A lightning deduction method and system based on spatiotemporal graph convolutional neural network

Abstract

The present invention discloses a lightning deduction method and system based on a spatiotemporal graph convolutional neural network, the method comprising obtaining lightning area information according to lightning location data, constructing a lightning area topology map; preprocessing the lightning location data, mapping the preprocessed data to the corresponding area, calculating the ground-to-ground lightning density in each area, and classifying the ground-to-ground lightning density according to a set threshold; mapping the ground-to-ground lightning density level of each area to the corresponding node of the graph, obtaining the ground-to-ground lightning density data within a set time, dividing it according to the time interval, combining the lightning area topology map, constructing a lightning density time series data set with a graph structure, inputting the data set into a spatiotemporal graph convolutional neural network model for training and parameter tuning, obtaining a final spatiotemporal graph convolutional neural network model, and using the model to complete lightning deduction. The present invention can more effectively learn the spatial and temporal laws of lightning and improve the accuracy of lightning deduction.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a lightning deduction method and system based on a spatiotemporal graph convolutional neural network. Background Art

[0002] Lightning is a natural phenomenon related to thunderstorms. Its essence is the discharge between thunderstorm clouds or between thunderstorm clouds and the ground. Lightning is caused by the rising airflow inside thunderstorm clouds. This rising momentum creates a potential difference between clouds. When this potential difference exceeds a certain threshold, it triggers the discharge phenomenon. As a natural disaster phenomenon, severe lightning weather will have a negative impact on people's daily lives and economic activities. Therefore, it is very important to provide an accurate, reliable and efficient lightning simulation method.

[0003] Since the causes of lightning activity are complex and the amount of data involved in lightning deduction is huge, lightning deduction has always been a challenging task. In recent years, with the improvement of machine learning, big data and the performance of meteorological observation equipment, artificial intelligence algorithms have been widely used in the field of meteorological forecasting. Through artificial intelligence algorithms, researchers have enabled meteorological models to learn patterns from a large amount of historical meteorological data, improving the accuracy and efficiency of lightning deduction. Now there is a technology that can rasterize lightning data and complete lightning deduction through LSTM (Long Short-Term Memory) and its variant models. This technology can currently complete some reliable lightning forecasting work.

[0004] However, the use of rasterization and LSTM methods for lightning deduction still has limitations. First, the rasterization method divides the geospatial data into regular grids, while lightning location data usually presents an irregular distribution. Affected by the natural geographical environment and meteorological conditions, the lightning frequency and activity range in each region vary greatly. After rasterization, this irregular distribution will lose a lot of local information and details in the fixed grid, thus affecting the accuracy of the deduction. Secondly, the data preprocessing and format conversion process of the rasterization method is complicated, and the data set generated after rasterization is often large in scale. When processing large-scale data, the LSTM model has high computational complexity, long training time, and is prone to overfitting and poor performance. Therefore, it is very necessary to propose an efficient lightning deduction method that can fully capture the spatiotemporal relationship. Summary of the invention

[0005] The purpose of the present invention is to provide a lightning deduction method and system based on spatiotemporal graph convolutional neural network, which can effectively solve the limitations of the prior art in using rasterization methods and LSTM models for lightning deduction. By combining the processing ability of graph convolutional networks for spatial features and the ability of gated recurrent units to capture time series dependencies, the spatiotemporal characteristics of lightning activities can be better reflected, and the accuracy and efficiency of deduction can be improved.

[0006] The present invention adopts the following technical solution:

[0007] A lightning deduction method based on spatiotemporal graph convolutional neural network includes the following steps:

[0008] S1. Obtain lightning area information based on lightning location data and construct a lightning area topology map.

[0009] S2. Preprocess the lightning location data, map the preprocessed data to the corresponding area, calculate the ground-to-ground lightning density in each area, and classify the ground-to-ground lightning density according to the set threshold.

[0010] S3. Map the ground-to-ground lightning density level of each area to the corresponding node of the graph, obtain the ground-to-ground lightning density data within the set time, divide it according to the time interval, and combine it with the lightning area topology map in step S1 to construct a lightning density time series dataset with a graph structure, and divide the dataset into a training set, a validation set, and a test set according to the set ratio.

[0011] S4. Input the data set obtained in step S3 into the spatiotemporal graph convolutional neural network model for training and parameter tuning to obtain the final spatiotemporal graph convolutional neural network model, and use the model to complete lightning deduction.

[0012] Furthermore, in step S1, constructing a lightning area topology map includes the following contents:

[0013] The research area is determined based on the lightning area information, and an undirected graph is used to describe the lightning area topology. The specific expression is:

[0014] ;

[0015] Where V represents the set of lightning zone nodes, , It represents the Nth node, where N represents the total number of nodes; E represents the set of edges; and G represents the overall logical structure of the lightning area topology map.

[0016] Use the adjacency matrix A to represent the geographical location relationship between lightning areas. .

[0017] Furthermore, in step S2, the classification of ground-to-ground lightning density includes the following contents:

[0018] S201. Screen the lightning location data in the study area, obtain the lightning location data with historical lightning intensity greater than 2KA, and map it to the corresponding area.

[0019] S202. Calculate the density of ground-to-ground lightning. The specific formula is:

[0020]

[0021] Where S represents the area of ​​the study area, N1 represents the number of ground-to-ground flashes in the study area, and N G Indicates the density of ground-to-ground lightning.

[0022] S203, divide the density of ground lightning into [0,1), [1,2), [2,3), [3,4] and The five levels represent the degree of lightning hazard in the study area from small to large.

[0023] Furthermore, in step S3, constructing a lightning density time series dataset with a graph structure includes the following contents:

[0024] S301, dividing the ground-to-ground lightning density sequence within a set time according to time intervals to obtain a lightning density feature matrix , where P represents the total number of node attribute characteristics, and the node attribute characteristics are the ground-to-ground lightning density levels.

[0025] S302. Combine the adjacency matrix A of the lightning area topology map, associate the lightning density feature matrix with the corresponding spatial nodes, and generate a lightning density time series dataset with a graph structure.

[0026] S303, dividing the time series data set in step S302 into a training set, a validation set and a test set in a ratio of 8:1:1.

[0027] Furthermore, in step S4, completing the lightning deduction includes the following:

[0028] S401, the spatiotemporal graph convolutional neural network model includes a spatial feature extraction module, a temporal feature extraction module and a fully connected layer; wherein the spatial feature extraction module includes a two-layer graph convolutional neural network, and the temporal feature extraction module includes a gated recurrent unit.

[0029] The training set is input into the spatiotemporal graph convolutional neural network model, the error is calculated through the loss function, and the parameters of the model are adjusted according to the error to train the model; the training process is iterated until the model converges; the formula of the loss function loss is:

[0030]

[0031] in, represents the lightning density value predicted by the model at time t, Y t represents the true value of lightning density at time t, L reg represents the regularization term, represents the training parameters, and n represents the total time.

[0032] The validation set is input into the trained spatiotemporal graph convolutional neural network model, the prediction performance indicators are calculated, and the hyperparameters with the best performance are selected.

[0033] The test set is input into the trained spatiotemporal graph convolutional neural network model after determining the optimal hyperparameters, and the prediction performance indicators are calculated to obtain the final spatiotemporal graph convolutional neural network model.

[0034] S402. Input the lightning density time series data set with graph structure into the final spatiotemporal graph convolutional neural network model. When passing through the spatial feature extraction module, the spatial information of the topological map of the lightning landing point area is extracted through the graph convolution operation to generate time series data with spatial characteristics. When the data passes through the time feature extraction module, the time dependency of the lightning landing point data is obtained through information transmission between gated recurrent units, and a high-dimensional representation with spatiotemporal characteristics is output. When the high-dimensional representation passes through the fully connected layer, it is processed by linear transformation, nonlinear activation and output mapping to obtain the final lightning deduction result.

[0035] Furthermore, the spatial feature extraction module is configured to perform the following actions:

[0036] Use the spectral method to perform graph convolution operations and extract spatial information between regions layer by layer. The specific formula is:

[0037]

[0038] in, Represents a time series with spatial characteristics; represents the normalized adjacency matrix, , represents the adjacency matrix after adding the self-loop, express The degree matrix of the input layer; ReLU represents the rectified linear unit; W0 represents the weight matrix from the input layer to the hidden layer, , B represents the length of the feature matrix, H represents the latitude of the hidden layer; W1 represents the weight matrix from the hidden layer to the output layer, , T represents the prediction length; Represents the Sigmoid activation function used for nonlinear models.

[0039] Furthermore, the temporal feature extraction module is configured to perform the following actions:

[0040] The calculation formula for extracting high-dimensional representations with spatiotemporal features using gated recurrent units is:

[0041]

[0042]

[0043]

[0044]

[0045] in, represents the hidden layer state at time t-1; Represents the output value at time t; represents the update gate, Reset gate. represents the input modulation gate, , and Both represent weight matrices used for linear transformation; , , Represents the deviation during training; tanh represents the nonlinear activation function; represents the matrix multiplication between the weight matrix and the latent vector; * represents the Hadamard multiplication.

[0046] Furthermore, the present invention also proposes a lightning deduction system based on a spatiotemporal graph convolutional neural network, which includes:

[0047] The topology map construction module is used to obtain lightning area information based on lightning location data and construct a lightning area topology map.

[0048] The data classification module is used to pre-process the lightning location data, map the pre-processed data to the corresponding area, calculate the ground-to-ground lightning density in each area, and classify the ground-to-ground lightning density according to the set threshold.

[0049] The data set acquisition module is used to map the ground-to-ground lightning density level of each area to the corresponding node of the graph, obtain the ground-to-ground lightning density data within a continuous period of time, divide it according to time intervals, and combine it with the lightning area topology map to construct a lightning density time series data set with a graph structure. The data set is then divided into a training set, a validation set, and a test set according to a set ratio.

[0050] The lightning deduction module is used to input the lightning density time series dataset with a graph structure into the spatiotemporal graph convolutional neural network model for training and parameter tuning to obtain the final spatiotemporal graph convolutional neural network model, and use the model to complete lightning deduction.

[0051] Furthermore, the present invention also proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the lightning deduction method based on the spatiotemporal graph convolutional neural network when executing the computer program.

[0052] Furthermore, the present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is run by a processor, it executes the lightning deduction method based on the spatiotemporal graph convolutional neural network.

[0053] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0054] 1. The present invention can effectively handle complex geographic dependencies and solve the shortcomings of rasterization methods in processing irregular spatial data. At the same time, the present invention processes time series through gated recurrent units to capture the temporal dynamic changes of lightning activities and enhance the ability to predict the timing of lightning occurrence.

[0055] 2. The present invention can process a large amount of historical lightning location data, automatically learn its spatial and temporal laws, and thus achieve efficient lightning trend deduction.

[0056] 3. The present invention can not only process irregularly distributed spatial data, but also effectively avoid overfitting problems, improve the prediction accuracy and computational efficiency of the model, and is suitable for short-term and long-term lightning prediction tasks.

[0057] 4. The present invention significantly improves the accuracy and efficiency of lightning deduction, provides technical support for the prevention of meteorological disasters, and reduces the losses caused by lightning disasters to the social economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is an overall implementation flow chart of the present invention.

[0059] Figure 2 It is a model diagram of the spatiotemporal graph convolutional neural network of the present invention.

[0060] Figure 3 It is a schematic diagram of extracting spatial features by the spatial feature extraction module of the present invention.

[0061] Figure 4 1 is a diagram of lightning deduction results in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0063] To achieve the above objectives, the present invention proposes a lightning deduction method based on spatiotemporal graph convolutional neural network, such as Figure 1 As shown, the specific steps are as follows:

[0064] S1. Collect lightning location data through lightning locators deployed at meteorological observation stations, obtain lightning area information based on the data, and construct a lightning area topology map; the specific contents are as follows:

[0065] Lightning area information includes longitude and latitude, intensity, steepness, error, positioning method, specific time and date, specific lightning-strike area and county, and the geographic grid to which it belongs.

[0066] The research area is determined based on the lightning area information, and an undirected graph is used to describe the lightning area topology. The specific expression is:

[0067] ;

[0068] Where V represents the set of lightning zone nodes, , represents the Nth node, N represents the total number of nodes; E represents the set of edges; G represents the overall logical structure of the lightning area topology map, which is an undirected graph composed of the node set V and the edge set E, and is used to describe the spatial topological relationship between lightning areas.

[0069] Use the adjacency matrix A to represent the geographical location relationship between lightning areas. ; The adjacency matrix A contains only two elements, 0 and 1. 0 indicates that there is no edge connection between the two lightning areas, and 1 indicates that there is an edge connection. The nodes V and edges E in the lightning area topology graph G are a graphical description of spatial relationships, and the adjacency matrix A is a matrix expression of this relationship, which is convenient for mathematical operations.

[0070] In this embodiment, the study area is divided into 100 districts and counties, and the districts and counties are used as nodes of the graph, that is, N=100; according to the geographic spatial relationship, a 100×100 adjacency matrix A is constructed to describe the spatial relationship between the districts and counties.

[0071] S2. Preprocess the lightning location data of the study area in August 2018, map the preprocessed data to the corresponding area, calculate the ground-to-ground lightning density in each area, and classify the ground-to-ground lightning density according to the set threshold; the specific contents are as follows:

[0072] S201. Screen the lightning location data in the study area, obtain the lightning location data with historical lightning intensity greater than 2KA, and map it to the corresponding districts and counties.

[0073] S202. Calculate the density of ground-to-ground lightning. The specific formula is:

[0074]

[0075] Where S represents the area of ​​the study area, in square kilometers; N1 represents the number of ground-to-ground flashes in the study area, in times; N G Indicates the density of ground-to-ground lightning in times per square kilometer per year.

[0076] S203. According to the ground-to-ground lightning density classification standard, the ground-to-ground lightning density is divided into five levels from small to large, representing the lightning danger level in the area from small to large. Specifically, for danger level I, the ground-to-ground lightning density range is [0,1); for danger level II, the ground-to-ground lightning density range is [1,2); for danger level III, the ground-to-ground lightning density range is [2,3); for danger level IV, the ground-to-ground lightning density range is [3,4); for danger level V, the ground-to-ground lightning density range is [4,5).

[0077] S3, map the ground-to-ground lightning density level of each region to the corresponding node of the graph, obtain the ground-to-ground lightning density data within a continuous month, divide it according to the time interval, combine it with the lightning area topology map in step S1, construct a lightning density time series dataset with a graph structure, and divide the dataset into a training set, a validation set and a test set according to the set ratio; the specific content is:

[0078] S301. Divide the lightning density sequence within a continuous period of time into 15-min time intervals to obtain the lightning density feature matrix. , where P represents the number of node attribute features. In the present invention, the attribute feature of the node is the density level of ground lightning in the area.

[0079] S302. Combine the adjacency matrix A of the lightning area topology map, associate the lightning density feature matrix with the corresponding spatial nodes, and generate a lightning density time series dataset with a graph structure.

[0080] S303, dividing the time series data set in step S302 into a training set, a validation set and a test set in a ratio of 8:1:1.

[0081] In this embodiment, P=1.

[0082] S4, input the data set obtained in step S3 into the spatiotemporal graph convolutional neural network model for training and parameter tuning, obtain the final spatiotemporal graph convolutional neural network model, and use the model to complete lightning deduction; Figure 2 As shown, the specific contents are:

[0083] S401, the spatiotemporal graph convolutional neural network model includes a spatial feature extraction module, a temporal feature extraction module and a fully connected layer; wherein the spatial feature extraction module includes a two-layer graph convolutional neural network, and the temporal feature extraction module includes a gated recurrent unit.

[0084] Xt-n+1 represents the lightning density feature vector at the t-n+1th moment, X t-1 represents the lightning density feature vector at time t-1, X t Represents the lightning density feature vector at time t.

[0085] The training set is input into the spatiotemporal graph convolutional neural network model, the error is calculated through the loss function, and the parameters of the model are adjusted according to the error to minimize the error between the predicted value and the true value, and the model is trained; the training process is iterated until the model converges; the formula of the loss function loss is:

[0086]

[0087] in, represents the lightning density value predicted by the model at time t, Y t represents the true value of lightning density at time t, L reg represents the regularization term, represents the training parameters, and n represents the total time.

[0088] The purpose of the loss function is to penalize the model parameters while calculating the error between the predicted value and the true value to prevent the model from overfitting and achieve the effect of training the model.

[0089] The validation set is used to tune the model to select the optimal hyperparameters. The validation set is input into the trained spatiotemporal graph convolutional neural network model to calculate the prediction performance indicators. By comparing the performance of different hyperparameter combinations on the validation set, the hyperparameters with the best performance are selected.

[0090] The test set is used to evaluate the model. The test set is input into the trained spatiotemporal graph convolutional neural network model after determining the optimal hyperparameters, and the prediction performance index is calculated. This result is used to judge the generalization ability of the model, that is, the prediction effect of the model in actual application scenarios; the final spatiotemporal graph convolutional neural network model is obtained.

[0091] The prediction performance indicators include root mean square error RMSE, mean absolute error MAE and determination coefficient R 2 , the specific calculation formula is:

[0092]

[0093]

[0094]

[0095] in, express With Y t The average value of the error between .

[0096] Among the three indicators, RMSE and MAE represent Y t and The smaller the value, the better the prediction ability of the model. 2 It is used to measure the ability of the predicted value to represent the true value. The larger the value, the higher the prediction ability of the model.

[0097] In this embodiment, the optimal hyperparameters are set as follows: learning rate is 0.0001, batch size is 64, hidden layer dimension is 64, number of training rounds is 3000, weight decay coefficient in L2 regularization is 0.0015, and ratio of Dropout regularization is 0.3; in terms of model time window setting, input sequence length is 8, and set prediction sequence lengths include 1, 2, 3, and 4, corresponding to 15 minutes, 30 minutes, 45 minutes, and 60 minutes in the future. During training, an early stopping strategy is adopted, and the patience value of the early stopping strategy is set to 500 rounds, that is, if the model performance does not improve within 500 rounds, the training is stopped.

[0098] S402: Input the lightning density time series data set with graph structure into the final spatiotemporal graph convolutional neural network model. When passing through the spatial feature extraction module, the spectral method is used to perform graph convolution operation to extract the spatial information of the lightning strike area topology map layer by layer to generate time series data with spatial features. The specific formula is:

[0099]

[0100] in, Represents a time series with spatial characteristics; represents the normalized adjacency matrix, , represents the adjacency matrix after adding the self-loop, express The degree matrix; ReLU represents the rectified linear unit, which is used as the activation layer; W0 represents the weight matrix from the input to the hidden layer, , B represents the length of the feature matrix, H represents the latitude of the hidden layer; W1 represents the weight matrix from the hidden layer to the output layer, , T represents the prediction length; Represents the Sigmoid activation function used for nonlinear models.

[0101] In a gated recurrent unit, the update gate Responsible for executing the functions of "forgetting" and "selective memory" at the same time, adjusting the impact of the state information of the previous moment on the current moment; at the same time, resetting the gate It determines how to fuse the state information of the previous moment with the current information; obtains the time dependency of lightning strike data through information transmission between gated recurrent units, and outputs a high-dimensional representation with spatiotemporal characteristics. The specific formula is:

[0102]

[0103]

[0104]

[0105]

[0106] in, represents the hidden layer state at time t-1; Represents the output value at time t; represents the update gate, Reset gate. represents the input modulation gate, , and Both represent weight matrices used for linear transformation; , , Both represent deviations during training; tanh represents nonlinear activation functions, which give neural networks the ability to handle nonlinear problems; represents the matrix multiplication between the weight matrix and the latent vector; * represents the Hadamard multiplication.

[0107] The above formula reflects that while capturing the lightning information at the current moment, it still retains the changing trend of historical lightning information and has the ability to capture time dependency.

[0108] When the high-dimensional representation with spatiotemporal features passes through the fully connected layer, it is processed by linear transformation, nonlinear activation and output mapping to integrate and transform the extracted spatiotemporal features to obtain the final lightning deduction result.

[0109] Figure 3 The process of extracting spatial information by the spatial feature extraction module is shown. Figure 3 In (a), it is assumed that the lightning area is centered on node 1. Figure 3 (b) shows the topological relationship between the central lightning area and its surrounding lightning areas. By encoding the topological structure of the lightning area graph and the attributes within the area, the geographical location dependency between the areas is obtained.

[0110] Among the set prediction sequence lengths, 15 minutes achieves the best prediction effect, and 30 minutes, 45 minutes, and 60 minutes predictions also have good results. Figure 4The lightning density prediction results of a county in Hunan Province when the prediction sequence length is 15 minutes are displayed. By comparing the changes between the predicted values ​​and the true values, it is shown that the present invention can achieve better results in lightning simulation tasks.

[0111] The embodiment of the present invention also proposes a lightning deduction system based on a spatiotemporal graph convolutional neural network, including a topology map construction module, a data partitioning module, a data set acquisition module, a lightning deduction module, and a computer program that can be run on a processor. It should be noted that each module in the above system corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0112] The embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, the specific steps of the method provided in the embodiment of the present invention are corresponding to the specific steps of the method provided in the embodiment of the present invention, and the processor has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0113] The embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. It should be noted that when the computer program is executed by the processor, it corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A lightning deduction method based on spatiotemporal graph convolutional neural network, characterized in that: include: S1. Obtain lightning area information based on lightning location data and construct a lightning area topology map; S2. Preprocess the lightning location data, map the preprocessed data to the corresponding area, calculate the ground-to-ground lightning density in each area, and classify the ground-to-ground lightning density according to the set threshold; S3, map the ground-to-ground lightning density level of each area to the corresponding node of the graph, obtain the ground-to-ground lightning density data within a set time, divide it according to the time interval, and build a lightning density time series dataset with a graph structure in combination with the lightning area topology map in step S1, and divide the dataset into a training set, a validation set, and a test set according to a set ratio; S4, inputting the data set obtained in step S3 into the spatiotemporal graph convolutional neural network model for training and parameter tuning to obtain the final spatiotemporal graph convolutional neural network model, wherein the spatiotemporal graph convolutional neural network model includes a spatial feature extraction module, a temporal feature extraction module and a fully connected layer; The lightning density time series dataset with a graph structure is input into the final spatiotemporal graph convolutional neural network model. When passing through the spatial feature extraction module, the spatial information of the topological map of the lightning landing point area is extracted through the graph convolution operation to generate time series data with spatial characteristics. When the data passes through the temporal feature extraction module, the time dependency of the lightning landing point data is obtained through information transmission between gated recurrent units, and a high-dimensional representation with spatiotemporal characteristics is output. When the high-dimensional representation passes through the fully connected layer, it is processed by linear transformation, nonlinear activation and output mapping to obtain the final lightning deduction result.

2. The lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: In step S1, constructing a lightning area topology map includes the following: The research area is determined based on the lightning area information, and an undirected graph is used to describe the lightning area topology. The specific expression is: G = (V, E); Where V represents the set of nodes in the lightning region, V = {v1, v2, …, v N },v N represents the Nth node, N represents the total number of nodes; E represents the set of edges; G represents the overall logical structure of the lightning area topology map; Use the adjacency matrix A to represent the geographical relationship between lightning areas, A∈R N×N .

3. The lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: In step S2, the density of ground-to-ground lightning is classified into the following levels: S201, screening the lightning location data in the study area, obtaining lightning location data with historical lightning intensity greater than 2KA, and mapping it to the corresponding area; S202. Calculate the density of ground-to-ground lightning. The specific formula is: Where S represents the area of ​​the study area, N1 represents the number of ground-to-ground flashes in the study area, and N G Indicates the density of ground-to-ground flashes; S203. The density of ground-to-ground flashes is divided into five levels from small to large: [0,1), [1,2), [2,3), [3,4) and [4,∞), representing the grading of lightning hazard levels in the study area from small to large.

4. The lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 2 is characterized in that: In step S3, constructing a lightning density time series dataset with a graph structure includes the following: S301, dividing the ground-to-ground lightning density sequence within the set time according to the time interval to obtain the lightning density feature matrix X N ×P , where N represents the total number of nodes, P represents the total number of node attribute characteristics, and the node attribute characteristics are the ground-to-ground lightning density level; S302, combining the adjacency matrix A of the lightning area topology map, associating the lightning density feature matrix with the corresponding spatial nodes, and generating a lightning density time series data set with a graph structure; S303, dividing the time series data set in step S302 into a training set, a validation set and a test set in a ratio of 8:1:

1.

5. The lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: In step S4, the final spatiotemporal graph convolutional neural network model includes the following contents: The spatial feature extraction module includes a two-layer graph convolutional neural network, and the temporal feature extraction module includes a gated recurrent unit; The training set is input into the spatiotemporal graph convolutional neural network model, the error is calculated through the loss function, and the parameters of the model are adjusted according to the error to train the model; the training process is iterated until the model converges; the formula of the loss function loss is: in, represents the lightning density value predicted by the model at time t, Y t represents the true value of lightning density at time t, L reg represents the regularization term, λ represents the training parameter, and n represents the total time; Input the validation set into the trained spatiotemporal graph convolutional neural network model, calculate the prediction performance indicators, and select the hyperparameters with the best performance; The test set is input into the trained spatiotemporal graph convolutional neural network model after determining the optimal hyperparameters, and the prediction performance indicators are calculated to obtain the final spatiotemporal graph convolutional neural network model.

6. The lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The spatial feature extraction module is configured to perform the following actions: Use the spectral method to perform graph convolution operations and extract spatial information between regions layer by layer. The specific formula is: Where f(X,A) represents a time series with spatial characteristics; X represents the lightning density feature matrix; A represents the adjacency matrix of the lightning area; represents the normalized adjacency matrix, represents the adjacency matrix after adding the self-loop, express degree matrix; ReLU represents the rectified linear unit; W0 represents the weight matrix from the input layer to the hidden layer, W0∈R B×H , B represents the length of the feature matrix, H represents the latitude of the hidden layer; W1 represents the weight matrix from the hidden layer to the output layer, W1∈R H×T , T represents the prediction length; σ() represents the Sigmoid activation function used for nonlinear models.

7. The lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: The temporal feature extraction module is configured to perform the following actions: The calculation formula for extracting high-dimensional representations with spatiotemporal features using gated recurrent units is: z t =σ(W z ·[f(X,A),h t-1 ]+b z ) r t =σ(W r ·[f(X,A),h t-1 ]+b t ) Where σ represents the Sigmoid activation function used for nonlinear models; f(X,A) represents the time series with spatial characteristics, X represents the lightning density feature matrix, A represents the adjacency matrix of the lightning area; h t-1 represents the hidden layer state at time t-1; h t represents the output value at time t; z t represents the update gate, r t Reset gate. represents the input modulation gate, W Z , W r and W represent weight matrices for linear transformation; b z 、b t , b both represent bias; tanh represents a nonlinear activation function; · represents the matrix multiplication between the weight matrix and the latent vector; * represents the Hadamard multiplication.

8. A system for the lightning deduction method based on spatiotemporal graph convolutional neural network according to claim 1, characterized in that: include: A topology map construction module is used to obtain lightning area information based on lightning location data and construct a lightning area topology map; The data classification module is used to pre-process the lightning location data, map the pre-processed data to the corresponding area, calculate the density of ground-to-ground lightning in each area, and classify the density of ground-to-ground lightning according to the set threshold; The data set acquisition module is used to map the ground-to-ground lightning density level of each area to the corresponding node of the graph, obtain the ground-to-ground lightning density data within a continuous period of time, divide it according to the time interval, and build a lightning density time series data set with a graph structure in combination with the lightning area topology map. The data set is then divided into a training set, a validation set, and a test set according to a set ratio. The lightning deduction module is used to input the lightning density time series dataset with a graph structure into the spatiotemporal graph convolutional neural network model for training and parameter tuning to obtain the final spatiotemporal graph convolutional neural network model, wherein the spatiotemporal graph convolutional neural network model includes a spatial feature extraction module, a temporal feature extraction module and a fully connected layer; the lightning density time series dataset with a graph structure is input into the final spatiotemporal graph convolutional neural network model, and when passing through the spatial feature extraction module, the spatial information of the topological map of the lightning landing point area is extracted through the graph convolution operation to generate time series data with spatial characteristics, and when the data passes through the temporal feature extraction module, the temporal dependency of the lightning landing point data is obtained through information transmission between gated recurrent units, and a high-dimensional representation with spatiotemporal characteristics is output, and when the high-dimensional representation passes through the fully connected layer, it is processed by linear transformation, nonlinear activation and output mapping to obtain the final lightning deduction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the lightning deduction method based on the spatiotemporal graph convolutional neural network described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the lightning deduction method based on a spatiotemporal graph convolutional neural network according to any one of claims 1 to 7 is executed.

Images