A method for evaluating sunspot region correlation based on cross-validation

Abstract

The application discloses a sunspot region correlation evaluation method based on cross-validation, and comprises the following steps: 1, obtaining a flare dataset, and performing dataset segmentation according to sunspot region numbers to obtain each data subset; 2, performing machine learning algorithm training on each data subset to obtain a machine learning model of the data subset; 3, calculating a corresponding AUC index for each machine learning model to obtain an A matrix; 4, performing correlation analysis by using the A matrix to obtain a correlation index Corr; and 5, judging the correlation index Corr to obtain a given region correlation determination conclusion. The application provides a calculation method of the correlation index, which can simplify the process of establishing a prediction algorithm to a certain extent, and in the process of using federal transfer learning, can aggregate algorithms of sunspot regions with high similarity to speed up the training process and then improve the prediction efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of flare prediction, and relates to a method for solving redundancy of sunspot region numbering, in particular to a sunspot region correlation evaluation method based on cross-validation for a flare prediction process. BACKGROUND

[0002] In the technical field of flare prediction, in order to facilitate observation and research of sunspots, scientists have adopted a numbering system to name each sunspot region, and the basic method is to identify and classify sunspots on the surface of the sun according to the size, shape and position of the sunspots, and then determine the number after an active region is observed by two independent observation stations at the same time. Each number is composed of a letter and four digits, such as "AR 5128" or "NOAA AR 5128". This numbering system has been implemented since January 5, 1972, and has been used until now. As of June 14, 2002, more than 10,000 active regions have been numbered. However, the same sunspot may be observed and recorded by different observers or instruments multiple times, resulting in the same sunspot being assigned different numbers. Therefore, this sunspot region numbering method has redundancy.

[0003] In order to solve this problem, the International Astronomical Union revised the sunspot numbering system in 2011 and introduced standardized measurement methods such as "relative sunspot number" and "international sunspot number" to better compare and analyze data between different observers and instruments. However, in fact, this revision method still cannot perfectly represent all sunspot regions. On the one hand, this redundant numbering can seriously affect the prediction task carried out by using machine learning methods; on the other hand, due to the observation and recording of sunspots multiple times, there are a large number of repeated flare samples in the data set, and this redundancy increases the complexity of the flare samples, so as to affect the analysis cost of the machine learning algorithm. The existing technology does not carry out detailed discussion and research on the redundancy of sunspot region numbering.

[0004] In summary, it is necessary to design a method suitable for solving the redundancy of sunspot region numbering to improve the analysis accuracy and prediction efficiency. SUMMARY

[0005] In order to effectively evaluate the correlation between sunspot region numbers and then solve the problem of numbering redundancy, the application provides a sunspot region correlation evaluation method based on cross-validation.

[0006] The purpose of the application is achieved by the following technical scheme:

[0007] A sunspot region correlation evaluation method based on cross-validation, comprising the following steps:

[0008] Step one: Obtain the flare dataset, and perform dataset segmentation according to the sunspot region number to obtain each data subset;

[0009] Step two: Perform machine learning algorithm training on each data subset obtained in step one to obtain the machine learning model of the data subset;

[0010] Step three: Calculate the corresponding AUC index for each machine learning model in step two to obtain the A matrix;

[0011] Step four: Perform correlation analysis using the A matrix calculated in step three to obtain the correlation index Corr;

[0012] Step five: Determine the correlation index Corr to obtain the correlation determination conclusion of the given region.

[0013] Compared with the prior art, the present application has the following advantages:

[0014] The present application establishes the relationship between the prediction algorithm and the sunspot region number, and provides a calculation method of the "correlation index" based on the cross-validation evaluation method. This calculation method can simplify the process of establishing the prediction algorithm to a certain extent, and in the process of using federal transfer learning, the sunspot regions with high similarity can be aggregated for algorithm to speed up the training process and improve the prediction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the sunspot region correlation evaluation method based on cross-validation of the present application;

[0016] Figure 2 The A matrix calculation results of seven sunspot regions taking the CNN-GRU algorithm as an example. DETAILED DESCRIPTION

[0017] The technical solutions of the present application will be further described below in conjunction with the drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present application without departing from the spirit and scope of the present application shall be covered in the protection scope of the present application.

[0018] The present application provides a sunspot region correlation evaluation method based on cross-validation, as shown in Figure 1 The method comprises the following steps:

[0019] Step one: Obtain the flare dataset, and perform dataset segmentation according to the sunspot region number to obtain each data subset.

[0020] Step one: Obtain the flare dataset, which can be expressed as:

[0021] {(N0,X0,y0),(N0,X1,y1),(N1,X2,y2),…,(N j ,X i ,y i )}

[0022] Wherein, X i represents the forecast factor of the i-th flare data, y i represents the label of whether the i-th flare data occurs in the future 48 hours, N j represents the sunspot region number j in the flare data.

[0023] Step two: data set segmentation is performed to obtain each data subset, and let D j represent the data set corresponding to the sunspot region number j, then the process can be represented as:

[0024] D j ={(N j ,X0,y0),(N j ,X1,y1),…,(N j ,X i ,y i )}

[0025] Step three: through the above process of step two, it is ensured that there is and only one flare data with the same number in a certain data set, so the entire data set is segmented into the form of {D0,D1,D2,…,D j}.

[0026] Step two: machine learning algorithm training is performed on each data subset obtained in step one to obtain the machine learning model of the data subset.

[0027] For each data subset D j , machine learning algorithm training is performed to obtain the machine learning model f j of the data subset, and the machine learning algorithm training process can be represented as:

[0028]

[0029] Wherein, is used to represent the machine learning algorithm.

[0030] In this step, the machine learning algorithm includes but is not limited to "neural network algorithm", "support vector machine algorithm" and the like. The present application does not limit the specific machine learning algorithm, and any model f j can be obtained.

[0031] Step three: calculate the corresponding AUC index of each machine learning model in step two to obtain the A matrix.

[0032] Step three one: for each machine learning model f j , compute its value of AUC metric in cross-validation A j,k , the specific process is:

[0033]

[0034] where, is the mapping process for calculating AUC metric using algorithm and dataset, D k represents all data subsets left after removing D j from the "flare dataset" in step one.

[0035] In this step, the AUC (Area Under the Curve) metric is a common measure to evaluate the performance of binary classification algorithms, with AUC ranging from 0 to 1, and the AUC of a perfect classifier is equal to 1, while the AUC of random guessing is equal to 0.5.

[0036] Step three two: traverse all data subsets by the method of step three one, get A matrix:

[0037]

[0038] Step four: use the A matrix calculated in step three to perform correlation analysis, get correlation index Corr.

[0039] For given flare region j and k, define Corr j,k represents the correlation of model f j and model f k , the calculation method is as follows:

[0040]

[0041] where, A (j,k) represents the AUC metric value calculated using model f j and data subset D k by the method of step three one, that is, the element value of the jth row and kth column in the A matrix; A (k,j) represents the AUC metric value calculated using model f k and data subset D j by the method of step three one, that is, the element value of the kth row and jth column in the A matrix; A (j,j) represents the AUC metric value calculated using model f j and data subset D j by the method of step three one, that is, the element value of the jth row and jth column in the A matrix; A (k,k) represents the AUC metric value calculated using model f kand data subset D k The calculated AUC index value, i.e. the element value of the kth row and kth column in the A matrix.

[0042] The advantage of designing the calculation formula of Corr is that on the one hand, the degree of completion of the model itself can be determined, and on the other hand, the difference between the model f j and the model f k can be expressed in numerical values.

[0043] Step five: judge the correlation index Corr, and draw the correlation determination conclusion of the given area.

[0044] This step screens the models with strong correlation by setting the threshold value of the correlation index Corr, which is defined as follows:

[0045] If the correlation index Corr j,k of the sunspot area j and the sunspot area k is greater than or equal to 1, it is proved that j and k have "strong correlation", and it is confirmed that they are the same sunspot area with repeated numbering;

[0046] If the correlation index Corr j,k of the sunspot area j and the sunspot area k is between 0.5 and 1, it is proved that j and k have "certain correlation", and it is suspected that they are the same sunspot area with repeated numbering;

[0047] If the correlation index Corr j,k of the sunspot area j and the sunspot area k is between 0 and 0.5, it is proved that j and k have "weak correlation", and they are not treated as the same sunspot area;

[0048] If the correlation index Corr j,k of the sunspot area j and the sunspot area k is less than 0, it cannot be explained that j and k have correlation, and it is proved that the machine learning algorithm is abnormal and needs to be re-executed in step two.

[0049] Embodiment:

[0050] This embodiment takes the seven sunspot areas numbered 2362, 3364, 1321, 4197, 2220, 1461 and 5956 as examples, and uses the CNN-GRU algorithm to establish a machine learning model for each data subset under each number.

[0051] The "CNN-GRU" is the full name of "Convolutional Neural Network-Gated Recurrent Unit", which is a typical machine learning algorithm.

[0052] The data acquisition and segmentation training process is omitted. The A matrix obtained by the method of step three is shown in Figure 2 .

[0053] Using the results of the above calculation of A matrix, on this basis, using step four to calculate the correlation between models Corr, the relationship can be obtained as shown in Table 1.

[0054] Table 1 Correlation analysis between sunspot regions

[0055]

[0056] According to the correlation judgment standard of step five, it can be known that after excluding the "1" on the diagonal line (the same number is the same sunspot region, and the correlation is necessarily 1), the correlation between 5956 and 1461 is 0.54, which is determined to have "certain correlation", and is suspected to be the same sunspot region being repeatedly coded. The correlation of 2362 and 3364, 3364 and 2220, 2220 and 1321, 5956 and 4197, 5956 and 2220 is negative, which proves that the machine learning algorithm training is abnormal, combined with the judgment of other non-negative relationships, it can be known that the sunspot region models of 3364, 2220 and 4197 need to be retrained. The correlation between other regions is between 0 and 0.5, which is determined to have weak correlation, and is not processed repeatedly.

Claims

1. A cross-validation based sunspot region correlation assessment method, characterized in that The method comprises the following steps: Step one: obtain the flare dataset, and perform dataset segmentation according to the sunspot region number to obtain each data subset; Step two: perform machine learning algorithm training on each data subset obtained in step one to obtain a machine learning model of the data subset; Step three: calculate the corresponding AUC index of each machine learning model in step two to obtain an A matrix, and the specific steps are as follows: Step three one: for each machine learning model f j , compute its value A j,k of the AUC metric on cross-validation, the specific procedure being: wherein, a mapping process for characterizing the computation of the AUC metric with an algorithm and a dataset, D k denotes the flare dataset at step one minus D j denotes all the data subsets left after step two, D j denotes the dataset corresponding to the blacked-out region numbered j; Step three two: through the method of step three one, all data subsets are traversed to obtain the A matrix: Step four: Use the A matrix calculated in step three to perform a correlation analysis to obtain a correlation index Corr, where: for a given sunspot region j and k, define Corr j,k represents the correlation of the model f j and the model f k , and is calculated as follows: wherein A (j，k) denotes the AUC metric value calculated using the model f j and the data subset D k , i.e. the element value in the j-th row and k-th column of the A matrix; A (k，j) denotes the AUC metric value calculated using the model f k and the data subset D j , i.e. the element value in the k-th row and j-th column of the A matrix; A (j,j) denotes the AUC metric value calculated using the model f j and the data subset D j , i.e. the element value in the j-th row and j-th column of the A matrix; A (k，k) denotes the AUC metric value calculated using the model f k and the data subset D k , i.e. the element value in the k-th row and k-th column of the A matrix; Step five: judge the correlation index Corr to obtain a given region correlation determination conclusion, and the specific steps are as follows: If the correlation index Corr j,k ≥ 1, it is proved that j and k have "strong correlation", and it is confirmed that they are the same sunspot region with repeated numbering. If the correlation index of sunspot region j and sunspot region k is 0.5 ≤ Corr j，k <1, it is proved that j and k have "certain correlation", and it is suspected that they are the same sunspot region with repeated numbering; If sunspot region j and sunspot region k have a correlation index 0 < Corr j,k If the correlation index is < 0.5, it is proven that j and k have a weak correlation and are not treated as the same sunspot region. If the correlation index Corr(j, k) of sunspot region j and sunspot region k is greater than 0 j,k If the correlation index Corr(j, k) of sunspot region j and sunspot region k is greater than 0, then the correlation of j and k is not able to be explained, the machine learning algorithm training is abnormal, and step two needs to be re-executed.

2. The cross-validation based sunspot region correlation assessment method of claim 1, wherein The specific steps of the step one are as follows: Step one one: obtain the flare dataset, and the flare dataset in a given time period is represented as: {(N0, X0, y0), (N0, X1, y1), (N1, X2, y2),..., (N j , X i , y i )} where X i represents the predictor of the i-th flare data, y i represents the label of whether the i-th flare data will flare within the next 48 hours, N j represents the sunspot number j in the flare data; Step one two: data set segmentation is performed to obtain each data subset, denoted as D j The data set corresponding to the black sub-region numbered j is denoted as D, and the process is represented as: D j = {(N j , X0, y0), (N j , X1, y1),..., (N j , X i , y i} Step 1-3: Through the process of step 1-2, the entire data set is divided into the form of {D0, D1, D2, …, D j} 3. The cross-validation based sunspot region correlation assessment method of claim 1, wherein The specific steps of the step two are as follows: for each data subset D j performing a machine learning algorithm training to obtain a machine learning model f for the data subset j , the machine learning algorithm training process is represented as: wherein, for characterizing a machine learning algorithm.

Images