Lightning frequency nowcasting method and system

Abstract

The invention provides a lightning frequency nowcasting method and system, and relates to the technical field of lightning nowcasting, and the method comprises the steps: obtaining the actual measurement data of severe convective weather of a target region in a current time period, the severe convective weather actual measurement data at least comprise weather radar echo actual measurement data, lightning frequency actual measurement data and stationary satellite multichannel actual measurement infrared brightness temperature; and based on a preset time forecasting interval, inputting the severe convective weather measured data into a multi-branch coding lightning frequency forecasting model to obtain weather radar echo forecasting data and lightning frequency forecasting data of the target area in the next time period output by the multi-branch coding lightning frequency forecasting model, the multi-branch coding lightning frequency forecasting model is obtained through training based on severe convection weather sample data marked with severe convection weather labels. According to the invention, lightning frequency forecasting is realized, and the accuracy of lightning forecasting is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lightning nowcasting, and in particular to a lightning frequency nowcasting method and system. Background Art

[0002] Lightning, as a disastrous weather phenomenon caused by severe convective activity, poses a serious threat to humans, buildings, power facilities, etc. Whether it is a direct hit or an indirect impact, it may cause significant casualties and property losses.

[0003] There are currently two main technical means for forecasting lightning. The first forecasting method is based on observational diagnosis and extrapolation, which relies on a variety of observation tools such as ground automatic stations, Doppler weather radars, geostationary satellites, and lightning locators. It uses statistical methods to explore the relationship between different forecasting factors and lightning, and combines convection identification technology, optical flow extrapolation, and machine learning to achieve the forecast of lightning activity. The other forecasting method is based on the parameterization scheme of the numerical weather forecast model, which focuses on the parameterization of the electrification and discharge process inside the cloud, or by analyzing the relationship between factors such as water particles and updrafts in the cloud and lightning, to build a forecasting model, thereby achieving the forecast of lightning.

[0004] However, the existing statistical diagnostic methods based on observations are difficult to accurately capture the complex nonlinear relationship between multiple prediction factors and lightning, resulting in insufficient accuracy in the final lightning forecast; and the parameterization scheme based on numerical weather forecasting is highly dependent on the parameter settings of microphysics and the discharge process, is sensitive to parameter selection, and has a very high computational cost. Therefore, a lightning frequency nowcasting method and system are urgently needed to solve the above problems. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a lightning frequency nowcasting method and system.

[0006] The present invention provides a lightning frequency nowcasting method, comprising: Acquire the severe convective weather measured data of the target area in the current period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; Based on the preset time forecast interval, the measured data of severe convective weather is input into the multi-branch coded lightning frequency forecast model to obtain the weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecast model, wherein the multi-branch coded lightning frequency forecast model is trained based on severe convective weather sample data marked with severe convective weather labels.

[0007] According to a lightning frequency nowcasting method provided by the present invention, the multi-branch coding lightning frequency forecasting model is trained by the following steps: Construct weather sample data based on weather radar echo sample data, lightning frequency sample data and geostationary satellite multi-channel sample infrared brightness temperature; Based on the preset time forecast interval, determining weather sample data for the next time period corresponding to the weather sample data; Using the weather sample data of the next period as the severe convective weather label, marking the weather sample data, and obtaining the severe convective weather sample data marked with the severe convective weather label, wherein the lightning frequency label in the severe convective weather label corresponds to different lightning frequency levels; The deep learning model is trained according to the severe convective weather sample data to obtain the multi-branch encoding lightning frequency forecasting model.

[0008] According to a lightning frequency nowcasting method provided by the present invention, weather sample data is constructed based on weather radar echo sample data, lightning frequency sample data and geostationary satellite multi-channel sample infrared brightness temperature, including: Performing maximum and minimum value normalization processing on the weather radar echo sample data to obtain normalized weather radar echo sample data; Performing logarithmic transformation and normalization processing on the lightning frequency sample data to obtain normalized lightning frequency sample data; Performing mean-standard deviation normalization processing on the geostationary satellite multi-channel sample infrared brightness temperature to obtain normalized geostationary satellite multi-channel sample infrared brightness temperature; The weather sample data is constructed according to the normalized weather radar echo sample data, the normalized lightning frequency sample data and the normalized geostationary satellite multi-channel sample infrared brightness temperature.

[0009] According to a lightning frequency nowcasting method provided by the present invention, the deep learning model includes a first encoder, a second encoder, an attention time fusion module and a decoder, wherein: The first encoder is used to extract features from the weather radar echo sample data and the lightning frequency sample data to obtain a first high-dimensional feature; The second encoder is used to extract features of the infrared brightness temperature of the multi-channel samples of the geostationary satellite to obtain a second high-dimensional feature; The attention time fusion module is used to perform feature fusion on the first high-dimensional feature and the second high-dimensional feature based on the attention mechanism to obtain a high-dimensional fusion feature; The decoder is used to predict lightning frequency according to the high-dimensional fusion features.

[0010] According to a lightning frequency nowcasting method provided by the present invention, a deep learning model is trained according to the severe convective weather sample data to obtain the multi-branch encoding lightning frequency forecasting model, including: The deep learning model is trained according to the severe convective weather sample data until the loss function value in the training process reaches a convergence condition, thereby obtaining the multi-branch coding lightning frequency forecast model, wherein the calculation formula of the loss function value is: in, represents the loss function value, Indicates the loss value corresponding to the weather radar echo, Indicates the loss value corresponding to the lightning frequency, The total number of tags indicating the lightning frequency level, Indicates The forecast value corresponding to the lightning frequency level label, Indicates The label value corresponding to the lightning frequency level label, Represents the loss value between the predicted value and the true label value, Indicates the number of time frames, Indicates Pixel coordinates in the frame radar echo The corresponding measured value is Indicates Pixel coordinates in the frame radar echo The corresponding forecast value is Indicates Pixel coordinates in the frame radar echo The corresponding weight value.

[0011] According to a lightning frequency nowcasting method provided by the present invention, the method further comprises: Based on the radar echo time and spatial resolution corresponding to the weather radar echo sample data, the lightning observation sample data is resampled to obtain the lightning frequency sample data that is aligned with the weather radar echo sample data in time and space, wherein the lightning observation sample data is discrete lightning observation data observed by a lightning locator.

[0012] The present invention also provides a lightning frequency nowcasting system, comprising: A severe convective weather data acquisition module is used to obtain the severe convective weather measured data of the target area in the current period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; The lightning frequency forecast module is used to input the measured severe convective weather data into the multi-branch coded lightning frequency forecast model based on a preset time forecast interval, and obtain the weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecast model, wherein the multi-branch coded lightning frequency forecast model is trained based on severe convective weather sample data marked with severe convective weather labels.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any one of the lightning frequency nowcasting methods described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the lightning frequency nowcasting method described in any one of the above is implemented.

[0015] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method for nowcasting lightning frequency as described above is implemented.

[0016] The lightning frequency nowcasting method and system provided by the present invention realizes the coordinated forecasting of radar echo and lightning frequency by considering the relationship between convection intensity and lightning frequency, thereby improving the accuracy of lightning forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A schematic diagram of the flow chart of the lightning frequency nowcasting method provided by the present invention; Figure 2 A schematic diagram of the structure of the multi-branch coding lightning frequency forecasting model provided by the present invention; Figure 3 A schematic diagram of the overall process of the lightning frequency nowcasting method provided by the present invention; Figure 4A comparison chart of the lightning frequency nowcasting method provided by the present invention and the existing lightning forecasting method at different times; Figure 5 A schematic diagram of the structure of the lightning frequency nowcasting system provided by the present invention; Figure 6 This is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Lightning is a disastrous weather phenomenon of severe convective activity. Whether it directly or indirectly hits people, buildings, power facilities, etc., it will cause serious casualties and property losses. There are currently two main ways to predict lightning: one is the diagnosis and extrapolation method based on observations (for example, ground automatic stations, Doppler weather radars, geostationary satellites, and lightning locators). This method is mainly based on statistical methods to analyze the relationship between different prediction factors and lightning, and combines convection identification, optical flow extrapolation, and machine learning methods to achieve lightning forecasts; the other is a parameterization method based on numerical weather forecast models. This method is mainly based on the parameterization scheme of electrification and discharge in the cloud, or based on the diagnosis scheme of the relationship between factors such as hydroponic particles and updrafts in the cloud and lightning, to achieve lightning forecasts.

[0021] In addition, existing technologies also utilize Doppler weather radar products (such as radar echoes, echo top heights, vertical accumulated liquid water content, etc.), geostationary satellites (infrared channels, water vapor channels, etc.) and lightning locators (imagers) observations, and combine them with deep learning models such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to carry out lightning probability forecasts for 0 to 60 minutes.

[0022] However, among the existing lightning forecasting methods, statistical diagnostic methods based on observations have difficulty in achieving high accuracy when dealing with the complex and nonlinear relationship between multiple forecasting factors and lightning. Although the parameterization method based on numerical weather forecast models provides a forecasting approach, it is highly dependent on the parameterization of microphysical processes and the initiation and discharge process, which makes the method very sensitive to parameter selection and is accompanied by huge computational requirements.

[0023] Although lightning nowcasting based on machine learning methods has improved the forecast accuracy to a certain extent, this improvement is mainly limited to single-task learning, that is, only the probability of lightning occurrence is predicted, and the forecast of lightning frequency is not involved. In addition, this method also ignores the intrinsic connection between the intensity of severe convective activity and lightning. It is worth noting that there is a good consistency between the changes in lightning frequency and severe convective weather: generally, convective systems with sudden changes in lightning frequency have a higher degree of organization, stronger intensity and longer life cycle. Therefore, lightning frequency can be used as an important indicator factor for the nowcast of severe convective weather (such as hail, tornadoes, thunderstorms, supercells, etc.).

[0024] In view of the problems existing in the above-mentioned prior art, the present invention constructs a multi-task learning lightning frequency nowcasting model (i.e., a multi-branch coding lightning frequency forecasting model) that integrates multi-source heterogeneous observations. By considering the relationship between convective intensity and lightning frequency, the coordinated forecast of convective intensity (radar echo) and lightning frequency is achieved, thereby improving the accuracy of lightning frequency forecast.

[0025] Figure 1 A schematic diagram of the flow chart of the lightning frequency nowcasting method provided by the present invention is shown in FIG. Figure 1 As shown, the present invention provides a lightning frequency nowcasting method, comprising: Step 101, obtaining the severe convective weather measured data of the target area in the current time period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature.

[0026] In the present invention, weather radar echo measured data (with a time resolution of 6 minutes and a spatial resolution of 1 km) are essential for identifying and analyzing severe convective weather systems (such as rainstorms, thunderstorms, tornadoes, etc.) because they can capture the movement of precipitation particles and precipitation intensity. During the data collection process, the present invention will perform quality control on the radar data, check the integrity and consistency of the data, and ensure that there are no missing data or outliers.

[0027] A lightning locator is a key tool for monitoring lightning activity. It can provide the precise location and time of lightning occurrence. By collecting lightning location data that matches the time range of weather radar data, it can obtain actual lightning frequency data in the target area.

[0028] At the same time, the present invention provides observation data of water vapor and infrared channels through the Sunflower 8 / 9 geostationary satellites, thereby obtaining the geostationary satellite multi-channel measured infrared brightness temperature, and these data have a time resolution of 10 minutes and a spatial resolution of 2 kilometers.

[0029] Preferably, since satellite observations may have data missing at certain times (such as cloud cover or satellite failure), the present invention will use appropriate methods (such as interpolation algorithms) to fill in these missing moments to ensure data continuity. The present invention collects multi-source observation data such as Doppler weather radar, geostationary satellite and lightning locator, and performs quality control and missing data filling, so as to obtain the measured data of severe convective weather in the target area in the current period, including weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature, which will be used for subsequent lightning frequency nowcasting.

[0030] Step 102, based on a preset time forecast interval, input the measured severe convective weather data into a multi-branch coded lightning frequency forecast model to obtain weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecast model, wherein the multi-branch coded lightning frequency forecast model is trained based on severe convective weather sample data marked with a severe convective weather label.

[0031] In the present invention, the preset time forecast interval indicates when to start forecasting future weather and determines the length of the forecast time. For example, if the lightning frequency forecast for the next 2 hours is to be forecasted, the preset time forecast interval is 2 hours from the current time period. The selection of this preset time forecast interval can be determined based on factors such as business requirements, model performance, and data availability.

[0032] It should be noted that in the present invention, the measured data of severe convective weather in the time period from the current moment to the previous 6 minutes is used as the measured data of severe convective weather in the current period for analysis. Therefore, when obtaining the measured data of severe convective weather, considering the highly temporal and spatial discrete characteristics of lightning observation, it is necessary to grid the lightning locator observation data, and perform temporal and spatial resampling of the lightning locator observation data according to the radar echo time and spatial resolution corresponding to the matched severe convective process, and sample the discrete lightning locator observation data into a lightning density grid product with a temporal resolution of 6 minutes and a spatial resolution of 1 km, thereby obtaining the measured data of lightning frequency.

[0033] In the present invention, a multi-branch encoding lightning frequency forecast model is used to forecast a machine learning model of lightning frequency and weather radar echo. The model adopts a multi-branch structure, which can simultaneously process multiple types of input data (such as weather radar echo, lightning frequency and infrared brightness temperature, etc.) and output corresponding forecast results. Among them, different branches of the model can process different types of input data, and each branch has its own processing logic and parameters, so that the model can more comprehensively capture the characteristics of weather changes.

[0034] In the present invention, the multi-branch coding lightning frequency forecast model uses coding technology (such as convolutional layers in convolutional neural networks, recurrent layers in recurrent neural networks, etc.) to extract the features of input data and convert these features into a format that the model can understand. Finally, the multi-branch coding lightning frequency forecast model outputs lightning frequency forecast data and weather radar echo forecast data, wherein the weather radar echo forecast data provides the possible situation of weather radar echoes in the future period of time, which is helpful to identify the future precipitation area and intensity; the lightning frequency forecast data provides the possible situation of lightning activities in the future period of time, which is helpful to evaluate the possibility and intensity of future severe convective weather.

[0035] In the present invention, the multi-branch coding lightning frequency forecast model is trained by a large amount of sample data marked with severe convective weather labels, including historical weather radar echo data, lightning frequency data, infrared brightness temperature data, etc. The severe convective weather label is an important guiding information in the model training process, which is used to guide the model which features are related to severe convective weather, thereby helping the model to better identify and learn these features. The trained multi-branch coding lightning frequency forecast model can receive the actual measured data of severe convective weather in the current period as input, and output the weather radar echo forecast data and lightning frequency forecast data of the target area in the next period.

[0036] The lightning frequency nowcasting method provided by the present invention realizes the coordinated forecasting of radar echo and lightning frequency by considering the relationship between convection intensity and lightning frequency, thereby improving the accuracy of lightning forecasting.

[0037] Based on the above embodiment, the multi-branch coding lightning frequency prediction model is trained by the following steps: Construct weather sample data based on weather radar echo sample data, lightning frequency sample data and geostationary satellite multi-channel sample infrared brightness temperature; Based on the preset time forecast interval, determining weather sample data for the next time period corresponding to the weather sample data; Using the weather sample data of the next period as the severe convective weather label, marking the weather sample data, and obtaining the severe convective weather sample data marked with the severe convective weather label, wherein the lightning frequency label in the severe convective weather label corresponds to different lightning frequency levels; The deep learning model is trained according to the severe convective weather sample data to obtain the multi-branch encoding lightning frequency forecasting model.

[0038] In the present invention, it is first necessary to collect weather sample data from multiple sources, including: 1. Doppler weather radar data, which has a time resolution of 6 minutes and a spatial resolution of 1 km, provides high-resolution weather radar echo information. These data are essential for identifying precipitation areas, estimating precipitation intensity, and monitoring the evolution of weather systems.

[0039] 2. Sunflower 8 / 9 geostationary satellite data, which include observations of water vapor and infrared channels (such as lower tropospheric water vapor channels, middle tropospheric water vapor channels, upper tropospheric water vapor channels, cloud phases and cloud top heights, etc.), with a temporal resolution of 10 minutes and a spatial resolution of 2 km. The water vapor channel helps identify the distribution of water vapor in the atmosphere, while the infrared channel provides temperature information on the surface and clouds.

[0040] 3. Lightning locator data, which provide information on the temporal and spatial distribution of lightning activity and are very important for identifying severe convective weather (such as thunderstorms) and their evolution.

[0041] After collecting these data, they are checked for completeness and consistency, and any possible data missing or errors are handled. For example, for geostationary satellite data, if data at certain times are missing, these missing data need to be supplemented by interpolation or other methods to ensure the integrity and continuity of the data set.

[0042] Furthermore, the preset time forecast interval is determined, that is, it is clear how long the model can predict the lightning frequency nowcast in the future. This time interval can be a 6-minute forecast interval within 0 to 2 hours. After determining the preset time forecast interval, the multi-threshold severe convection identification and tracking method is used to continue to identify and track the severe convection process based on the Doppler weather radar echo, including monitoring the changes in the severe convection area, the direction and speed of movement, and the increase or decrease in intensity. Then, according to the identified and tracked severe convection process, the weather radar echo sample data, lightning frequency sample data and geostationary satellite multi-channel sample infrared brightness temperature corresponding to the current weather sample data for the next time period are matched, so as to obtain the weather sample data for the next time period corresponding to the weather sample data.

[0043] In the present invention, the weather sample data of the next period is used as a label of severe convective weather, and the label covers a variety of information related to severe convective weather, such as convective intensity (radar echo), lightning frequency and precipitation intensity. In order to describe the lightning frequency more finely, the present invention divides the lightning frequency data corresponding to the lightning frequency label according to the number of lightning activities in the observation period to obtain the label value of the lightning frequency label in different lightning frequency threshold intervals. For example, 1 to 2 lightnings appear in the observation period, which is a low lightning frequency, indicating that there is almost no or only a small amount of lightning activity in the observation period; 3 to 5 lightnings appear in the observation period, which is a medium lightning frequency, indicating that there are relatively frequent lightning activities in the observation period, but not reaching extreme levels; more than 6 lightnings appear in the observation period, which is a high lightning frequency, indicating that lightning activities are very frequent in the observation period, and may be accompanied by strong convective weather. Correspondingly, the label values ​​of the lightning frequency levels corresponding to the lightning frequency labels in different lightning frequency threshold intervals are also different.

[0044] Furthermore, the input part of each weather sample data (radar echo, infrared brightness temperature, lightning density, etc.) is associated with its corresponding weather sample data of the next period (as a severe convective weather label). During the labeling process, in order to ensure that the training, verification and testing process of the model is not disturbed, the training set, verification set and test set can be strictly divided according to the time order, that is, when dividing the data set, it is necessary to ensure that all sub-datasets (training set, verification set, test set) are completely independent, that is, there is no time overlap or dependency between them, to ensure that the model will not be exposed to any information in the test set or verification set during the training process.

[0045] In the present invention, the hyperparameters of the deep learning model are set, such as learning rate, batch size, number of iterations, etc., and then a suitable loss function is designed according to the task requirements to measure the difference between the model prediction results and the actual observed data. The weights of the model are continuously adjusted during the training process until the performance of the model on the validation set reaches an acceptable level.

[0046] On the basis of the above embodiment, the weather sample data is constructed according to the weather radar echo sample data, the lightning frequency sample data and the geostationary satellite multi-channel sample infrared brightness temperature, including: Performing maximum and minimum value normalization processing on the weather radar echo sample data to obtain normalized weather radar echo sample data; Performing logarithmic transformation and normalization processing on the lightning frequency sample data to obtain normalized lightning frequency sample data; Performing mean-standard deviation normalization processing on the geostationary satellite multi-channel sample infrared brightness temperature to obtain normalized geostationary satellite multi-channel sample infrared brightness temperature; The weather sample data is constructed according to the normalized weather radar echo sample data, the normalized lightning frequency sample data and the normalized geostationary satellite multi-channel sample infrared brightness temperature.

[0047] In the present invention, different methods are used to normalize these data in view of the differences between radar echoes, multi-channel infrared brightness temperature of geostationary satellites, and lightning density (i.e., lightning frequency). Specifically, for weather radar echo sample data, maximum and minimum value normalization is used to normalize the radar echo to between 0 and 1, retaining the relative size and distribution characteristics of the data; for geostationary satellite multi-channel sample infrared brightness temperature, mean-standard deviation normalization is used to calculate the mean and standard deviation of each channel respectively, and normalize them separately, thereby eliminating the differences between channels and helping the model learn the common features between different channels; for lightning frequency sample data, considering that lightning density has obvious long-tail distribution characteristics, logarithmic transformation normalization is used to reduce the impact of extreme high values ​​on model training, making the model more robust and reliable.

[0048] Based on the above embodiment, the deep learning model includes a first encoder, a second encoder, an attention time fusion module and a decoder, wherein: The first encoder is used to extract features from the weather radar echo sample data and the lightning frequency sample data to obtain a first high-dimensional feature; The second encoder is used to extract features of the infrared brightness temperature of the multi-channel samples of the geostationary satellite to obtain a second high-dimensional feature; The attention time fusion module is used to perform feature fusion on the first high-dimensional feature and the second high-dimensional feature based on the attention mechanism to obtain a high-dimensional fusion feature; The decoder is used to predict lightning frequency according to the high-dimensional fusion features.

[0049] Figure 2 The schematic diagram of the structure of the multi-branch coding lightning frequency forecast model provided by the present invention can be referred to Figure 2As shown, in the present invention, in order to solve the problem of inconsistent time and spatial resolutions of weather radar, geostationary satellite and lightning locator observations, different encoding modules are used to extract features, wherein the radar echo and lightning density have the same time and spatial resolution, and the same encoding module (i.e., the first encoder) is used to extract high-dimensional features (i.e., the first high-dimensional features); for geostationary satellite multi-channel infrared brightness temperature observations, a separate encoding module (i.e., the second encoder) is used to extract high-dimensional features (i.e., the second high-dimensional features); then, an attention mechanism (i.e., the attention time fusion module) is introduced to fuse the above high-dimensional features; finally, the above high-dimensional features and spatial features are further fused in the decoding and prediction module (i.e., the decoder), and the coordinated prediction of convective intensity and lightning density is achieved.

[0050] Specifically, in the present invention, the first encoder is mainly responsible for processing weather radar echo sample data and lightning frequency sample data. Since these two types of data have the same temporal and spatial resolutions (corresponding resampling processing has been performed in the early stage), they can be processed by the same encoding module. The first encoder extracts high-dimensional features from the input data through a series of complex operations (such as convolution, pooling, etc.). These features are called first high-dimensional features. The first high-dimensional features not only contain the spatial information of the original data, but also imply the correlation and regularity between the data.

[0051] The second encoder is specifically used to process the multi-channel sample infrared brightness temperature data of geostationary satellites. Since these data have different temporal and spatial resolutions from radar echoes and lightning frequency data, they require a separate encoding module for processing. Similar to the first encoder, the second encoder also extracts high-dimensional features from the input data through a series of operations. These features are called second high-dimensional features. These second high-dimensional features also contain the spatial information of the original data and the correlation between the data, but they reflect more of the characteristics of the infrared brightness temperature data.

[0052] The attention time fusion module is a key component in the model, responsible for fusing the first high-dimensional feature with the second high-dimensional feature. By introducing the attention mechanism, the module can dynamically adjust the importance of different features to optimize the fusion effect. During the fusion process, the attention time fusion module performs weighted summation according to the importance of each feature (i.e., attention weight) to obtain a high-dimensional fusion feature. The high-dimensional fusion feature not only contains all the information of the original data, but also realizes the optimized combination of information through the attention mechanism.

[0053] The decoder is the output part of the model and is responsible for forecasting lightning frequency based on high-dimensional fusion features. The decoder converts high-dimensional fusion features back to the spatial resolution of the original data through a series of inverse operations (such as deconvolution and upsampling), and outputs the lightning frequency forecast results. In this process, the decoder will make full use of the information in the high-dimensional fusion features to achieve coordinated forecasting of convective intensity and lightning density, thereby improving the accuracy of the forecast.

[0054] On the basis of the above embodiment, the deep learning model is trained according to the severe convective weather sample data to obtain the multi-branch coding lightning frequency forecast model, including: The deep learning model is trained according to the severe convective weather sample data until the loss function value in the training process reaches a convergence condition, thereby obtaining the multi-branch coding lightning frequency forecast model, wherein the calculation formula of the loss function value is: in, represents the loss function value, Indicates the loss value corresponding to the weather radar echo, Indicates the loss value corresponding to the lightning frequency, The total number of tags indicating the lightning frequency level, Indicates The forecast value corresponding to the lightning frequency level label, Indicates The label value corresponding to the lightning frequency level label, Represents the loss value between the predicted value and the true label value, Indicates the number of time frames, Indicates Pixel coordinates in the frame radar echo The corresponding measured value is Indicates Pixel coordinates in the frame radar echo The corresponding forecast value is Indicates Pixel coordinates in the frame radar echo The corresponding weight value.

[0055] In the present invention, based on the severe convective weather sample data and deep learning model constructed in the above embodiments, reasonable model hyperparameters are set, and class weighted MAE loss and class imbalance loss DiceLoss are introduced to optimize the convective intensity (radar echo) and lightning density, respectively, to achieve coordinated forecasting of convective intensity (radar echo) and lightning frequency, so that the trained model can finally output kilometer-level and minute-level convective intensity (radar echo) and lightning frequency nowcasting results.

[0056] In the present invention, reference may be made to Figure 2 As shown in the figure, the observation data of the past hour is used as the input of the multi-branch encoding deep learning model, and the radar echo and lightning frequency observations of the next 2 hours are used as the label training model. During the training process, the radar echo and lightning frequency are used as the input of the first encoder to extract the high-dimensional features of radar echo and lightning frequency; then the water vapor and infrared channels of the geostationary satellite are used as the input of the second encoder to extract the high-dimensional features of multi-channel infrared brightness temperature; then, the spatiotemporal attention fusion module is used to further fuse the high-dimensional features extracted by the two encoder branches and use them as the input of the decoder. Finally, the model will output radar echo and lightning frequency at the same time.

[0057] Furthermore, the prediction results of the model and the corresponding labels are used to calculate the corresponding losses. In the present invention, the weighted MAE loss function is used for radar echoes: Different weights are set for radar echoes of different intensities to calculate the loss of radar echoes. The weight information corresponding to the radar echo threshold is as follows: in, Represents the reflectivity factor, which is used to quantify the radar echo strength.

[0058] For lightning frequency, the present invention adopts the DiceLoss category imbalance loss function: in, represents the total number of grid points, For lightning frequency forecast, is the threshold corresponding to the lightning frequency level, is the probability value after conversion by sigmoid function, It is the label of the real lightning frequency level.

[0059] Furthermore, consider the orderliness of lightning frequency levels: in, is the lightning frequency level, For the Grid point The lightning frequency level label value, the total DiceLoss loss is: Finally, the radar echo and lightning frequency losses are directly summed to get the final loss .

[0060] Based on the above embodiment, the method further includes: Based on the radar echo time and spatial resolution corresponding to the weather radar echo sample data, the lightning observation sample data is resampled to obtain the lightning frequency sample data that is aligned with the weather radar echo sample data in time and space, wherein the lightning observation sample data is discrete lightning observation data observed by a lightning locator.

[0061] Figure 3 The overall flow chart of the lightning frequency nowcasting method provided by the present invention can be referred to as follows: Figure 3 As shown, in the present invention, radar echoes, lightning locator data and multi-channel infrared brightness temperature (water vapor channel and infrared channel) of a Doppler weather radar network in a certain area are collected, and then continuous time observations related to typical severe convective processes are extracted from the above variables. Then, the data is quality controlled and continuity checked, and abnormal data in radar echoes, lightning locators and geostationary satellite observations are removed. At the same time, it is checked whether the geostationary satellite observations are missing. If there are missing moments in the middle, the missing data are supplemented by observations at adjacent moments; further, the discrete lightning locator observations are resampled to align the time and space resolution with the radar echo observations.

[0062] Next, the grid data is normalized, where the radar echo is normalized to a value range of 0 to 1 using the maximum and minimum values; the lightning frequency is normalized using logarithmic transformation; the geostationary satellite water vapor and infrared brightness temperature observations are normalized using the mean and standard deviation methods, thereby generating normalized grid data through the above normalization methods. After quality control, gridding and normalization, a sample of severe convective processes is generated according to 1 hour input and 2 hours output (label), where the temporal and spatial resolutions of radar echoes and lightning frequencies are 6 minutes and 1 km each; the temporal and spatial resolutions of geostationary satellite water vapor and infrared channels are 10 minutes and 2 km each.

[0063] In the present invention, the feature extraction module extracts key information from radar echo and lightning density data respectively, and then the high-dimensional feature fusion module fuses and optimizes the features extracted from different data sources. Finally, the decoding and forecasting module uses the fused feature set to perform lightning density nowcasting to generate kilometer-level and minute-level lightning density forecast products.

[0064] The present invention solves the problem of feature extraction and fusion of observations such as weather radar, geostationary satellite and lightning locator with inconsistent temporal and spatial resolutions by designing a multi-branch coding structure attention mechanism model; and realizes the nowcasting of kilometer-level and minute-level lightning frequency by fusing multi-source heterogeneous observations. In addition, the present invention realizes the coordinated forecasting of convective intensity (radar echo) and lightning frequency by introducing a multi-task optimization module, effectively improving the forecasting accuracy of convective intensity (radar echo) and lightning frequency.

[0065] Figure 4 The comparison diagram of the lightning frequency nowcasting method provided by the present invention and the existing lightning forecasting method at different times can be referred to Figure 4 As shown in the figure, the images of multiple variables (i.e., observation data from weather radar, geostationary satellite, and lightning locator) in the past hour are input into the multi-branch coded lightning frequency forecast model. The model will output normalized radar echoes and lightning frequency forecasts with a time resolution of 6 minutes and a spatial resolution of 1 km, and then use the corresponding normalization parameters for denormalization to obtain the final radar echo and lightning frequency forecast. Figure 4 As shown, compared with the existing optical flow method, the present invention can accurately predict the location and intensity of lightning frequency, especially the initial prediction of lightning at subsequent moments, while the optical flow method cannot predict the initial occurrence of subsequent lightning. The present invention improves the prediction accuracy of kilometer-level lightning frequency every 6 minutes from 0 to 2 hours, which has important indicative significance for the monitoring and forecasting of subsequent severe convective processes. Specifically, it is embodied as follows: The lightning frequency in the present invention (≥1 flash / km 2 ) forecasts have improved TS (ThreatScore) scores by 122% compared to existing optical flow methods; for strong lightning frequency (≥8 flash / km 2 ) forecasts have a 16.8% improvement in TS score compared to existing optical flow methods. The lightning frequency of multi-task learning collaborative forecasts (forecasting radar echoes and lightning frequency at the same time) is better than that of single-task learning forecasts (forecasting lightning frequency only), and the lightning frequency (≥1 flash / km 2 ) TS score increased by 5.6%, and the frequency of strong lightning (≥8 flash / km 2 )’s TS score increased by 41.6%.

[0066] The lightning frequency nowcasting system provided by the present invention is described below. The lightning frequency nowcasting system described below and the lightning frequency nowcasting method described above can be referred to each other.

[0067] Figure 5 The schematic diagram of the structure of the lightning frequency nowcasting system provided by the present invention is as follows: Figure 5 As shown, the present invention provides a lightning frequency nowcasting system, including a severe convective data acquisition module 501 and a lightning frequency forecasting module 502, wherein the severe convective data acquisition module 501 is used to obtain the severe convective weather measured data of the target area in the current time period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; the lightning frequency forecasting module 502 is used to input the severe convective weather measured data into a multi-branch encoding lightning frequency forecasting model based on a preset time forecast interval, and obtain the weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch encoding lightning frequency forecasting model, wherein the multi-branch encoding lightning frequency forecasting model is trained based on severe convective weather sample data marked with severe convective weather labels.

[0068] The lightning frequency nowcasting system provided by the present invention realizes the coordinated forecasting of radar echo and lightning frequency by considering the relationship between convection intensity and lightning frequency, thereby improving the accuracy of lightning forecasting.

[0069] The system provided in the embodiment of the present invention is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for the specific process and detailed contents, which will not be repeated here.

[0070] Figure 6 A schematic diagram of the structure of an electronic device provided by the present invention, such as Figure 6 As shown, the electronic device may include: a processor (Processor) 601, a communication interface (Communications Interface) 602, a memory (Memory) 603 and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 complete mutual communication through the communication bus 604. The processor 601 may call the logic instructions in the memory 603 to execute the lightning frequency nowcasting method, which includes: obtaining the severe convective weather measured data of the target area in the current period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; based on the preset time forecast interval, inputting the severe convective weather measured data into the multi-branch coding lightning frequency forecast model, and obtaining the weather radar echo forecast data and lightning frequency forecast data of the target area in the next period output by the multi-branch coding lightning frequency forecast model, wherein the multi-branch coding lightning frequency forecast model is trained based on the severe convective weather sample data marked with severe convective weather labels.

[0071] In addition, the logic instructions in the above-mentioned memory 603 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0072] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the lightning frequency nowcasting method provided by the above-mentioned methods, and the method includes: obtaining the severe convective weather measured data of the target area in the current time period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; based on a preset time forecast interval, inputting the severe convective weather measured data into a multi-branch coded lightning frequency forecast model to obtain the weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecast model, wherein the multi-branch coded lightning frequency forecast model is trained based on severe convective weather sample data marked with a severe convective weather label.

[0073] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the lightning frequency nowcasting method provided in the above-mentioned embodiments, the method comprising: obtaining measured data of severe convective weather in a target area in a current time period, wherein the measured data of severe convective weather at least includes measured weather radar echo data, measured lightning frequency data and multi-channel measured infrared brightness temperature of a geostationary satellite; based on a preset time forecast interval, inputting the measured data of severe convective weather into a multi-branch coded lightning frequency forecasting model, and obtaining weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecasting model, wherein the multi-branch coded lightning frequency forecasting model is trained based on severe convective weather sample data marked with a severe convective weather label.

[0074] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0075] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0076] 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 it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightning frequency nowcasting method, characterized in that: include: Acquire the severe convective weather measured data of the target area in the current period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; Based on the preset time forecast interval, the measured data of severe convective weather is input into the multi-branch coded lightning frequency forecast model to obtain the weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecast model, wherein the multi-branch coded lightning frequency forecast model is trained based on severe convective weather sample data marked with severe convective weather labels.

2. The lightning frequency nowcasting method according to claim 1, characterized in that: The multi-branch coding lightning frequency prediction model is trained by the following steps: Construct weather sample data based on weather radar echo sample data, lightning frequency sample data and geostationary satellite multi-channel sample infrared brightness temperature; Based on the preset time forecast interval, determining weather sample data for the next time period corresponding to the weather sample data; Using the weather sample data of the next period as the severe convective weather label, marking the weather sample data, and obtaining the severe convective weather sample data marked with the severe convective weather label, wherein the lightning frequency label in the severe convective weather label corresponds to different lightning frequency levels; The deep learning model is trained according to the severe convective weather sample data to obtain the multi-branch encoding lightning frequency forecasting model.

3. The lightning frequency nowcasting method according to claim 2, characterized in that: The weather sample data is constructed according to the weather radar echo sample data, the lightning frequency sample data and the geostationary satellite multi-channel sample infrared brightness temperature, including: Performing maximum and minimum value normalization processing on the weather radar echo sample data to obtain normalized weather radar echo sample data; Performing logarithmic transformation and normalization processing on the lightning frequency sample data to obtain normalized lightning frequency sample data; Performing mean-standard deviation normalization processing on the geostationary satellite multi-channel sample infrared brightness temperature to obtain normalized geostationary satellite multi-channel sample infrared brightness temperature; The weather sample data is constructed according to the normalized weather radar echo sample data, the normalized lightning frequency sample data and the normalized geostationary satellite multi-channel sample infrared brightness temperature.

4. The lightning frequency nowcasting method according to claim 2, characterized in that: The deep learning model includes a first encoder, a second encoder, an attention time fusion module and a decoder, wherein: The first encoder is used to extract features from the weather radar echo sample data and the lightning frequency sample data to obtain a first high-dimensional feature; The second encoder is used to extract features of the infrared brightness temperature of the multi-channel samples of the geostationary satellite to obtain a second high-dimensional feature; The attention time fusion module is used to perform feature fusion on the first high-dimensional feature and the second high-dimensional feature based on the attention mechanism to obtain a high-dimensional fusion feature; The decoder is used to predict lightning frequency according to the high-dimensional fusion features.

5. The lightning frequency nowcasting method according to claim 4, characterized in that: The deep learning model is trained according to the severe convective weather sample data to obtain the multi-branch coding lightning frequency forecast model, including: The deep learning model is trained according to the severe convective weather sample data until the loss function value in the training process reaches a convergence condition, thereby obtaining the multi-branch coding lightning frequency forecast model, wherein the calculation formula of the loss function value is: in, represents the loss function value, Indicates the loss value corresponding to the weather radar echo, Indicates the loss value corresponding to the lightning frequency, The total number of tags indicating the lightning frequency level, Indicates The forecast value corresponding to the lightning frequency level label, Indicates The label value corresponding to the lightning frequency level label, Represents the loss value between the predicted value and the true label value, Indicates the number of time frames, Indicates Pixel coordinates in the frame radar echo The corresponding measured value is Indicates Pixel coordinates in the frame radar echo The corresponding forecast value is Indicates Pixel coordinates in the frame radar echo The corresponding weight value.

6. The lightning frequency nowcasting method according to claim 2, characterized in that: The method further comprises: Based on the radar echo time and spatial resolution corresponding to the weather radar echo sample data, the lightning observation sample data is resampled to obtain the lightning frequency sample data that is aligned with the weather radar echo sample data in time and space, wherein the lightning observation sample data is discrete lightning observation data observed by a lightning locator.

7. A lightning frequency nowcasting system, characterized in that: include: A severe convective weather data acquisition module is used to obtain the severe convective weather measured data of the target area in the current period, wherein the severe convective weather measured data at least includes weather radar echo measured data, lightning frequency measured data and geostationary satellite multi-channel measured infrared brightness temperature; The lightning frequency forecast module is used to input the measured severe convective weather data into the multi-branch coded lightning frequency forecast model based on a preset time forecast interval, and obtain the weather radar echo forecast data and lightning frequency forecast data of the target area in the next time period output by the multi-branch coded lightning frequency forecast model, wherein the multi-branch coded lightning frequency forecast model is trained based on severe convective weather sample data marked with severe convective weather labels.

Images