution is 6 min.

(2) FY-4A AGRI radiation imaging data

FY-4A is the first star of the second generation stationary orbit meteorological satellite Fengyun No. four series in China, successfully transmits in the year of 2016 and 12 and 11, formally performs business operation in the east longitude of 104.7 degrees, and is loaded with four observation instruments, namely an advanced stationary orbit radiation imager (Advanced Geostationary Radiation Imager, AGRI), an interference atmosphere vertical detector (Geostationary Interferometric Infrared Sounder, GIIRS), a lightning imager (Lightning Mapping Imager, LMI) and a space environment monitoring instrument (Space Environment Monitoring Instrument Package, SEP). Compared with the existing wind cloud No. two static satellites, the AGRI and SEP loads are obviously improved, and a lightning imager aiming at the observation in the asiatai area is installed.

AGRI carried by FY-4A has 14 channels in visible light, near infrared, short wave infrared, medium wave infrared, water vapor and long infrared wave bands, and the number of the AGRI is nearly three times that of the No. 5 channels of the Fengyun. Since only the visible light channel and the near-short wave infrared channel can observe effective data in daytime, in order to train a model available for 24 h a whole day, in this embodiment, we choose 7 channels of water vapor, mid-infrared and far-infrared bands, and observe different physical characteristics at different wavelengths, so that the thermal conditions and water vapor conditions of thunderstorm occurrence can be well indicated, and the detailed information of channel selection of the AGRI for lightning prediction is shown in Table 1 below. The spatial resolution of these 7 channels in the northern hemisphere was 0.04 ° and the temporal resolution was 15 min.

TABLE 1

Besides basic channel observation products, FY-4A also provides rich secondary development products, but the research requirements of the embodiment are different from those of the embodiment in update time and forecast time, and deep learning has strong feature learning capability, and convection features can be directly extracted from the basic products, so that the embodiment only uses primary products as feature data.

(3) Lightning observation data

The lightning imager LMI mounted on FY-4A can detect the lightning activity of China and surrounding areas 3-9 months each year, and the principle is that a CCD (Charge-coupled Device) area array and an optical imaging technology are adopted to perform staring observation on the lightning light radiation in a detection area, and diffuse reflection light emitted by lightning illuminating a cloud top is mainly received, so that an observation object is mainly cloud flash. The field of view of FY-4A LMI approximates a trapezoid. The spatial resolution of the LMI at the point-under-the-satellite pixels is highest and is 7.8 km. The spatial resolution of the pixels gradually decreases with the increase of the latitude; meanwhile, in the longitudinal direction, the spatial resolution of the pixels far from the central axis is also reduced. At the two extreme corners of the LMI observation range, the spatial resolution of the pixel is as low as 20 km.

Since the spatial resolution of LMI is difficult to meet the model requirement, the present embodiment also uses another type of lightning observation data, namely the national lightning location data of the meteorological detection center of the chinese meteorological office, and the system for collecting this data is herein referred to as CNLDN (China National Lightning Detection Network). ADTD type equipment, namely an ADTD type lightning positioning instrument, works in a VLF/LF frequency band, adopts a foundation multi-station direction finding technology and a time difference method to position, the system mainly detects high-power ground flashback, and the positioning result mainly comprises the information of ground flashback time, longitude and latitude, intensity and the like, and is updated every minute. The DDW1 type equipment, namely the DDW1 type lightning locator, realizes the detection of partial cloud flash activity and fills the blank of cloud flash monitoring for a long time. The detection rate distribution condition of the CNLDN is not studied at present, but the detection rate of the areas in the east and south of China with dense site distribution is higher than that of the areas in the west and north of China with sparse site distribution according to the detection principle.

All three types of data can be acquired by a weather big data cloud platform.

(4) Surface coverage type and height data

In addition, the Chinese region is wide, the distribution difference of the height of the underlying surface and the ground surface type is large, and the two factors are also important factors influencing the meteorological characteristics, so the embodiment also uses the two types of geographic information data to participate in training together with the meteorological factors. The 30 m high resolution earth surface coverage type data produced by the land map geographic information agency of the people's republic of China using Landsat satellite images was collected, ten types including water, wetland, artificial earth surface, cultivated land, forest, bush, grassland, bare land, moss and permanent snow/ice. Land surface height data produced from observations of the new generation of earth-facing observation satellite Terra of NASA were collected with a vertical resolution of 20 m and a horizontal resolution of 30 m. Both types of data are global coverage.

2. Selection of research area, heterogeneous data scale conversion and space-time fusion

Considering that the research area is to be cross-covered by three kinds of observation data, the number of grid points is required to be whole hundred to facilitate image segmentation in the model, and meanwhile, superposition of newly divided grids and grids of source data is guaranteed to the greatest extent to reduce interpolation errors, the longitude range of the research area selected by the embodiment is 83-127 degrees E, and the latitude range is 15.5-51.5 degrees N, so that the research area can completely cover the southern area and the northern most area with high lightning incidence in China.

(1) Meteorological and geographic factor data fusion

Weather radar with high space-time resolution and stationary satellite data (namely Doppler radar combined reflectivity jigsaw and satellite radiation imaging data) are important carriers for recording the occurrence and development processes of thunderstorm activities. In addition, the type of ground surface and the height of the ground surface are also external factors that affect lightning. Radar reflectivity, 7 channels of radiation imaging data, and two types of geographic information data were taken as 10 factors for lightning prediction. Radar and satellite data have different spatial-temporal resolutions, with the resolution of radar data being higher. The radar combined reflectivity is converted by bilinear interpolation to the same spatial-temporal resolution as the AGRI data, i.e. 0.04 ° and 15 min. Meanwhile, the earth surface coverage type and the height data are subjected to scale conversion, spatial interpolation, time dimension expansion and the like. The investigation region thus constitutes a 1100×900 two-dimensional grid, and the observation data of the past n times constitutes a 1100×900×n×10 four-dimensional matrix, as shown by I in fig. 2.

(2) Lightning data fusion

The present embodiment uses lightning data as a tag for the training of the LightGBM model. Firstly, performing quality control on CNLDN lightning data: in order to ensure positioning accuracy, positioning results of three stations below are removed (namely lightning positioning data of lightning existing in a place is positioned by the three stations below are removed); to reduce spurious data, isolated positioning results within 50, km, 30 min are removed. And then the lightning positioning data are latticed, and the foundation lightning positioning data are converted into 'image' data similar to radar echo by calculating the total flash density in the space-time grid with the spatial resolution of 0.04 degrees and the time resolution of 15 minutes. Because a single lightning has a certain spatial scale, generally several kilometers to tens kilometers, lightning in the radius range of 10 km around each lattice point is included in density calculation of the lattice point when the lightning is lattice-treated, the method can ensure that the calculated density of the lightning is greater than the actual condition, but the model training is not influenced, the influence of the randomness of the lightning discharge can be reduced to a certain extent, and the forecasting accuracy is improved.

The CNLDN site has low detection efficiency in Qinghai-Tibet plateau and other places in the research area due to the non-uniformity of distribution, and has detection dead areas in sea areas and other land areas. Although the full coverage of the research area can be realized by the LMI mounted by the FY-4A, the spatial resolution can not meet the requirement, but the LMI mounted by the FY-4A has the advantages that the distribution of the spatial resolution, the detection rate and the detection error is in a fixed rule, and the LMI mounted by the FY-4A can be used for calibrating lightning density deviation caused by uneven distribution of sites of lightning positioning data of a foundation. The present embodiment innovatively proposes a foundation based on LMI data The lightning density correction method comprises the following specific steps: a. by combining the placement angle of CCD (charge coupled device) surface array, the earth curvature and the satellite height, fitting an empirical formula of the distribution of the detection rate of LMI (least mean squares) in a research area along with longitude and latitudeThe method comprises the steps of carrying out a first treatment on the surface of the b. Selecting a lattice point with most dense CNLDN site distribution and flat underlying surface to approximate the lightning density to be a true value, calculating the ratio of the lightning density to LMI lightning densityαThen->Or the ratio distribution of the densities of the two different positions; c. multiplying the distribution formula of the LMI lightning density by +.>The lightning distribution situation closest to the real density in the LMI irregular space resolution grid can be obtained; d. the spatial resolution of LMI is 7.8 km at the point under the satellite, the distortion is as low as 20 km at the most distal end, the spatial resolution of CNLDN after the latticed is 0.04 DEG, the detection rate difference of CNLDN lattice points in the lattice points with irregular spatial resolution of the same LMI is ignored, the lattice point data of CNLDN is utilized to downscale the irregular coarse grid density distribution obtained in the step c according to the weight, and the lightning density distribution condition after the deviation correction in the coverage area of CNLDN site can be obtained (for example, one lattice point in the irregular coarse grid density distribution is divided into a plurality of sub-lattice points, and the lightning density distribution condition after the deviation correction in the coverage area of CNLDN site is firstly based on the density ratio corresponding to each sub-lattice point Determining the weight of each sub-grid point, and multiplying the lightning density value of each sub-grid point by the weight of each sub-grid point to obtain the lightning density value of the corresponding sub-grid point, namely, the lightning density value of the sub-grid point is the product of the lightning density value of the grid point and the weight of the lightning density value of the grid point); e. and c, for detection dead zones of sea areas, other countries and the like, the coarse grid density distribution obtained in the step c can be directly subjected to linear interpolation to obtain uniform high-resolution lightning density distribution. It should be noted that, for the CNLDN probe blind area, the interpolation refinement method is not realized in a true senseDownscaling, but these areas are used only as deep learning training and are not the lightning forecast areas of great interest. The content of this part of the lightning observation correction fusion has been completed in the preamble work, as shown in fig. 2 as II.

3. AFNO-ViT model development and construction

The present embodiment uses ViT as the backbone network because it can well simulate the spatial dependence of the image parts. The advantage of combining ViT backbone network with AFNO, AFNO-ViT for short, AFNO-ViT in this embodiment is that the mixing operation of token is set to a continuous global convolution, which can be effectively implemented with fast fourier transforms (Fast Fourier Transform, FFT) in the fourier domain. With such a design, the complexity of spatial mixing is reduced to This allows the model to be flexibly extended in terms of spatial resolution and number of channels, and speed up in parallel when computed. Based on the above, in order to fully extract the time evolution information of the flow system in the observed data, the embodiment innovatively expands the ViT channel by one dimension, namely, changing the variable channel into the outer product of the variable channel and the time channel.

The flow of the AFNO-ViT model is shown in FIG. 2 as III, and is described in detail in the following steps:

(1) Dividing each 1100×900 two-dimensional picture intoNon-overlapping patches (i.e. image blocks), so the middle layer feature of the model can be expressed as +.>Wherein->Representing the number of channels that are to be processed,crepresenting the number of variables, in this embodimentc=10，nRepresenting the time of past observations. The three-dimensional size of each patch is +.>It is projected as a one-dimensional vector, called token in deep learning (i.e., each image block is converted into a token). First%i,j) The token can be expressed as +.>. In order to save the use of symbols +.>Represent the firstsBy token, useRepresent the firsttA token, wherein->。

(2) To record the positional relationship between patches, the token is added to the temporal and spatial position codes (using summation rather than stitching), and the position vector can be a relative position or an absolute position. Thus, the intermediate layer feature (composed of a plurality of token after the position coding) after the position coding can be obtained.

(3) In ViT original model, feature aggregation is performed among patches through self-attention mechanism, and definition of self-attention mechanism：

Wherein,respectively, matrix of Query, key and Value, definitionThe self-attention mechanism can be considered as an asymmetric matrix coreWritten->The self-attention mechanism can rewrite the nucleation function:

spatial blending based on a multi-headed self-attention mechanism can introduce complexity into the number of tokens to the power of two, and is not suitable for high resolution grids. Therefore, we extend the kernel summation method in the above formula to continuous kernel integration, tensor of inputXNo longer is Euclidean spaceIs defined in the spatial domain +.>Spatial function of (3). In this continuous representation, the network is changed into a kernel integration operator that extracts the features of the input function，

Wherein,as a continuous kernel function. Use of green's nucleus->Global convolution can be implemented, with the green kernel being one of the continuous kernel functions. Convolution is less complex than integration, and green's kernel has beneficial regularization effects and is able to capture global relationships. Global convolution may be implemented by FFT.

Defining Fourier neural operators(Fourier Neural Operators, FNO) for continuous input And (2) core->First, thesThe kernel integral of the token can be expressed as

And->Representing the continuous fourier transform and its inverse, respectively. Inspired by FNO, for a limited dimension picture that can be converted into a discrete grid, a discrete Fourier transform (Discrete Fourier Transform, DFT) can be used to applyDiscrete into fourier domain, i.e. token mixing process: />

Wherein,m,nthe spatial sequence number and the temporal sequence number, respectively. Defining complex weightsThe above can be transformed into:

on the basis, in order to reduce the number of parameters, the embodiment is intended toWAdding a rectangular diagonal structure to the matrix to obtain complex weightsWDecomposition into a stack of shared weights, each of which has dimensions ofThe above formula can be run in parallel:

inspired by the self-attention mechanism, in order to make the token adaptive, interact between different token, and be able to decide the passing mode of different frequencies (Multiply Frequency), dynamic filtering is performed using a three-Layer Multi-Layer perceptron (MLP) structure, which can be approximated as any function with sufficiently large hidden layers:

wherein,is the complex weight of the first layer, +.>Is the complex weight of the second layer, +.>Is the complex weight value of the third layer, bIs a bias parameter. />、/>、bAre shared by all token, which can effectively reduce the number of parameters. />Is a standard deviation scaled lightning density map.

On the basis, a soft threshold layer is further overlapped，λIs control ofThe sparsity tuning parameters are given by:

finally, the fourier domain is transformed back into the patch spatial domain by an inverse discrete fourier transform (Inverse Discrete Fourier Transform, IDFT). DFT assumes that global convolution is applied on periodic images, which is not applicable to real images, and to compensate for local features and non-periodic boundaries, a residual term can be added to token：

After the token is subjected to AFNO space mixing, the channel is subjected to double-layer MLP structuredMixing%dFrom the number of variables and time, data for different variables, different times are mixed). The entire mixing process is repeated L times to predict the token for the next time. Finally, the token is converted into patches through a linear decoding layer, and the patches are spliced into an image through position decoding (each patch comprises size and position information, so that a complete image can be spliced).

The framework can be built based on Python language Python, and can realize the past nThe radar and satellite images of the next time and the geographic factor data of the next time are used as input, and the total formula is as follows:

the present embodiment does not directly take lightning data as the extrapolation result, because the input data and the output data class need to be kept consistent to achieve iterative extrapolation for multiple times.

4. AFNO-ViT model training

Deep learning modelThe model needs to input a large amount of historical data for training. FY-4A radiation imaging data was published from 12 months 19 days 2018, so this example uses observation data from 60 months 2019 to 2023. One sample every 15 min, a total of 175320 samples. And the effective sample number is increased by a data augmentation method, and the sample number is changed into 5 times by horizontal overturning, 90-degree rotation, 180-degree rotation and 270-degree rotation respectively, so that 876600 samples are finally obtained, and the deep learning training requirement can be met. The samples were randomly partitioned into training, testing, validation sets in a 6:2:2 ratio. The characteristic data of each sample is pastnThe radar, satellite image and geographic factor data of the next time are used as marks.

In the training process, we choose a cross entropy loss function to measure the difference between the model predicted value and the actual value:

In the middle ofAs a positive sample class weighting factor, the present embodiment intends to set it as a function related to longitude and latitude, earth surface type, earth surface height.

Training is carried out tonThe radar and satellite data at each moment are input into the AFNO-ViT network, and an extrapolated image taking thunderstorm characteristics as focus of attention at one moment in the future can be predicted. Through the process ofmIteration is carried out for the futuremPrediction of the individual moments. By continuously debugging the model, the realization ofmMaximization.

5. LightGBM model tuning and training

Future generation using AFNO-ViT modelmAfter the radar satellite data of the thunderstorm is weighted and marked every time, the corresponding relation with thunder needs to be established. In conventional lightning potential forecasting methods, radar reflectivity up to 30 dBz is often used as an indicator of lightning occurrence (Seroka et al 2012). However, the thunderstorm structure is complex and changeable, and the rough estimation methodThe lightning refined forecasting requirement is difficult to meet. The machine learning algorithm can establish a plurality of meteorological factors and nonlinear models between the geographic factors and thunder and lightning under the condition that a thunderstorm discharge mechanism is not clearly known, and can achieve more accurate forecasting effect than experience forecasting under the support of big data.

The embodiment finally selects the LightGBM model by comparing a plurality of machine learning algorithms, which is a strong open source gradient lifting framework and is the most advanced integrated tree model at present, and the main idea is that a strategy based on single-side sampling of gradient and exclusive feature binding is used for iterative training by using a weak classifier (decision tree) to obtain the optimal effect. It supports efficient parallel training and has the advantages of faster training speed, lower memory consumption, better accuracy, support for distributed fast processing of mass data, etc., see IV in fig. 2. The LightGBM does not need to be built layer by layer like an AFNO-ViT deep neural network, and the LightGBM library in Python has made an integrated construction for the whole algorithm.

Since wind four LMI only observes in 3-9 months towards northern hemisphere, CNLDN covers cloud flash data from 1 st 2021, so this embodiment selects data of 3-9 rd 2021-2023 for training of LightGBM model. Unlike the deep learning model, the data of each lattice point is taken as one sample, so the number of samples isEach sample contains 10 features reflecting thunderstorm thermal, moisture and geographical conditions, and the lightning data after grid correction is used as a mark. The samples were also randomly partitioned into training, testing, validation sets in a 6:2:2 ratio. The large sample and limited features make fitting and visualization of the model easier. Since the number of negative samples (the points where no lightning occurs) is much larger than the number of positive samples (the points where lightning occurs) during training, the number of samples can be balanced using undersampling techniques. Then minimizing the hyper-parameter set of the cross entropy loss function by Bayes optimization search (i.e., computing posterior probability distribution using Bayes' formulas, finding globally optimal solutions from a given finite function evaluation) And combining to obtain a training model.

Using the 8 radar satellite variable factors and 2 geographic factors at the next time of the AFNO-ViT network output as inputs, the LightGBM model can output the lightning occurrence probability for each grid point. These probabilities can be used to form lightning falls within the investigation region and probability distribution predictions can be made within the falls.

In order to preserve the continuity of the current lightning live data, the lightning prediction data and the lightning live data are simultaneously connected into a Fusion layer (namely a mixed layer, the purpose is to fuse the lightning prediction data and the lightning live data according to a formula 13 through matrix Fusion), and the contribution of the lightning live data and the lightning prediction data is dynamically calibrated by adopting a parameter matrix Fusion method, wherein the contribution is shown in the formula 13. Wherein,representation->Lightning prediction data of moment (which is equal to +.>Live lightning data at the moment),the representation input is +.>Time-of-day multi-source heterogeneous prediction data +.>The output of the LightGBM model,L _obs,t representation oftLive data of lightning at the moment, parameter matrix->Training is carried out in historical data, and the weight of the live lightning data is continuously reduced along with the increase of the predicted time.

In summary, as shown in fig. 3, the overall data processing flow of the embodiment realizes lightning landing areas and probability forecast at m times in the future based on the radar satellite observation data at n times in the past.

The embodiment of the invention has the following beneficial effects:

the Visual Transformer model is applied to lightning forecasting for the first time, an adaptive Fourier neural operator AFNO is introduced, a one-dimensional characteristic channel is innovatively expanded into a two-dimensional characteristic channel, and lightning spectrum is used for carrying out marking weighting in the channel mixing process. Compared with models such as CNN, RNN, GRU, LSTM which are commonly used for time and space prediction, the method is better in extracting the characteristics of thunderstorm extinction evolution, so that the prediction resolution and accuracy are improved.

The AFNO-ViT neural network constructed by the embodiment has the following advantages:

a. the ViT model inherits the strong characteristic recognition and long-distance dependency capability in the computer vision task. b. Compared with ViT original network, the multi-head self-attention layer with quadratic complexity of number of token is replaced by global convolution based on AFNO, so that the space-time resolution of the model can be flexibly expanded, and the model has better expressive and generalization. c. Compared with the previous radar image extrapolation method based on CNN, each convolution kernel of CNN can only capture the characteristics of local pixels where the convolution kernel is located, and AFNO-ViT can capture the relation between global pixels, so that the characteristics are extracted in a larger range, and the limitation of information is effectively avoided; AFNO-ViT is better able to understand the rotation, expansion and contraction of the thunderstorm, while translational invariance of CNN makes it unintelligible about these changes. AFNO-ViT is therefore more capable of capturing the spatial development and temporal variations of thunderstorms. d. The model can fully extract the information of the evolution of the thunderstorm system along with time in the observed data, learn the actions and changes of different meteorological factors in the thunderstorm occurrence and development process and the complex association among the meteorological factors, realize the research of the principle of the thunderstorm mechanism and lead the machine learning to be no longer a 'black box'. e. After the early training is finished, the calculation cost in real-time prediction is several orders of magnitude lower than that of the most advanced numerical prediction mode at present, and the prediction result can be given in time in a few seconds.

The LightGBM is an open source framework which is fully tested and verified, and has the advantages of high efficiency, expandability, accuracy, interpretability, usability and the like.

Corresponding to the lightning approach forecasting method, the embodiment of the invention provides a lightning approach forecasting device. Referring to fig. 4, a schematic structural diagram of a lightning proximity prediction apparatus is shown, the apparatus includes:

an acquisition module 401, configured to acquire current multi-source heterogeneous acquisition data of a target research area; the current multi-source heterogeneous acquisition data comprise Doppler radar combined reflectivity jigsaw in a current sampling period, satellite radiation imaging data of a plurality of preset channels, earth surface type data and altitude data, and the current sampling period comprises a plurality of moments before a current forecast starting moment;

the adjustment module 402 is configured to perform adjustment of a preset spatial-temporal resolution on the current multi-source heterogeneous acquired data to obtain current multi-source meteorological grid data;

the prediction module 403 is configured to input the current multi-source weather grid data into a trained prediction network model to obtain multi-source heterogeneous prediction data in a current prediction period; wherein the current prediction period includes the current prediction starting time and a plurality of times after the current prediction starting time; the prediction network model is a Visual Transformer model introduced with an adaptive Fourier neural operator AFNO;

A determining module 404, configured to input the multi-source heterogeneous prediction data into a trained lightning prediction model, to obtain a lightning prediction result of the target research area in the current prediction period; the lightning prediction model is trained based on historical multi-source heterogeneous acquisition data and lightning observation data thereof.

Further, the preset space-time resolution is the space-time resolution of the satellite radiation imaging data; the adjustment module 402 is specifically configured to:

Further, the device further comprises a first training module, configured to:

Further, the device further comprises a second training module, configured to:

Further, the second training module is specifically configured to:

The lightning approach prediction device provided in this embodiment has the same implementation principle and technical effects as those of the lightning approach prediction method embodiment, and for a brief description, reference may be made to corresponding contents in the lightning approach prediction method embodiment where the lightning approach prediction device embodiment is not mentioned.

As shown in fig. 5, an electronic device 500 according to an embodiment of the present invention includes: the lightning proximity forecasting method comprises a processor 501, a memory 502 and a bus, wherein the memory 502 stores a computer program capable of running on the processor 501, and when the electronic device 500 runs, the processor 501 and the memory 502 communicate through the bus, and the processor 501 executes the computer program to realize the lightning proximity forecasting method.

Specifically, the memory 502 and the processor 501 can be general-purpose memories and processors, which are not particularly limited herein.

The embodiment of the invention also provides a computer storage medium, and a computer program is stored on the computer storage medium, and when the computer program is run by a processor, the lightning approach prediction method in the previous method embodiment is executed. The computer storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a RAM, a magnetic disk, or an optical disk, etc., which can store program codes.

Any particular values in all examples shown and described herein are to be construed as merely illustrative and not a limitation, and thus other examples of exemplary embodiments may have different values.

The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

In the several embodiments provided in this application, it should be understood that the disclosed apparatus and method may be implemented in other ways. The above-described apparatus embodiments are merely illustrative, for example, the division of the units is merely a logical function division, and there may be other manners of division in actual implementation, and for example, multiple units or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be through some communication interface, device or unit indirect coupling or communication connection, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.

Claims

1. A lightning approach prediction method, comprising:

2. The lightning approach forecast method of claim 1, wherein the predetermined spatial-temporal resolution is a spatial-temporal resolution of the satellite radiation imaging data; the adjusting the preset space-time resolution of the current multi-source heterogeneous acquisition data to obtain current multi-source meteorological grid data comprises the following steps:

3. The lightning proximity forecasting method of claim 1, further comprising:

4. The lightning approach forecast method of claim 1, wherein the current multisource weather grid data is comprised of a plurality of two-dimensional images; the processing procedure of the prediction network model on the current multi-source meteorological grid data comprises the following steps:

5. The lightning approach forecast method of claim 4, wherein the spatial mixing operation includes a discrete fourier transform, dynamic filtering of the first multi-layer perceptron, superposition of soft threshold layers, and an inverse discrete fourier transform; the channel mixing operation includes dynamic filtering of the second multi-layer perceptron.

6. The lightning proximity forecasting method of claim 1, further comprising:

7. The lightning approach forecast method of claim 6, wherein the obtaining a plurality of historical multi-source heterogeneous acquisition data of the second research region and lightning observation data thereof comprises:

8. A lightning proximity prediction apparatus, comprising:

9. An electronic device comprising a memory, a processor, the memory having stored thereon a computer program executable on the processor, wherein the processor, when executing the computer program, implements the lightning approach forecast method of any of claims 1-7.

10. A computer storage medium having a computer program stored thereon, which, when executed by a processor, performs the lightning approach prediction method of any of claims 1-7.