Lightning proximity prediction method and system based on space-time separation convolutional neural network

Abstract

The invention provides a lightning proximity prediction method based on a space-time separation convolutional neural network. The method comprises the following steps: S1, acquiring satellite infrared brightness temperature observation data and lightning positioning data; s2, performing spatial resolution adjustment and standardization processing on the satellite infrared brightness temperature observation data; gridding and de-noising processing is carried out on the lightning positioning data; s3, defining a prediction task; s4, establishing a space-time sequence prediction model; based on the processed data, storing the processed data in a numpy file format according to a time point sequence, dividing the processed data into a training set, a verification set and a test set, and loading the training set, the verification set and the test set to a space-time sequence prediction model for model training to obtain an optimization model; and based on the prediction task, using the optimization model to carry out lightning approaching prediction to obtain lightning prediction data. According to the method, the development of cumulus cloud is captured by using satellite infrared brightness temperature observation data and brightness temperature difference data, so that the problem that thunder and lightning inception cannot be effectively predicted is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of lightning imminent prediction, and in particular to a lightning imminent prediction method and system based on a time-space separation convolutional neural network. Background Art

[0002] Lightning is a natural phenomenon that occurs within cumulonimbus clouds. When the charge distribution within the cloud is uneven and the electric field strength reaches a certain level, discharge occurs. Lightning is often accompanied by strong winds, heavy rain, and hail, significantly impacting the human environment and even threatening human life. Therefore, accurately predicting lightning strikes can reduce the resulting loss of life and property, while also raising public awareness and ensuring public safety. With the advancement of meteorological monitoring, weather radars and geostationary meteorological satellites with high temporal and spatial resolution have become essential tools for observing lightning development. Weather radars analyze the structure and evolution of lightning by measuring the reflectivity of large ice-water mixtures. Meteorological satellites, on the other hand, measure cumulus cloud information using infrared radiation. Based on cloud images, they infer cloud top information and updraft strength to determine whether growing cumulus clouds will form thunderstorms. Consequently, various high-resolution observational data from satellites and radars are widely used in lightning tracking and prediction methods.

[0003] Traditional lightning prediction methods include extrapolation-based lightning tracking and prediction and numerical prediction models. Using thunderstorm-related weather data collected through extensive weather radar networks and lightning location systems, researchers use extrapolation algorithms to analyze and forecast lightning trends. Dixon et al. first proposed the TITAN (Thunderstorn Identification, Tracking, Analysis, and Noecasting) algorithm, which employs a single threshold method to identify thunderstorm cells. This algorithm uses radar data to identify thunderstorms, then matches new thunderstorms that emerge over time with existing ones. Finally, geometric reasoning is used to handle merging and classification. The SCIT (Storm Cell Identification and Tracking) algorithm proposed by Johnson et al. uses seven reflectivity thresholds to identify the centroid of thunderstorm cells. However, because it automatically discards low-threshold identification results, it results in a significant loss of thunderstorm structure information. Because extrapolation algorithms are based on historical data, they cannot provide accurate predictions for sudden and random lightning.

[0004] Numerical forecasting methods for predicting lightning occurrence are based on the mechanism of lightning generation. Numerical models are used to model and calculate various parameters in the atmosphere, providing quantitative lightning prediction results. Price and Rind proposed the famous PR92 scheme, which discovered the relationship between maximum vertical velocity and lightning frequency. They used this parameterization method to simulate the global lightning distribution on global and local data sets, and achieved good experimental performance. Lynn and Yair proposed the Lightning Potential Index (LPI), which is calculated from the maximum mixing ratio of snow, cloud ice, and graupel, but they did not propose a method to directly obtain lightning density. The lightning parameterization method based on numerical forecasting is more effective for medium- and long-term forecasts but is not suitable for nowcasting of lightning, and its parameterization scheme itself has certain errors.

[0005] Machine learning predicts future behavior, outcomes, and trends by identifying features in complex data sets. This approach has brought new research to weather forecasting. By training models to identify the characteristics of thunderstorm and lightning activity, machine learning models are able to predict the occurrence of lightning. Azad et al. proposed a hybrid model to predict the monthly frequency of lightning in Bangladesh. The model uses random forests to select 11 influential predictive features, then uses ensemble empirical mode decomposition to deconstruct the original time series data into a set of functions and residuals. Finally, the selected functions and other parameters are combined and input into an artificial neural network or support vector machine. Rufus et al. used historical lightning and meteorological data from the southern part of Peninsular Malaysia to evaluate the effectiveness of five machine learning algorithms in thunderstorm prediction tasks. The results showed that the use of stratified K-fold cross-validation and gradient boosting models is the best algorithm for thunderstorm prediction in this region.

[0006] However, traditional machine learning methods rely heavily on manual feature extraction. Deep learning methods can address these shortcomings of machine learning because they can autonomously learn and extract meaningful features from complex datasets. For meteorological data with spatiotemporal characteristics, deep learning models use convolutional neural networks and recurrent neural networks to extract spatial and temporal features from the data, respectively. Yangli-Ao et al. constructed a new model, LightNet, using a convolutional LSTM for lightning prediction. Experimental results showed that the model achieved good results in 6-hour lightning prediction. Lu et al. used three-dimensional weather radar data to predict lightning strike zones. They compared CNNs with machine learning methods such as logistic regression and random forests, demonstrating that CNNs performed better. Cintineo used GOES-16 satellite data to predict lightning and proposed the LightningCast model to learn the relationship between lightning and satellite imagery. Experimental results demonstrated the potential advantages of satellite data in lightning prediction.

[0007] Despite extensive research on lightning prediction, utilizing deep learning models for lightning prediction remains challenging. First, traditional deep learning models in meteorological forecasting often utilize recurrent neural networks to learn the temporal characteristics of meteorological data. However, cumulative errors in recurrent neural networks often lead to poor prediction results. Second, current spatiotemporal series prediction models based on fully convolutional neural networks only consider the changing trends of different frames in the temporal dimension, which can lead to noise or information discontinuity. Therefore, prediction accuracy still needs to be improved. Finally, most deep learning-based lightning predictions directly utilize all observed data without performing any analysis or selection, which can lead to the model being more sensitive to noisy or redundant data. Summary of the Invention

[0008] This paper provides a lightning nowcasting method and system based on a spatiotemporal separation convolutional neural network to address the technical problem of poor accuracy in lightning nowcasting. Based on a fully convolutional neural network (STFNet) that extracts spatiotemporal features, and utilizing high-resolution satellite infrared brightness temperature observations and lightning location data, accurate lightning nowcasting is achieved. By parallelizing the spatial long-range dependency information of a single frame in the time and channel dimensions and the motion characteristics between different frames, the designed temporal fusion block (TFBlock) processes the proposed model. This model can more completely model the temporal evolution of lightning, and by integrating the spatial features of the observation data, it significantly improves the accuracy of lightning nowcasting. Furthermore, the present invention uses satellite infrared brightness temperature observations and brightness temperature difference data to capture the development of cumulus clouds, thereby addressing the problem of ineffective prediction of lightning initiation.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] The lightning approach prediction method based on time-space separation convolutional neural network includes:

[0011] S1. Obtain satellite infrared brightness temperature observation data and lightning location data;

[0012] S2. Adjusting the spatial resolution and standardizing the satellite infrared brightness temperature observation data; and gridding and denoising the lightning location data;

[0013] S3. Define the prediction task: select data from the past N hours to predict the location of lightning in the next N hours;

[0014] S4. Establish a spatiotemporal sequence prediction model with an encoding-decoding structure based on the time fusion block TFBlock. Based on the processed satellite infrared brightness temperature observation data and lightning location data, save them in numpy file format in time order, divide them into training set, validation set and test set in a ratio of 7:1:2, load them into the spatiotemporal sequence prediction model for model training, and obtain the optimized model. Based on the prediction task, use the optimized model to perform lightning impending prediction and obtain lightning prediction data.

[0015] In this manual, satellite infrared brightness temperature observation data includes data from the water vapor channel and cumulus information-related channels recorded by the Sunflower-8 meteorological satellite, as well as dual-channel or triple-channel brightness temperature difference information; lightning location data includes scattered data obtained by measuring the time, intensity, longitude, latitude, and altitude of lightning using a lightning locator.

[0016] In this specification, in S2, the required spatial range is intercepted from the longitude and latitude range of the original satellite infrared brightness temperature observation data, and the spatial range is the starting longitude 106.4°E, the ending longitude 110.4°E, the starting latitude 21.4°N, and the ending latitude 25.4°N; and the original spatial resolution is reduced to 0.04°*0.04°. Specifically, the original grid data is traversed by a submatrix with a step size of 2, and the maximum value in the submatrix is taken as the new grid data. After the original grid data is traversed, the size of the new matrix is 1 / 2 of the original matrix.

[0017] In this specification, in S2, the brightness temperature value range of the channel observation data of the satellite infrared brightness temperature observation data is adjusted to (0, 70) K to accelerate the convergence speed of the model and improve the training efficiency; and the following formula is used for standardization;

[0018]

[0019] Among them, m1 and m2 represent the maximum and minimum values of the original values, n1 and n2 represent the maximum and minimum values of the standardized values, x is the original value, and y is the standardized value. The original brightness temperature value x is processed to obtain the standardized value y.

[0020] In this specification, in S2, the longitude, latitude, and time information in the lightning location data are extracted one by one, all time points are grouped and superimposed at 10-minute intervals, and the longitude and latitude at the time of the lightning occurrence are mapped to a grid to obtain the final lightning gridded data. The calculation formula for the longitude and latitude mapping is as follows:

[0021]

[0022] Among them, Lat_proj and Lon_proj represent the row and column values of the latitude and longitude values after being projected onto the grid, respectively, res represents the spatial resolution, Lon is longitude, Lat is latitude, Lat_end is the end latitude, and Lon_start is the start longitude.

[0023] In this manual, in S2, based on the Band 13 infrared brightness temperature observation channel, 235K is selected as the brightness temperature threshold to identify lightning noise points. The location (x, y) of the lightning occurrence is first obtained from the lightning grid data. The channel brightness temperature difference data BTD1513 and TTD, which represent cloud optical thickness and cloud top phase, are used to filter out cloud-free areas and cirrus objects. Then, it is determined whether there is data with a Band 13 brightness temperature value lower than 235K within an 8km range. If so, the lightning recorded at this point is determined to be real lightning rather than noise. Otherwise, the value at this point is set to 0.

[0024] In this specification, in S3, the location of lightning in the next hour is predicted based on the satellite infrared brightness temperature observation data and lightning location data of the past hour. The current time is t, and the input data of the past "i" frame and the output result of the future "j" frame are defined as:

[0025]

[0026] Among them, [S T ] and [L T ] represent the input satellite infrared brightness temperature observation data and lightning location data for the past hour, F(;θ) represents the model and the learnable parameter θ. Represents the output flash forecast value for the next hour.

[0027] In this specification, S4 includes:

[0028] S4-1,STFNet model establishment;

[0029] The STFNet model is a spatiotemporal sequence prediction model with an encoding-decoding structure built based on the time fusion block TFBlock. It uses pure convolution operations to separate and extract the spatiotemporal features of meteorological data to more accurately capture the underlying patterns of the data.

[0030] S4-1-1, construct model space coding structure;

[0031] The encoder part of the model is composed of N s It is composed of a series of (Conv2d+GroupNorm+SiLU) blocks to extract high-dimensional spatial feature information of the data; the input is X i-1 , F i is the output of the encoder, then the spatial encoding process is expressed as: F i=SiLU(GroupNorm(Conv2d(X i-1 )));

[0032] S4-1-2, build model time modeling structure;

[0033] As the basic unit for capturing the temporal characteristics of lightning, TFBlock consists of two core modules: the spatiotemporal processing module and the channel processing module. These modules respectively extract the shape and structural features of lightning in a single frame and learn the changing process of lightning between different frames. The two modules process the temporal evolution of lightning in parallel, combining the static and motion characteristics of lightning. The output results of the two modules are then element-wise multiplied with the input data, and then subjected to a 1x1 convolution to fuse the features. Finally, a residual connection is made with the input data that has not been processed by TFBlock to improve the feature extraction effect.

[0034] The spatiotemporal processing module is responsible for capturing the static features of lightning to understand the interdependencies between distant pixels. After reducing the spatial dimension using average pooling, the module uses a 1x1 convolution to obtain local region features. Two depthwise convolutions are then used to approximate standard large-kernel convolutions to reduce the number of parameters. The two convolution kernels are 1x11 and 11x1, respectively. By learning a wider range of pixel dependencies within a large receptive field, the module captures a more accurate lightning structure. Finally, a 1x1 convolution and a sigmoid activation function are used to generate attention weights to further improve the output of the spatiotemporal processing module.

[0035] The channel processing module captures the motion characteristics of lightning by learning channel weights between different frames in the time and channel dimensions. The channel processing module performs average pooling and maximum pooling operations respectively. Maximum pooling is used to focus on the salient features of the image. Average pooling compresses the spatial dimensions of the input feature map to preserve the overall image characteristics. Both pooling operations are used simultaneously to enhance the characteristic expression of the lightning motion process. The two features after the pooling operation are respectively sent to a shared fully connected layer. Finally, the output channel features are added based on the element-wise operation to obtain the final channel attention weight.

[0036] The spatiotemporal processing module is represented as:

[0037] A j =sigmoid(conv2d(DWconv(DWconv(av□_pool(F j-1 )))));

[0038] The channel processing module is represented as:

[0039]

[0040] Among them, F j-1 is the output of the spatial encoding part, Indicates a connection operation;

[0041] S4-1-3, construct the model's spatiotemporal decoding structure;

[0042] The model's spatiotemporal decoder fuses the encoder's spatial feature map with the temporal features of TFBlock to predict future lightning development and, in the process, reconstruct the resolution of lightning features. The number of decoding blocks in the decoder is the same as that in the encoder. Each block includes unConv2d, GroupNorm, and SiLU, expressed as:

[0043] Z k =SiLU(GroupNorm(unConv2d(F k-1 ,T k-1 )));

[0044] Among them, F k-1 and T k-1 The outputs of the encoder and TFBlock are respectively, using ConvTransposed2d as the unConv2d operator;

[0045] S4-1-4, build a complete model;

[0046] The input data first passes through the model's spatial encoding structure to extract high-dimensional spatial features. The feature map output by the encoding structure is then input into the model's temporal modeling structure TFBlock in S4-1-2 to learn the temporal evolution of lightning. The encoder's feature map and the features of TFBlock are then input into the spatiotemporal decoder for decoding. Finally, a convolution operation is performed to obtain an output of the same size as the input, and a binary classification prediction is performed on the grid points of the output feature map.

[0047] S4-2, load the dataset;

[0048] The processed satellite infrared brightness temperature observation data and lightning location data in S2 are saved in numpy file format in time order. When the model loads the dataset, it is divided into training, validation, and test sets in a ratio of 7:1:2. The training set is used to learn the patterns and weights in the data. The validation set is used to verify or adjust the model's hyperparameters during training to prevent overfitting and select the model with the best training results as the final model. The test set is used to test the model's final performance and evaluate its generalization ability. The satellite data and lightning data from the previous hour are read from the dataset by traversing the time points and input into the model.

[0049] S4-3, training model;

[0050] S4-3-1, parameter setting;

[0051] A fixed random seed is set during model training to ensure the repeatability of the model training results. During the training process, the AdamW optimizer is used, and the learning rate is set to 10 -4 The rest of the settings are default values; the batch size is set to 16, the iteration cycle is set to 1000, and the early stopping strategy is adopted. When the validation set loss stops decreasing for more than 10 consecutive iteration cycles and the model score on the validation set no longer improves, the training is stopped; when traversing the dataset, the batch size is loaded in batches of 16 data each time, all the data of the previous hour are spliced into a tensor by frame and fed into the model to obtain the predicted data after one hour;

[0052] S4-3-2, loss function;

[0053] The Focal loss function is used as the loss function for model training. When the data categories are unbalanced, the Focal loss function assigns different weights to samples according to the difficulty of successful sample classification. The Focal loss function is expressed as:

[0054] FL(P,y)=-α t (1-P t ) γ log(P t );

[0055] Where P is the probability that lightning is correctly classified, (1-P t ) γ is the loss adjustment factor, α is the weighting coefficient;

[0056] The predicted data and the real data are input into the loss function to calculate the average loss of the predicted results, and then the model parameters are updated through back propagation;

[0057] S4-4, Model Validation and Comparison;

[0058] After the model has been trained with one round of training data, the data in the validation set are traversed and input into the model in the same way to obtain lightning prediction data. The number of hits (NA), null reports (NB), missed reports (NC), and non-reports (ND) in the lightning grid data are counted. pre and I act They represent the model prediction output and the real lightning data respectively. The statistical formula of each data is expressed as:

[0059] NA=∑(I pre =1∧I act =1);

[0060] NB=∑(I pre =1∧I act =0);

[0061] NC=∑(I pre =0∧I act =1);

[0062] ND=∑(I pre =0∧I act =0);

[0063] After the statistics are completed, the verification indicators of this round of verification set will be calculated: threat score TS, fair threat score ETS, false alarm rate FAR and hit rate POD; the calculation formulas for each evaluation indicator are as follows:

[0064]

[0065]

[0066]

[0067]

[0068] in,

[0069] Finally, the model with the highest validation set score was selected and compared on the test set data with four classic models in spatiotemporal sequence prediction: the convolutional long short-term memory network ConvLSTM, the spatiotemporal sequence prediction model PredRNN based on the ST-LSTM unit, the 3D image segmentation model 3DUnet improved based on unet, and the latest spatiotemporal sequence prediction model Simvp based entirely on CNN. The TS, ETS, FAR, and POD scores of each model were calculated to compare the performance of the STFNet model in lightning nowcasting.

[0070] A lightning approach prediction system based on a time-space separation convolutional neural network is used to implement any of the above-mentioned lightning approach prediction methods based on a time-space separation convolutional neural network. The lightning approach prediction system based on a time-space separation convolutional neural network includes:

[0071] Data acquisition module, used to obtain satellite infrared brightness temperature observation data and lightning location data;

[0072] A data processing module is used to adjust the spatial resolution and standardize the satellite infrared brightness temperature observation data; and to grid and denoise the lightning location data;

[0073] The task module is used to define the prediction task: select the data of the past N hours to predict the location of lightning in the next N hours;

[0074] The prediction module is used to build a spatiotemporal series prediction model with an encoding-decoding structure based on the time fusion block TFBlock. Based on the processed satellite infrared brightness temperature observation data and lightning location data, it is saved in numpy file format in time point order, divided into training set, validation set, and test set in a ratio of 7:1:2, and loaded into the spatiotemporal series prediction model for model training to obtain the optimized model.

[0075] Based on the prediction task, the optimization model is used to perform lightning forecasting and obtain lightning prediction data.

[0076] In summary, the present invention has at least the following beneficial effects:

[0077] In order to improve the accuracy of lightning nowcasting while effectively modeling the evolution of lightning in the time dimension, the present invention proposes a spatiotemporal sequence prediction model STFNet based on the time fusion block TFBlock to achieve accurate lightning nowcasting at high spatiotemporal resolution. The model adopts full convolution operation, which has a simple structure, low computational overhead, and avoids the cumulative error problem that may exist when using recurrent neural networks. On the basis of spatiotemporal feature separation and extraction, the model processes the long-range dependency information of a single frame and the change characteristics between frames in parallel in the time and channel dimensions, integrates the shape characteristics and motion trends of lightning, and better learns the temporal evolution law of lightning. Compared with traditional deep learning models in the field of meteorological forecasting, STFNet not only has a significant improvement in prediction accuracy, but also can better process satellite cloud images with high spatiotemporal resolution, which is crucial for the field of lightning nowcasting. By using the proposed model, the accuracy of lightning forecasting is greatly improved, especially the forecast performance of lightning in the initial stage of lightning is far better than other deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0079] Figure 1 Schematic diagram of the lightning approach prediction method based on time-space separation convolutional neural network involved in the present invention.

[0080] Figure 2 This is a schematic diagram illustrating an example of the resolution of satellite data involved in the present invention.

[0081] Figure 3 Schematic diagram of the lightning data denoising process involved in the present invention.

[0082] Figure 4 Schematic diagram of the temporal modeling structure of the STFNet model involved in the present invention.

[0083] Figure 5 Schematic diagram of the overall structure of the STFNet lightning nowcasting model involved in the present invention. DETAILED DESCRIPTION

[0084] Hereinafter, only certain exemplary embodiments are briefly described. As will be appreciated by those skilled in the art, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and description are to be regarded as illustrative in nature and not restrictive.

[0085] The disclosure below provides many different embodiments or examples for implementing different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the components and configurations of specific examples are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. In addition, the embodiments of the present invention may repeat reference numerals and / or reference letters in different examples. Such repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or configurations discussed.

[0086] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0087] like Figure 1 As shown, this embodiment provides a lightning approach prediction method based on a spatiotemporal separation convolutional neural network, including:

[0088] S1. Obtain satellite infrared brightness temperature observation data and lightning location data;

[0089] S2. Adjusting the spatial resolution and standardizing the satellite infrared brightness temperature observation data; and gridding and denoising the lightning location data;

[0090] S3. Define the prediction task: select data from the past N hours to predict the location of lightning in the next N hours;

[0091] S4. Establish a spatiotemporal sequence prediction model with an encoding-decoding structure based on the time fusion block TFBlock. Based on the processed satellite infrared brightness temperature observation data and lightning location data, save them in numpy file format in time point order, divide them into training set, validation set and test set in a ratio of 7:1:2, load them into the spatiotemporal sequence prediction model for model training, and obtain the optimized model.

[0092] Based on the prediction task, the optimization model is used to perform lightning forecasting and obtain lightning prediction data.

[0093] This embodiment provides a lightning approach prediction system based on a time-space separation convolutional neural network, which is used to implement the above-mentioned lightning approach prediction method based on a time-space separation convolutional neural network. The lightning approach prediction system based on a time-space separation convolutional neural network includes:

[0094] Data acquisition module, used to obtain satellite infrared brightness temperature observation data and lightning location data;

[0095] A data processing module is used to adjust the spatial resolution and standardize the satellite infrared brightness temperature observation data; and to grid and denoise the lightning location data;

[0096] The task module is used to define the prediction task: select the data of the past N hours to predict the location of lightning in the next N hours;

[0097] The prediction module is used to build a spatiotemporal series prediction model with an encoding-decoding structure based on the time fusion block TFBlock. Based on the processed satellite infrared brightness temperature observation data and lightning location data, it is saved in numpy file format in time point order, divided into training set, validation set, and test set in a ratio of 7:1:2, and loaded into the spatiotemporal series prediction model for model training to obtain the optimized model.

[0098] Based on the prediction task, the optimization model is used to perform lightning forecasting and obtain lightning prediction data.

[0099] To further illustrate the technical solution of the present invention, the technical concept of the present invention is as follows:

[0100] S1. Data selection

[0101] The data used in the present invention include satellite infrared brightness temperature observation data and lightning location data. Satellite infrared brightness temperature observation data is data from the water vapor channel and cumulus information related channel recorded by the Sunflower-8 meteorological satellite. Dual-channel or triple-channel brightness temperature difference information can also represent the physical properties of cumulus growth, including cloud top thickness, convective cloud height and rising intensity, etc. Therefore, the present invention newly constructs four channel combination data combined with channel observation data to achieve more accurate lightning imminent prediction. The original time resolution of the satellite data is 10 minutes, and the spatial resolution is 0.02°*0.02°. The lightning location data is scattered data obtained by measuring the time (T), intensity, longitude (Lon), latitude (Lat), altitude and other information of lightning using a lightning locator, where the time resolution is at the minute level. The data information used is described in Table 1.

[0102] Table 1 Selected satellite and lightning data information

[0103]

[0104]

[0105] S2. Data Processing

[0106] S2-1. Satellite data processing

[0107] S2-1-1. Adjustment of satellite data spatial resolution

[0108] The spatial range required by the present invention is intercepted from the longitude and latitude range of the original satellite data (starting longitude Lon_start = 106.4°E, ending longitude Lon_end = 110.4°E; starting latitude Lat_start = 21.4°N, ending latitude Lat_end = 25.4°N). Since lightning is sudden and random, it will cause sparse and unbalanced lightning in the grid data. In order to reduce the negative impact of such problems, we reduce the original spatial resolution to 0.04°*0.04°. The main operation is to traverse the original grid data with a submatrix with a step size of 2, and then take the maximum value in the submatrix as the new grid data. The operation example is as follows Figure 2 After traversing the original grid data, the new matrix size is 1 / 2 of the original matrix, and all lightning locations will be relatively dense.

[0109] S2-1-2. Satellite data standardization

[0110] The brightness temperature range of satellite channel observation data is (180, 310) K. Lower values indicate stronger convection and a greater probability of lightning. Because this range differs significantly from the values of lightning location data, it can slow model convergence or even prevent convergence. Therefore, we normalize the satellite data to (0, 70) K to accelerate model convergence and improve training efficiency. The normalization formula is:

[0111]

[0112] Among them, m1 and m2 represent the maximum and minimum values of the original values, n1 and n2 represent the maximum and minimum values of the normalization, and the original brightness temperature value x is processed to obtain the normalized value y.

[0113] S2-2. Lightning Location Data Processing

[0114] S2-2-1. Gridding of Lightning Location Data

[0115] The longitude, latitude and time information in the lightning location data are taken out one by one, and all time points are grouped and superimposed at intervals of 10 minutes, and the longitude and latitude when the lightning occurs are mapped to the grid. Specifically, based on the time resolution of the satellite data, the one-hour time range is first divided into 6 groups, each group is ten minutes apart, and the time of lightning occurrence corresponds to each group one by one. Then, the read longitude and latitude information is projected to the corresponding grid according to the spatial resolution modified in S2-1-1. The lightning data of the corresponding time period at the same grid point are superimposed to obtain the final lightning gridded data. The calculation formula for longitude and latitude projection is as follows:

[0116]

[0117] Among them, Lat_proj and Lon_proj respectively represent the row and column values of the latitude and longitude values after being projected onto the grid, and res represents the spatial resolution.

[0118] S2-2-2. Lightning Location Data Denoising

[0119] The monitoring of the lightning locator may be interfered with and produce noise data, so the lightning grid data obtained by S2-2-1 needs to be denoised. Previously, lightning discrimination based on satellite brightness temperature observation data usually selected a brightness temperature threshold between 221-250K. Based on previous studies and combined with the characteristics of the designated area, the present invention selects 235K as the brightness temperature threshold based on the Band13 infrared brightness temperature observation channel to discriminate lightning noise points. First, the location (x, y) where the lightning occurred is obtained from the lightning grid data, and the channel brightness temperature difference data BTD1513 and TTD representing the cloud optical thickness and cloud top phase are used to filter out cloudless areas and cirrus objects, and then it is determined whether there is data with a Band13 brightness temperature value lower than 235K within a range of 8Km. Only when all conditions are met, it is determined that the lightning recorded at the point is real lightning rather than noise, otherwise the value of the point is set to 0. The specific discrimination process is as follows Figure 3 shown.

[0120] S3, task definition

[0121] In this paper, we predict the location of lightning within the next hour based on real-world meteorological observation data from the past hour. Lightning grid data provides essential information for predicting future lightning motion and evolution trends, while satellite infrared observations provide rich cumulus cloud information about the lightning development process, particularly during its incipient stages. Assuming the current time is t, the input data for the past "i" frame and the output for the next "j" frame are defined as:

[0122]

[0123]

[0124]

[0125] Among them, [S T ] and [L T ] represent the input satellite data and lightning data of the past hour respectively, F(;θ) represents the model and the learnable parameters θ, Represents the output flash forecast value for the next hour.

[0126] S4. Model construction

[0127] S4-1. Model establishment

[0128] The STFNet model is a spatiotemporal series prediction model with an encoder-decoder architecture based on the time fusion block TFBlock. The model uses pure convolutional operations to separate and extract the spatiotemporal features of meteorological data, enabling it to more accurately capture the underlying patterns of the data.

[0129] S4-1-1. Constructing the model space coding structure

[0130] The encoder part of the model is composed of N s The (Conv2d+GroupNorm+SiLU) blocks are connected in series to extract the high-dimensional spatial feature information of the data. Assume that the input is X i-1 , F i is the output of the encoder, then the spatial encoding process can be expressed as:

[0131] F i =SiLU(GroupNorm(Conv2d(X i-1 ))) (7)

[0132] S4-1-2. Constructing model time modeling structure

[0133] As the basic unit for capturing lightning temporal features, TFBlock is designed to include two core modules: STmodule (time-space processing module) and TCmodule (channel processing module). These two modules respectively extract the shape and structural features of lightning in a single frame and learn the changing process of lightning between different frames. Figure 3 As shown in (a), two modules process the temporal evolution of lightning in parallel. The outputs of the two modules are then element-wise multiplied with the input data before undergoing a 1x1 convolution to fuse the features. This combination of static and motion features gives the model a more comprehensive understanding of lightning development, providing rich information for more accurate subsequent predictions. Finally, a residual connection is made with the input data that has not been processed by TFBlock to improve feature extraction.

[0134] The STmodule is mainly responsible for capturing the static characteristics of lightning, aiming to grasp the interdependence between distant pixels. Figure 3 As shown in (b), the ST module uses average pooling to reduce spatial dimensionality and then employs 1x1 convolution to obtain local features. Two depthwise convolutions are then used to approximate standard large-kernel convolutions, reducing the number of parameters. The two convolution kernels, 1x11 and 11x1, respectively, learn broader pixel dependencies within a large receptive field, thereby capturing more accurate lightning structure. Finally, a 1x1 convolution and a sigmoid activation function are used to generate attention weights, further enhancing the ST module's output.

[0135] The TCmodule captures the motion characteristics of lightning by learning the channel weights between different frames in the time and channel dimensions, such as Figure 4 As shown in (c), the TC module performs average pooling and max pooling. Average pooling compresses the spatial dimensions of the input feature map, preserving the overall image characteristics. Max pooling focuses on salient image features, facilitating the inference of important information or features in key channels. Using both pooling operations simultaneously enhances the representation of lightning motion. The two pooled features are then fed into a shared fully connected layer. Finally, the output channel features are subjected to an element-wise addition operation to obtain the final channel attention weights.

[0136] STmodule can be expressed as:

[0137] A j =sigmoid(conv2d(DWconv(DWconv(avg_pool(F j-1 )))))(8)

[0138] TCmodule can be expressed as:

[0139]

[0140] Among them, F j-1 is the output of the spatial encoding part, Represents a join operation.

[0141] S4-1-3. Constructing the model's spatiotemporal decoding structure

[0142] The model's spatiotemporal decoder fuses the encoder's spatial feature maps with the temporal features of TFBlock to predict future lightning development and, in the process, reconstruct the resolution of lightning features. The number of decoding blocks in the decoder is the same as in the encoder, and each block includes unConv2d, GroupNorm, and SiLU, which can be expressed as:

[0143] Z k =SiLU(GroupNorm(unConv2d(F k-1 ,T k-1 )))(10)

[0144] Among them, F k-1 and T k-1 are the outputs of the encoder and TFBlock respectively, and we use ConvTransposed2d as the unConv2d operator.

[0145] S4-1-4. Build a complete model

[0146] The input data first passes through the S4-1-1 model spatial encoding structure to extract high-dimensional spatial features, and then the feature map output from the encoding structure is input into the S4-1-2 model temporal modeling structure TFBlock to learn the temporal evolution of lightning. The encoder feature map and the features of TFBlock are then input into the spatiotemporal decoder for decoding. Finally, a convolution operation is performed to obtain an output of the same size as the input, and a binary classification prediction is performed on the output feature map grid. The STFNet model structure is as follows: Figure 5 shown.

[0147] S4-2. Loading the dataset

[0148] The satellite data and lightning location data obtained from S2 are saved in numpy file format in time order. When the model loads the dataset, it is divided into training, validation, and test sets in a ratio of 7:1:2. The training set allows the model to learn the patterns and weights in the data. The validation set is used to verify or adjust the model's hyperparameters during training to prevent overfitting and select the best trained model as the final model. The test set is used to test the model's final performance and help evaluate its generalization ability. The satellite data and lightning data for the hour before a specific time point are read from the specified dataset by traversing the time points and input into the model.

[0149] S4-3. Training Model

[0150] S4-3-1. Parameter setting

[0151] A fixed random seed is set during model training to ensure the repeatability of the model training results. During the training process, the AdamW optimizer is used, and the learning rate is set to 10 -4 The rest of the settings are default. The batch size is set to 16, the iterations are set to 1000, and the early stopping strategy is used. Training stops when the validation set loss stops decreasing for more than 10 consecutive iterations and the model score on the validation set no longer improves.

[0152] When traversing the dataset in S4-2, a batch size of 16 data is loaded each time. All data from the previous hour are concatenated into a tensor by frame and then fed into the model to obtain the predicted data for one hour later.

[0153] S4-3-2. Loss Function

[0154] Lightning itself is instantaneous and random, resulting in a significant imbalance in the number of lightning and non-lightning samples. Therefore, this paper uses the Focal Loss function as the loss function for model training. When the data categories are unbalanced, the Focal Loss function assigns different weights to samples based on the difficulty of successful classification, thereby solving the problem of extremely unbalanced sample numbers in lightning classification tasks. The Focal Loss function is expressed as:

[0155] FL(P,y)=-α t (1-P t ) γ log(P t )(11)

[0156] Where P is the probability that lightning is correctly classified, (1-P t ) γ is the loss adjustment factor and α is the weighting coefficient.

[0157] Input the predicted data and the real data into the loss function in S4-3-2 to calculate the average loss of the predicted results, and then update the model parameters through back propagation.

[0158] S4-4. Model Verification and Comparison

[0159] After the model is trained with a round of training data, the data in the validation set are traversed and input into the model in the same way to obtain lightning prediction data. The number of hits (NA), null reports (NB), missed reports (NC) and non-reports (ND) in the lightning grid data are counted. pre and I act Represent the model prediction output and the actual lightning data respectively, then the statistical formula of each data can be expressed as:

[0160] NA4=∑(I pre =1∧I act =1) (12)

[0161] NB=∑(I pre =1∧I act =0) (13)

[0162] NC=∑(I pre =0∧I act =1) (14)

[0163] ND=∑(I pre =0∧I act =0) (15)

[0164] After the statistics are completed, the validation indicators of this round of validation set will be calculated: Threat Score (TS), Fair Threat Score (ETS), False Alarm Rate (FAR), and Probability of Detection (POD). The TS score is a comprehensive score of the model's prediction results, with larger values indicating better prediction results. ETS is an improvement on TS that penalizes false alarms and missed alarms, making the scoring more fair. FAR represents statistics for lightning forecasts that did not actually occur, while POD represents statistics for lightning that occurred both in the forecast and in reality. The calculation formulas for each evaluation indicator are as follows:

[0165]

[0166]

[0167]

[0168]

[0169] in,

[0170] Finally, we selected the model with the highest validation score and compared it with four classic models for spatiotemporal sequence prediction on the test data: the Convolutional Long Short-Term Memory (ConvLSTM) network; the spatiotemporal sequence prediction model PredRNN based on ST-LSTM units; the 3D image segmentation model 3DUnet, which is improved based on Unet; and the latest spatiotemporal sequence prediction model Simvp, which is based entirely on CNNs. We calculated the TS, ETS, FAR, and POD scores for each model and compared the performance of the STFNet model in lightning nowcasting. The evaluation results for each model are shown in Table 2.

[0171] Table 2 Comparison of prediction scores of STFNet and other models

[0172] Model TS ETS FAR POD ConvLSTM 0.0967 0.0959 0.8484 0.2106 PredRNN 0.0761 0.0737 0.7913 0.1670 3DUnet 0.1173 0.1166 0.8121 0.2418 Simvp 0.1244 0.1236 0.8216 0.2911 STFNet 0.1304 0.1296 0.8127 0.3002

[0173] The key points of the present invention are:

[0174] (1) Data denoising method

[0175] Lightning locators may generate noisy data due to interference during monitoring. Directly inputting this noise into model learning can severely impact model performance and accuracy. The lightning data denoising method proposed in this paper utilizes satellite channel brightness temperature data and channel brightness temperature difference data to rigorously screen for non-real lightning within satellite cloud images. This combined satellite channel data filters out cloudless and cirrus cloud conditions, then determines whether the brightness temperature observation within a specified range is below a selected threshold to determine if the lightning is real. This dual discrimination method more accurately identifies real lightning, ensuring more reliable lightning data and providing more accurate information for subsequent lightning predictions.

[0176] (2) Model building method

[0177] To address the challenges of poor lightning forecasting accuracy, low spatiotemporal resolution, and inability to effectively predict lightning onset, we designed a fully convolutional neural network (STFNet) that separates and extracts spatiotemporal features. This network leverages high-resolution satellite and lightning data to achieve accurate lightning nowcasting. After convolving multiple spatial features, the model uses a temporal fusion block (TFBlock) to model the temporal evolution of lightning. This model captures both the morphology and structure of lightning while focusing on its motion, significantly improving the accuracy of the model's predictions of lightning strike locations.

[0178] The above embodiments are intended to illustrate the present invention, not to limit the present invention. Therefore, changes in illustrative values or substitutions of equivalent components should still fall within the scope of the present invention.

[0179] From the above detailed description, it will be clear to those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0180] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as covering the preferred embodiments and all changes and modifications that fall within the scope of the invention. The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

[0181] It should be noted that the above description of the relevant processes is for illustration and purpose only and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the processes under the guidance of this specification. However, such modifications and changes are still within the scope of this specification.

[0182] The basic concepts have been described above. It will be apparent to those skilled in the art after reading this application that the above disclosures are merely illustrative and do not constitute limitations on this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and amendments to this application. Such modifications, improvements, and amendments are suggested in this application and remain within the spirit and scope of the exemplary embodiments of this application.

[0183] At the same time, this application uses specific terms to describe the embodiments of this application. For example, "one embodiment," "an embodiment," and / or "some embodiments" refer to a certain feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "one embodiment," "an embodiment," or "an alternative embodiment" mentioned twice or more in different places in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application may be appropriately combined.

[0184] In addition, it will be understood by those skilled in the art that various aspects of the present application can be illustrated and described by a number of patentable categories or situations, including any new and useful combination of processes, machines, products or substances, or any new and useful improvements thereto. Therefore, various aspects of the present application can be implemented entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software can all be referred to as "units", "modules" or "systems". In addition, various aspects of the present application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0185] The computer program code required for the operation of each part of the application can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, conventional procedural programming languages such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages such as Python, Ruby and Groovy or other programming languages. The program code can be run completely on the user's computer, or run on the user's computer as an independent software package, or run partly on the user's computer and partly on a remote computer, or run completely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or be connected to an external computer (such as by the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0186] In addition, unless expressly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some embodiments of the invention that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0187] Similarly, it should be noted that in order to simplify the presentation of this disclosure and thereby facilitate understanding of one or more of the invention's embodiments, the foregoing descriptions of the embodiments of this disclosure sometimes combine multiple features into a single embodiment, figure, or description thereof. However, this approach should not be interpreted as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject matter of the invention may possess fewer features than the single embodiment described above.

Claims

1. A lightning approach prediction method based on spatiotemporal separation convolutional neural network, characterized by: include: S1. Obtain satellite infrared brightness temperature observation data and lightning location data; S2. performing spatial resolution adjustment and standardization processing on the satellite infrared brightness temperature observation data; Gridding and denoising of lightning location data; S3. Define the prediction task: select data from the past N hours to predict the location of lightning in the next N hours; S4. Establish a spatiotemporal sequence prediction model with an encoding-decoding structure based on the time fusion block TFBlock. Based on the processed satellite infrared brightness temperature observation data and lightning location data, save them in numpy file format in time point order, divide them into training set, validation set and test set in a ratio of 7:1:2, load them into the spatiotemporal sequence prediction model for model training, and obtain the optimized model. Based on the prediction task, the optimization model is used to perform lightning forecasting and obtain lightning prediction data.

2. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 1 is characterized in that: Satellite infrared brightness temperature observation data include data from the water vapor channel and cumulus information-related channels recorded by the Sunflower-8 meteorological satellite, as well as dual-channel or triple-channel brightness temperature difference information; lightning location data include scattered data obtained by measuring the time, intensity, longitude, latitude, and altitude of lightning using a lightning locator.

3. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 1 is characterized in that: In S2, the required spatial range is intercepted from the longitude and latitude range of the original satellite infrared brightness temperature observation data. The spatial range is the starting longitude 106.4°E, the ending longitude 110.4°E, the starting latitude 21.4°N, and the ending latitude 25.4°N. The original spatial resolution is reduced to 0.04°*0.04°. Specifically, the original grid data is traversed with a submatrix with a step size of 2, and the maximum value in the submatrix is taken as the new grid data. After traversing the original grid data, the size of the new matrix is 1 / 2 of the original matrix.

4. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 1 is characterized in that: In S2, the brightness temperature value range of the channel observation data of the satellite infrared brightness temperature observation data is adjusted to (0, 70) K to accelerate the convergence speed of the model and improve the training efficiency; And normalized using the following formula; Among them, m1 and m2 represent the maximum and minimum values of the original values, n1 and n2 represent the maximum and minimum values of the normalization, and the original brightness temperature value x is processed to obtain the normalized value y.

5. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 1 is characterized in that: In S2, the longitude, latitude, and time information in the lightning location data are extracted one by one. All time points are grouped and superimposed at 10-minute intervals. The longitude and latitude at the time of the lightning occurrence are then mapped to a grid to obtain the final gridded lightning data. The calculation formula for the longitude and latitude mapping is as follows: Among them, Lat_proj and Lon_proj represent the row and column values of the latitude and longitude values after being projected onto the grid, respectively, res represents the spatial resolution, Lon represents longitude, Lat represents latitude, Lat_end represents the end latitude, and Lon_start represents the start longitude.

6. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 5 is characterized in that: In S2, based on the Band 13 infrared brightness temperature observation channel, 235K is selected as the brightness temperature threshold to identify lightning noise points. The location of the lightning is first obtained from the lightning grid data. The channel brightness temperature difference data BTD1513 and TTD, which represent cloud optical thickness and cloud top phase, are used to filter out cloud-free areas and cirrus objects. Then, it is determined whether there is data with a Band 13 brightness temperature value lower than 235K within an 8km range. If so, the lightning recorded at that point is determined to be real lightning rather than noise. Otherwise, the value at that point is set to 0.

7. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 1 is characterized in that: In S3, the location of lightning in the next hour is predicted based on the satellite infrared brightness temperature observation data and lightning location data from the past hour. The current time is t, and the input data of the past "i" frame and the output result of the future "j" frame are defined as: Among them, [S T ] and [L T ] represent the input satellite infrared brightness temperature observation data and lightning location data for the past hour, F(;θ) represents the model and the learnable parameter θ. Represents the output flash forecast value for the next hour.

8. The lightning approach prediction method based on time-space separation convolutional neural network according to claim 1 is characterized in that S4 include: S4-1,STFNet model establishment; The STFNet model is a spatiotemporal sequence prediction model with an encoding-decoding structure built based on the time fusion block TFBlock. It uses pure convolution operations to separate and extract the spatiotemporal features of meteorological data to more accurately capture the underlying patterns of the data. S4-1-1, construct model space coding structure; The encoder part of the model is composed of N s It is composed of blocks in series to extract high-dimensional spatial feature information of the data. Each block includes Conv2d, GroupNorm and SiLU. The input is X i-1 , F i is the output of the encoder, then the spatial encoding process is expressed as: F i =SiLU(GroupNorm(Conv2d(X i-1 ))); S4-1-2, build model time modeling structure; As the basic unit for capturing lightning temporal features, TFBlock consists of two core modules: a spatiotemporal processing module and a channel processing module. These modules respectively extract the shape and structural features of lightning in a single frame and learn how lightning changes between frames. The two modules process the temporal evolution of lightning in parallel, combining both static and motion features. The outputs of the two modules are then element-wise multiplied with the input data, followed by a 1x1 convolution to fuse the features. Finally, a residual connection is performed with the input data that has not been processed by TFBlock to improve feature extraction. The spatiotemporal processing module is responsible for capturing the static features of lightning to understand the interdependencies between distant pixels. After reducing the spatial dimension using average pooling, the module uses a 1x1 convolution to obtain local region features. Two depthwise convolutions are then used to approximate standard large-kernel convolutions to reduce the number of parameters. The two convolution kernels, 1x11 and 11x1, respectively, learn a wider range of pixel dependencies over a large receptive field, thereby capturing a more accurate lightning structure. Finally, a 1x1 convolution and a sigmoid activation function are used to generate attention weights to further improve the output of the spatiotemporal processing module. The channel processing module captures the motion characteristics of lightning by learning channel weights between different frames in the time and channel dimensions. The channel processing module performs average pooling and maximum pooling operations respectively. Maximum pooling is used to focus on the salient features of the image. Average pooling compresses the spatial dimensions of the input feature map to preserve the overall image characteristics. Both pooling operations are used simultaneously to enhance the characteristic expression of the lightning motion process. The two features after the pooling operation are respectively sent to a shared fully connected layer. Finally, the output channel features are added based on the element-wise operation to obtain the final channel attention weight. The spatiotemporal processing module is represented as: A j =sigmoid(conv2d(DWconv(DWconv(avg_pool(F j-1 ))))); The channel processing module is represented as: Among them, F j-1 is the output of the spatial encoding part, Indicates a connection operation; S4-1-3, construct the model's spatiotemporal decoding structure; The model's spatiotemporal decoder fuses the encoder's spatial feature map with the temporal features of TFBlock to predict future lightning development and, in the process, reconstruct the resolution of lightning features. The number of decoding blocks in the decoder is the same as that in the encoder. Each block includes unConv2d, GroupNorm, and SiLU, expressed as: From k =SiLU(GroupNorm(unConv2d(F k-1 ,T k-1 ))); Among them, F k-1 and T k-1 The outputs of the encoder and TFBlock are respectively, using ConvTransposed2d as the unConv2d operator; S4-1-4, build a complete model; The input data first passes through the model's spatial encoding structure to extract high-dimensional spatial features. The feature map output from the encoding structure is then input into the model's temporal modeling structure TFBlock in S4-1-2 to learn the temporal evolution of lightning. The encoder's feature map and the features of TBBlock are then input into the spatiotemporal decoder for decoding. Finally, a convolution operation is performed to obtain an output of the same size as the input, and a binary classification prediction is performed on the output feature map grid points. S4-2, load the dataset; The processed satellite infrared brightness temperature observation data and lightning location data in S2 are saved in numpy file format in time order. When the model loads the dataset, it is divided into training set, validation set, and test set in a ratio of 7:1:

2. The patterns and weights in the data are learned through the training set. The validation set is used to verify or adjust the model's hyperparameters during training to prevent overfitting and select the model with the best training effect as the final model. The test set is used to test the final performance of the model and evaluate its generalization ability. The satellite data and lightning data one hour before a certain time point are read from the dataset by traversing the time points and input into the model. S4-3, training model. S4-3-1, parameter setting; A fixed random seed is set during model training to ensure the repeatability of the model training results. During the training process, the AdamW optimizer is used, and the learning rate is set to 10 -4 The rest of the settings are default values; the batch size is set to 16, the iteration cycle is set to 1000, and the early stopping strategy is adopted. When the validation set loss stops decreasing for more than 10 consecutive iteration cycles and the model score on the validation set no longer improves, the training is stopped; when traversing the dataset, the batch size is loaded in batches of 16 data each time, all the data of the previous hour are spliced into a tensor by frame and then fed into the model to obtain the predicted data one hour later; S4-3-2, loss function; The Focal loss function is used as the loss function for model training. When the data categories are unbalanced, the Focal loss function assigns different weights to samples according to the difficulty of successful sample classification. The Focal loss function is expressed as: FL(P,y)=-α t (1-P t ) γ log(P t ); Where P is the probability that lightning is correctly classified, (1-P t ) γ is the loss adjustment factor, α is the weighting coefficient; The predicted data and the real data are input into the loss function to calculate the average loss of the predicted results, and then the model parameters are updated through back propagation; S4-4, Model Validation and Comparison; After the model has been trained with one round of training data, the data in the validation set are traversed and input into the model in the same way to obtain lightning prediction data. The number of hits (NA), null reports (NB), missed reports (NC), and non-reports (ND) in the lightning grid data are counted. pre and I act They represent the model prediction output and the real lightning data respectively. The statistical formula of each data is expressed as: NA=∑(I pre =1∧I act =1); NB=∑(I pre =1∧I act =0); NC=∑(I pre =0∧I act =1); ND=∑(I pre =0∧I act =0); After the statistics are completed, the verification indicators of this round of verification set will be calculated: threat score TS, fair threat score ETS, false alarm rate FAR and hit rate POD; the calculation formulas for each evaluation indicator are as follows: in, Finally, the model with the highest validation set score was selected and compared on the test set data with four classic models in spatiotemporal sequence prediction: the convolutional long short-term memory network ConvLSTM, the spatiotemporal sequence prediction model PredRNN based on the ST-LSTM unit, the 3D image segmentation model 3DUnet improved based on unet, and the latest spatiotemporal sequence prediction model Simvp based entirely on CNN. The TS, ETS, FAR, and POD scores of each model were calculated to compare the performance of the STFNet model in lightning nowcasting.

9. A lightning approach prediction system based on a time-space separation convolutional neural network is characterized by: For implementing the lightning approach prediction method based on a time-space separation convolutional neural network according to any one of claims 1 to 8, the lightning approach prediction system based on a time-space separation convolutional neural network comprises: Data acquisition module, used to obtain satellite infrared brightness temperature observation data and lightning location data; A data processing module is used to adjust the spatial resolution and standardize the satellite infrared brightness temperature observation data; and to grid and denoise the lightning location data; The task module is used to define the prediction task: select the data of the past N hours to predict the location of lightning in the next N hours; The prediction module is used to build a spatiotemporal series prediction model with an encoding-decoding structure based on the time fusion block TFBlock. Based on the processed satellite infrared brightness temperature observation data and lightning location data, it is saved in numpy file format in time point order, divided into training set, validation set, and test set in a ratio of 7:1:2, and loaded into the spatiotemporal series prediction model for model training to obtain the optimized model. Based on the prediction task, the optimization model is used to perform lightning forecasting and obtain lightning prediction data.

Images