A lightning warning method based on atmospheric electric field networking data

Abstract

The present invention discloses a lightning warning method based on atmospheric electric field networking data, which relates to the field of lightning warning technology, including dividing electric field meter data into time windows of a certain length, integrating and splicing effective data, and dividing the data into cross-overlapping areas and non-cross-overlapping areas, determining whether lightning activity has occurred in each area, and obtaining a data set; constructing a graph convolutional network model with an adjacency matrix, obtaining whether different electric fields cross according to the adjacency matrix, and directly performing independent training if they do not cross; if they cross, the similarity of the electric field meter data is m≥0.6, which is a cross-overlapping area, and the data of several electric field meters involved in the cross are fused and trained; if m<0.6, it is a non-cross-overlapping area, and the data of different electric field meters are trained separately to obtain a lightning warning model. The present invention improves the judgment and fusion capabilities of signal confidence in the overlapping area, and improves the accuracy and reliability of lightning warning.

Description

Technical Field

[0001] The present invention relates to the technical field of lightning early warning, and in particular to a lightning early warning method based on atmospheric electric field networking data. Background Art

[0002] At present, lightning warning technology mainly relies on traditional means such as radar, satellites and meteorological stations, combined with meteorological observation data for analysis and prediction. However, these methods still have certain limitations in the monitoring and early warning of lightning. Although radar can monitor precipitation and cloud movement in real time, its direct monitoring ability for lightning is weak and there are certain detection blind spots. Satellite remote sensing data is widely used around the world, but it can usually only provide a rough forecast of lightning activity and the response speed is relatively slow. The monitoring range of meteorological stations is limited and cannot provide a large range of real-time electric field data. In addition, the meteorological conditions in different regions vary greatly, resulting in poor timeliness and accuracy of early warning.

[0003] With the development of meteorological observation technology, especially the maturity of wireless sensor network technology, the collection of atmospheric electric field networking data has gradually become an emerging technical means for lightning warning. As one of the precursors of lightning, the changes in the atmospheric electric field can reflect the charge accumulation and discharge risk of thunderstorm clouds. By deploying atmospheric electric field sensors into a network, electric field data in a large area can be obtained in real time, thereby providing a more accurate basis for the monitoring and early warning of lightning activities. In particular, when the atmospheric electric field changes suddenly or exceeds the preset threshold, it may indicate the imminent occurrence of lightning activity, thereby achieving early warning.

[0004] However, the collection and use of atmospheric electric field data still faces some challenges. First, the electric field data itself has strong dynamic changes. How to accurately capture the electric field changes and associate them with the actual situation of lightning is the key to improving the accuracy of early warning. In addition, in practical applications, the detection range of the electric field meters at each station usually has overlapping areas, such as Figure 1As shown in the figure, assuming there are three electric field meters, the three circles represent the detection radius of the three electric field meters, and the red center (*) of the three circles represents the location of the three electric field meters. Electric field meters No. 1 and No. 2 overlap, and No. 3 is relatively independent. Some electric field meters are far apart (such as electric field meter No. 3 and electric field meter No. 1, and electric field meter No. 3 and electric field meter No. 2), which exceeds the detection range of the two. It is impossible for the two to detect the same lightning, which means that the data of the electric field meters are independent and unrelated; and if the detection ranges of the two (or three or more) overlap, it means that the data of the electric field meters are not independent and are related to each other (such as electric field meters No. 1 and No. 2). In other words, in the overlapping areas of the detection range of the electric field meters, the prediction of the lightning warning probability comes from two electric field meters (or three electric field meters and more) (such as the lightning warning probability in area B comes from the fusion warning of electric field meters No. 1 and No. 2). In these intersection areas, how to fuse the data of multiple electric field meters to achieve warning in the intersection area is the key. The warning accuracy and reliability of the warning methods in the prior art for the intersection area are low. Summary of the invention

[0005] In order to overcome the above problems existing in the prior art, the present invention proposes a lightning warning method based on atmospheric electric field networking data.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a lightning early warning method based on atmospheric electric field networking data, comprising:

[0007] S1, select the atmospheric electric field instrument data from multiple stations and align them uniformly through temporal and spatial resolution;

[0008] S2, dividing the electric field meter data into time windows of a certain length, and processing the electric field meter data in each window intensively;

[0009] S3, integrating and splicing valid data from different stations, and dividing the spliced ​​data into cross-overlapping areas and non-cross-overlapping areas, determining whether lightning activity has occurred in each area, and obtaining a data set;

[0010] S4, construct a graph convolutional neural network model, and build an adjacency matrix in the model to distinguish whether the electric field meter data is crossed;

[0011] S5, divide the data set obtained in S3 into a training set, a validation set, and a test set, and evenly distribute the data of the cross-overlapping area and the non-cross-overlapping area in the training set, the validation set, and the test set. Input the training set data into the model at one time, and the model obtains whether different electric fields cross according to the adjacency matrix. If not, independent training is directly performed; if cross, similarity analysis is performed on the cross-over data of multiple electric field meters. If the similarity m of the electric field meter data is greater than or equal to 0.6, it is a cross-overlapping area. At this time, the data of several electric field meters involved in the cross-over are fused for training; if m is less than 0.6, it is a non-cross-overlapping area. At this time, the data of different electric field meters are independent and are trained separately to obtain a lightning warning model;

[0012] S6, inputting the atmospheric electric field instrument data into the lightning warning model obtained in S5, and the model outputs the probability of lightning occurrence within the detection range of each measuring station.

[0013] In the above-mentioned lightning warning method based on atmospheric electric field networking data, S2 specifically includes:

[0014] S21, setting the time window for data processing to 360 seconds, and performing segmented analysis on the electric field instrument data;

[0015] S22, using the sliding window method to select data in segments, where the time length of each window is 180 seconds; the data window slides in the time series at a fixed step length, and the step length is set to 180 seconds.

[0016] In the above-mentioned lightning warning method based on atmospheric electric field networking data, S3 specifically includes:

[0017] S31, integrating and splicing the data collected by each electric field meter according to the timestamp, integrating the measurement data of all electric field meters into a unified time series matrix, wherein each row of the matrix corresponds to the data of an atmospheric electric field meter, and each column corresponds to the data value of a different electric field meter;

[0018] S32, in the time series matrix obtained in S31, the number of valid data points is checked for each row. If the number of valid data points in a row is less than 360, it is considered that the data in this row does not meet the processing requirements, and the entire matrix will be discarded and will not participate in subsequent analysis; if all rows in the matrix meet the condition that the number of valid data points is greater than or equal to 360, linear interpolation processing is performed on the missing data to ensure that each row of electric field meter data has one data per second;

[0019] S33, performing a differential operation on each row of data;

[0020] S34, in each time window, comparing with the lightning location data to determine whether a lightning event occurs in each area. If there is at least one lightning location data in the area in the time window, it is determined that a lightning event occurs, and the area is labeled 1; if no lightning event occurs, the area is labeled 0;

[0021] S35, check the validity of the label columns in the matrix. If all label values ​​in the matrix are 0, that is, no relevant events are detected in all time windows or the data does not meet the predetermined conditions, then the data matrix is ​​considered to lack valid information and is discarded and does not participate in subsequent analysis and processing.

[0022] In the above-mentioned lightning warning method based on atmospheric electric field networking data, the S33 is specifically: using forward difference to calculate the difference between adjacent elements, reducing the data points from 360 to 359, and the specific calculation formula is:

[0023] ;

[0024] in, is the forward difference of the nth term, is the nth term of the sequence, is the n+1th item in the sequence.

[0025] In the above-mentioned lightning warning method based on atmospheric electric field networking data, S4 specifically includes:

[0026] S41, constructing an adjacency matrix for representing the connection relationship between nodes in the graph;

[0027] S42, the graph convolutional network model models the topological structure between nodes through the adjacency matrix, and uses graph convolution operations to perform information aggregation and feature learning. Each layer of the model propagates information between nodes through the adjacency matrix, updates the feature representation of the nodes layer by layer, and finally outputs task-specific prediction results.

[0028] In the above-mentioned lightning warning method based on atmospheric electric field networking data, the S41 is specifically as follows: each element A in the adjacency matrix A ij Indicates the connection relationship between station i and station j. If the actual distance between station i and station j is less than twice the effective detection radius, the element at the corresponding position in the adjacency matrix is ​​set to 1, indicating that there is a connection between the two stations; if the distance between the two stations is greater than or equal to twice the effective detection radius, the element at the corresponding position is set to 0, indicating that there is no direct connection between the two stations.

[0029] The adjacency matrix A is a symmetric matrix, that is, A ij =A ji , and the diagonal element A ii=0, and its value in the matrix is ​​determined by calculating the distance between any two measuring stations.

[0030] In the above-mentioned lightning warning method based on atmospheric electric field networking data, when training the model in S5, the model parameters are optimized to minimize the loss function, and during the training process, the weights of the convolutional layer are adjusted using the back propagation algorithm;

[0031] After the training is completed, the model is evaluated, and its generalization ability and prediction accuracy are verified using the validation set and test set to obtain the spatiotemporal lightning warning model.

[0032] The beneficial effect of the present invention is that the present invention adopts a graph convolutional network (GCN) model to construct a network structure of an atmospheric electric field meter. By processing and analyzing the collected electric field data, and based on the feature learning ability of the graph convolutional network, model training is performed, and finally the goal of lightning warning is achieved. As a deep learning model that can effectively process graph structured data, the graph convolutional network (GCN) can effectively solve the problem of cross-region signal fusion within the detection range of the atmospheric electric field meter by mining the relationship between nodes and edge information. The introduction of the graph convolutional network (GCN) can make full use of the spatial dependence and data correlation between each measuring station, thereby improving the judgment and fusion capabilities of signal confidence in overlapping areas, and improving the accuracy and reliability of lightning warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the intersection and non-intersection of the detection range of the measuring station of the present invention;

[0034] Figure 2 It is a flow chart of the technical solution of the present invention;

[0035] Figure 3 1 is a schematic diagram of an atmospheric electric field measurement station according to an embodiment of the present invention, wherein (a) is a schematic diagram of the layout of the atmospheric electric field measurement station, and (b) is an adjacency matrix corresponding to the atmospheric electric field measurement station;

[0036] Figure 4 It is a structural diagram of the model of the present invention. DETAILED DESCRIPTION

[0037] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0038] like Figure 2 As shown, this embodiment provides a lightning warning method based on atmospheric electric field networking data, including:

[0039] Step 1: Select atmospheric electric field instrument data from multiple stations and align them in a unified manner with respect to time and space resolution to ensure the consistency and comparability of the data, which will serve as the basic data for subsequent model training and prediction.

[0040] Step 1 specifically includes:

[0041] Determine the number of stations required and obtain the atmospheric electric field data of these stations from databases or relevant data sources (such as meteorological monitoring stations, remote sensing platforms, etc.). Extract the raw data of each station, including time, coordinates, and electric field strength. Ensure that the selected stations have sufficient coverage and spatiotemporal data density to represent the overall situation of the study area.

[0042] Step 2: divide the electric field data into time windows of a certain length. In this way, the electric field data in each window will be processed centrally.

[0043] In this embodiment, 6 minutes is used as the length of each time window, and overlapping segmentation is adopted: adjacent time windows overlap, and the data of the overlapping parts are reused.

[0044] Step 3: Integrate and splice the valid data from different stations, and divide the spliced ​​data into cross-overlapping areas and non-cross-overlapping areas to determine whether lightning activity has occurred in each area and obtain a data set.

[0045] Step 3 specifically includes:

[0046] Step A specifically includes: integrating and splicing the data collected by each electric field meter according to the timestamp.

[0047] The time series data from different electric field meters are aligned in chronological order to ensure the consistency of timestamps of each data source. Through this process, the measurement data of all electric field meters are integrated into a unified time series matrix, where each row of the matrix corresponds to the data of an atmospheric electric field meter, and each column corresponds to the data value of an electric field meter. This matrix structure provides a unified input format for subsequent data analysis, processing and model training.

[0048] Step B: In the spliced ​​time series matrix, for each row (corresponding to the electric field data at each time point), firstly check the number of valid data points.

[0049] Step B specifically includes: if the number of valid data points in a row is less than 360, that is, the valid data duration of the row is less than 6 minutes, the row data is considered to not meet the processing requirements, and the entire matrix will be discarded and will not participate in subsequent analysis. If all rows in the matrix meet the condition that the number of valid data points is greater than or equal to 360, the missing data is processed by linear interpolation. Through the linear interpolation method, the missing values ​​are calculated and supplemented based on the existing valid data points to ensure the integrity and continuity of each row of data.

[0050] Step C: perform a differential operation on each row of data.

[0051] Step C specifically includes: using forward difference to calculate the difference between adjacent elements, reducing the number of data points from 360 to 359. Forward difference refers to the difference between the current item and the next item in the sequence. Its mathematical expression is:

[0052] ;

[0053] in, is the forward difference of the nth term, is the nth term of the sequence, is the n+1th item in the sequence.

[0054] Step D, in each time window, compare the lightning location data of the large network and the small network to determine whether a lightning event occurs in each area. If there is at least one lightning location data in the area within the time window, it is determined that a lightning event has occurred and the area is labeled 1; if no lightning event has occurred, the area is labeled 0.

[0055] Step E: Check the validity of the label column in the matrix.

[0056] Step E specifically includes: if all label values ​​in the matrix are 0, that is, no relevant events are detected in all time windows or the data do not meet the predetermined conditions, then it is considered that the data matrix lacks valid information. At this time, the data matrix will be discarded.

[0057] Step 4: Build a model based on graph convolutional network (GCN) for learning and analyzing networking data.

[0058] The structure of the network is as follows Figure 4As shown in the figure, it includes two layers of graph convolution layers. The first layer maps the features of the nodes from the original dimension to 16 dimensions, and the second layer maps it to the final number of categories. During the forward propagation process, the input data first passes through the first layer of graph convolution, then the activation function is applied to introduce nonlinearity, and then dropout is used to prevent overfitting. Next, the data is passed to the second layer of graph convolution, and finally the function outputs the logarithmic probability of each node belonging to each category. In the end, the model outputs the probability distribution of each node belonging to each category. The advantage of GCN is that it can effectively utilize the topological structure information of the graph and learn the representation of the nodes in the graph.

[0059] Step A, when building a model based on graph convolutional network (GCN), we first need to build an adjacency matrix, which is used to represent the connection relationship between nodes in the graph, such as Figure 3 (a) Figure 3 (b) shown.

[0060] Step A specifically includes: In order to represent the adjacency relationship between the stations, an adjacency matrix is ​​defined, in which each element A ij Indicates the connection relationship between station i and station j. Specifically, if the actual distance between station i and station j is less than twice the effective detection radius, the element at the corresponding position in the adjacency matrix is ​​set to 1, indicating that there is a connection between the two stations; if the distance between the two stations is greater than or equal to twice the effective detection radius, the element at the corresponding position is set to 0, indicating that there is no direct connection between the two stations.

[0061] The adjacency matrix A is a symmetric matrix, that is, A ij =A ji , and the diagonal element A ii = 0 (usually does not represent a connection between itself). The construction of this adjacency matrix is ​​based on the geographical coordinates of the stations, and the distance between any two stations is calculated to determine its value in the matrix.

[0062] In step B, the graph convolutional network (GCN) model models the topological structure between nodes through the adjacency matrix and uses graph convolution operations for information aggregation and feature learning.

[0063] Step B specifically includes: each layer of the model propagates information between nodes through the adjacency matrix, updates the feature representation of the nodes layer by layer, and finally outputs task-specific prediction results.

[0064] S5, divide the data set obtained in S3 into a training set, a validation set, and a test set, and evenly distribute the data of the overlapping area and the non-overlapping area in the training set, the validation set, and the test set. Input the training set data into the model at one time, and the model obtains whether different electric fields are crossed according to the adjacency matrix. If not, independent training is directly performed; if crossed, similarity analysis is performed on the crossed multiple electric field meter data. If the similarity of the electric field meter data is m≥0.6, it is a crossing overlapping area. At this time, the data of several electric field meters involved in the crossing are fused for training; if m<0.6, it is a non-crossing overlapping area. At this time, the data of different electric field meters are independent and are trained separately. After the training is completed, the model is evaluated, and its generalization ability and prediction accuracy are verified using the validation set and the test set to obtain a spatiotemporal lightning warning model.

[0065] In this embodiment, the graph convolutional neural network model is trained using the training data set, and the model parameters are optimized to minimize the loss function. All electric field meters in the training set are input into the model, and the adjacency matrix is ​​used to obtain whether different electric fields intersect or overlap. Similarity analysis is performed on the data of multiple intersecting electric field meters. If the similarity exceeds 0.6, it means that lightning may occur in the overlapping area (such as Figure 1 Otherwise, if the similarity does not exceed 0.6, it means that lightning may occur in area A or C. The model will only use the data of electric field meter No. 1 to train the warning results in area A, and only use the data of electric field meter No. 2 to train the warning results in area C. In addition, since electric field meter No. 3 is independent, the corresponding training is also independent.

[0066] Step 6: Input the atmospheric electric field meter data into the graph convolutional network model constructed by S4, and the model outputs the probability of lightning occurrence within the detection range of each measuring station.

[0067] The above embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. A lightning warning method based on atmospheric electric field networking data, characterized in that: include: S1, select the atmospheric electric field instrument data from multiple stations and align them uniformly through temporal and spatial resolution; S2, dividing the electric field meter data into time windows of a certain length, and processing the electric field meter data in each window intensively; S3, integrating and splicing valid data from different stations, and dividing the spliced ​​data into cross-overlapping areas and non-cross-overlapping areas, determining whether lightning activity has occurred in each area, and obtaining a data set; S4, constructing a graph convolutional neural network model, and building an adjacency matrix in the model to determine whether the electric field meter data is crossed; S5, divide the data set obtained in S3 into a training set, a validation set, and a test set, and evenly distribute the data of the cross-overlapping area and the non-cross-overlapping area in the training set, the validation set, and the test set. Input the training set data into the model at one time, and the model obtains whether different electric fields cross according to the adjacency matrix. If not, independent training is directly performed; if cross, similarity analysis is performed on the cross-over multiple electric field meter data. If the similarity of the electric field meter data is m≥0.6, it is a cross-overlapping area. At this time, the data of several electric field meters involved in the cross-over are fused for training; If m < 0.6, it is a non-overlapping area. At this time, the data of different electric field meters are independent and are trained separately to obtain a lightning warning model. S6, inputting the atmospheric electric field instrument data into the lightning warning model obtained in S5, and the model outputs the probability of lightning occurrence within the detection range of each station; The S3 specifically includes: S31, integrating and splicing the data collected by each electric field meter according to the timestamp, integrating the measurement data of all electric field meters into a unified time series matrix, wherein each row of the matrix corresponds to the data of an atmospheric electric field meter, and each column corresponds to the data value of a different electric field meter; S32, in the time series matrix obtained in S31, the number of valid data points is checked for each row. If the number of valid data points in a row is less than 360, it is considered that the data in this row does not meet the processing requirements, and the entire matrix will be discarded and will not participate in subsequent analysis; if all rows in the matrix meet the condition that the number of valid data points is greater than or equal to 360, linear interpolation processing is performed on the missing data to ensure that each row of electric field meter data has one data per second; S33, performing a differential operation on each row of data; S34, in each time window, comparing with the lightning location data to determine whether a lightning event occurs in each area. If there is at least one lightning location data in the area in the time window, it is determined that a lightning event occurs, and the area is labeled 1; if no lightning event occurs, the area is labeled 0; S35, check the validity of the label columns in the matrix. If all label values ​​in the matrix are 0, that is, no relevant events are detected in all time windows or the data does not meet the predetermined conditions, then the data matrix is ​​considered to lack valid information and is discarded and does not participate in subsequent analysis and processing.

2. A lightning early warning method based on atmospheric electric field networking data according to claim 1, characterized in that: The S2 specifically includes: S21, setting the time window for data processing to 360 seconds, and performing segmented analysis on the electric field instrument data; S22, using the sliding window method to select data in segments, where the time length of each window is 180 seconds; the data window slides in the time series at a fixed step length, and the step length is set to 180 seconds.

3. The lightning early warning method based on atmospheric electric field networking data according to claim 1 is characterized in that: The S33 is specifically: using forward difference to calculate the difference between adjacent elements, reducing the data points from 360 to 359, and the specific calculation formula is: in, is the forward difference of the nth term, is the nth term of the sequence, is the n+1th item in the sequence.

4. The lightning early warning method based on atmospheric electric field networking data according to claim 1 is characterized in that: The S4 specifically includes: S41, constructing an adjacency matrix for representing the connection relationship between nodes in the graph; S42, the graph convolutional network model models the topological structure between nodes through the adjacency matrix, and uses graph convolution operations to perform information aggregation and feature learning. Each layer of the model propagates information between nodes through the adjacency matrix, updates the feature representation of the nodes layer by layer, and finally outputs task-specific prediction results.

5. A lightning early warning method based on atmospheric electric field networking data according to claim 4, characterized in that: The S41 is specifically as follows: each element A in the adjacency matrix A ij Indicates the connection relationship between station i and station j. If the actual distance between station i and station j is less than twice the effective detection radius, the element at the corresponding position in the adjacency matrix is ​​set to 1, indicating that there is a connection between the two stations; if the distance between the two stations is greater than or equal to twice the effective detection radius, the element at the corresponding position is set to 0, indicating that there is no direct connection between the two stations. The adjacency matrix A is a symmetric matrix, that is, A ij =A ji , and the diagonal element A ii =0, and its value in the matrix is ​​determined by calculating the distance between any two measuring stations.

6. The lightning early warning method based on atmospheric electric field networking data according to claim 1 is characterized in that: When training the model in S5, the model parameters are optimized to minimize the loss function, and during the training process, the weights of the convolutional layer are adjusted using a back propagation algorithm; After the training is completed, the model is evaluated, and its generalization ability and prediction accuracy are verified using the validation set and test set to obtain the spatiotemporal lightning warning model.

Images