CN105372723A - Solar flare forecasting method based on convolutional neural network model - Google Patents

Abstract

The invention discloses a solar flare forecasting method based on a convolutional neural network model, and the method comprises the steps: A, preparing observation raw data of an active region; B, building a depth forecasting model, extracting features from the observation data through employing a convolutional neural network, and forecasting whether the active region generates solar flare or not. The method can directly enable the observation raw data to serve as the input of the model, automatically extract a forecasting factor for the forecasting of the solar flare from the raw data through employing the strong learning capability of a depth neural network, builds a corresponding forecasting model, and achieves an ideal forecasting capability through the forecasting model.

Description

Solar flare prediction method based on convolutional neural network model

technical field

The present invention relates to the research technology of solar activity, in particular to a method for predicting solar flares based on a convolutional neural network model.

Background technique

The sun is the source of space weather. Intense solar activity may cause satellite failures, communication interruptions, navigation failures, and even power transmission network paralysis. If the situation of solar eruption activities can be accurately predicted in the future, disaster protection and treatment can be carried out in time.

Solar flare is a phenomenon of intense solar activity, and the prediction of solar flare has a long history of research and application. Based on the observation of sunspots, McIntosh proposed the morphological classification of sunspots (McIntosh classification) in "McIntosh, P.S.1990, Sol.Phys., 125, 251" in 1990. Based on the McIntosh classification of sunspots, an artificial establishment of Solar flare forecast expert system (Theo) with more than 500 rules. In 1994, Bornmann and Shaw calculated the relationship between the McIntsho typing of sunspots and solar flares in "Bornmann, P.L. & Shaw, D.1994, Sol. Phys., 150, 127", and established the regression model of McIntosh typing and solar flares . In 2007, Li Rong et al. used sunspots in "Li, R., Wang, H.-N., He, H., Cui, Y.-M., & Du, Z.-L.2007, ChJAA, 7,441" The area, magnetic classification, McIntosh classification, and 10 cm radio flux are used as model inputs, and the solar flare prediction model is established by using the support vector machine method. In 2009, Colak and Qahwaji used image processing technology and machine learning technology to establish an automatic solar activity prediction system (ASAP) in "Colak, T. & Qahwaji, R. 2009, SpaceWeather, 7, 06001", which automatically detects sunspots and predicts Automatically identify its McIntosh classification, and on this basis, use the neural network method to establish a solar flare forecast model. In 2012, Bloomfield et al. used the McIntosh classification of sunspots in "Bloomfield, D.S., Higgins, P.A., McAteer, R.T.J., & Gallagher, P.T. 2012, ApJL, 747, L41", and used Poisson statistical techniques to establish a solar flare prediction model.

Based on the solar flare events that have occurred, Wheatland in 2005 in "Wheatland, M.S. 2005, SpaceWeather, 3, 07003" only used the historical observation data of the solar flare itself, and established a solar flare prediction model using the Bayesian method.

Based on the solar photosphere magnetic field observation, Cui Yanmei et al. extracted from the MDI longitudinal magnetic field data in "Cui, Y.M., Li, R., Wang, H.N., & He, H.2007, Sol. The maximum horizontal gradient of the longitudinal magnetic field in the active area, the length of the center line, and the number of isolated singularities are three physical quantities, and the relationship between these physical quantities and solar flares is calculated. Based on these physical parameters, Wang Huaning et al. used the artificial neural network method in "Wang, H.N., Cui, Y.M., Li, R., Zhang, L.Y., & Han, H. forecast model. In 2007, Georgoulis and Rust defined the effective connectivity parameter Beff of the magnetic field in the active region in "Georgoulis, M.K. & Rust, D.M.2007, ApJ, 661, 109", which reflects the photosphere magnetic flux distribution in the active region and the connection characteristics of the photosphere magnetic field. In the article "Schrijver, C.J. 2007, ApJ, 655, 117" in 2007, Schrijver found that large solar flares are related to the strong gradient neutral line in the active region, and the energy of the flares comes from the free magnetic energy carried by the emerging fibrils. In order to describe the characteristics of the current carried by the magnetic fiber through the emergence of the photosphere, the author defined the unsigned magnetic field flux R value near the neutral line with a strong magnetic field and a large gradient within 15Mm, and calculated the relationship between this physical quantity and solar flares. In 2007, Leka and Barnes calculated a large number of magnetic field parameters (including magnetic inclination, magnetic field horizontal gradient, longitudinal current density, winding parameters, current helicity, magnetic shear angle, free magnetic energy density, etc.), it is found that among these parameters, the photosphere magnetic free energy has the strongest predictive ability, but any single photosphere magnetic field parameter is not enough to judge the active region Whether to produce large solar flares. In 2008, Barnes and Leka tested the literature "Georgoulis, M.K. & Rust, D.M.2007, ApJ, 661,109" and the literature "Schrijver, C.J.2007, ApJ, 655,117", pointing out that these physical parameters do not have significant differences when used in solar flare prediction, and the prediction ability of a single parameter is limited of. In 2010, Mason and Hoeksema used MDI longitudinal magnetic field observation in "Mason, J.P. & Hoeksema, J.T.2010, ApJ, 723, 634" to calculate the unsigned total magnetic flux in the active area, the length of the main center line, the effective segmentation distance of the active area, and the gradient weighting It is found that the gradient-weighted neutral line length combines the neutral line information reflecting the structure and the magnetic field gradient information reflecting the shear, so it can better reflect the occurrence of flares in the active region. In "McAteer, R.T.J., Gallagher, P.T., & Conlon, P.A. 2010, AdSpR, 45, 1067" in 2010, McAteer et al. described the complexity of the magnetic field in the active region of the photosphere from the two perspectives of the magnetic field energy spectrum and multifractal characteristics. And use these parameters to predict the occurrence of solar flares. In 2013, Ahmed et al. in "Ahmed, O.W., Qahwaji, R., & Colak, T.etal.2013, SoPh, 283, 157" used the Sun Monitoring Active Region Tracker (SMART) to extract the magnetic field characteristics of the active region, and used the neural network method to establish Solar flare forecasting model. In 2015, Bobra and Couvidat used the vector magnetic field observation of SDO in "BobraM.G.andCouvidatS.2015ApJ798135", extracted 25 physical parameters reflecting the characteristics of the active area, and used the support vector machine to establish a solar flare prediction model.

In summary, the modeling process of the existing solar flare forecast model is shown in Figure 1. Existing forecast models need to extract the parameters of the active area first to describe the characteristics of the active area, and use the extracted active area parameters as the input of the forecast model, and then give the forecast results.

However, the physical mechanism of solar flares is not very clear at present, and it is difficult to artificially construct the parameters of the active region, and the extracted parameters of the active region have not yet reached the ideal forecasting ability.

Contents of the invention

In view of this, the main purpose of the present invention is to provide a solar flare forecasting method based on a convolutional neural network model, which directly uses the observed original data as the input of the model, and utilizes the powerful learning ability of the deep neural network to automatically learn from the original The predictors used for solar flare forecasting are extracted from the data, and the corresponding forecasting model is established, so as to use it to achieve the ideal forecasting ability.

In order to achieve the above object, technical solution of the present invention is achieved in that way:

A solar flare forecasting method based on a convolutional neural network model, the forecasting method comprising:

A. The preparation steps of the original observation data in the active area;

B. Establish a deep prediction model, use convolutional neural network to extract features from observation data, and predict whether solar flares will occur in the active area.

Wherein, before step A, a step of extracting the parameters of the active area to characterize the characteristics of the active area is further included.

Described step A specifically comprises:

A1. Obtain the original data of the solar active region, that is, SOHO/MDI full sun longitudinal magnetic map;

A2. Steps for obtaining solar flare samples;

A3. Steps to determine the intensity of solar flares;

A4. Divide the data into training data and test data according to time, and convert the observation data in the active area into images of the same size between 0 and 1;

A5. Keep all the flare samples in the training data, and randomly select non-flare samples with the same number as the flare samples from the non-flare samples to form a new two-type balanced training data set.

The determination of the solar flare intensity is specifically:

A31. The intensity of solar flares is determined by the weighted sum of solar flares occurring within a specified time period, and its expression is:

I _tot =∑c+10∑m+100∑x

Among them: c, m and x represent the coefficients of C-class, M-class and X-class flares respectively.

The convolutional neural network described in step B is composed of 6 layers, specifically:

The first layer is the input layer, and the input layer is 100×100 photosphere magnetic field observation data;

The second layer is the convolutional layer. The convolutional layer includes a total of 100 filters, the filter size is 7, and the step size is 5; the output of the convolutional layer is 100 sets of 19×19 feature maps;

The third layer is the pooling layer, the filter size of the pooling layer is 3, the step size is 2, and the pooling method is to take the maximum value in the filter; the output of the pooling layer is 100 groups of 9×9 images;

The fourth layer is the first fully connected layer, and the number of nodes is 200;

The fifth layer is the second fully connected layer, and the number of nodes is 20;

The sixth layer is the output layer, the number of nodes is 2, corresponding to the two output states of the model respectively, the two output states are: solar flares will be generated in the future and solar flares will not be generated;

In the above model training process, the learning rate is set to 0.01, the momentum is set to 0.9, and the maximum number of cycles is set to 45000.

The extraction of features from the observation data, and forecasting whether solar flares will occur in the active area, specifically:

The prediction of whether the active area produces solar flares greater than a certain threshold is a typical binary prediction problem. For a binary prediction system, the prediction results are the following four possible results:

A sample that is itself a positive class and is correctly predicted as a positive class is called a correct positive; a sample that is itself a negative class and is correctly predicted as a negative class is called a correct negative; a sample that is itself a positive class and is incorrectly predicted as a positive class A sample of the negative class is called a false negative; a sample that is itself a negative class but is incorrectly predicted as a positive class is called a false positive.

In the solar flare forecast, the flare samples are regarded as positive samples, and the non-flare samples are regarded as negative samples; according to the four types of output of the forecast results, the following four indicators are defined to describe the performance of the forecast model:

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}$

Among them: N _TP is the number of correct positive samples, N _FN is the number of wrong negative samples;

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}$

Among them: N _TN is the number of correct negative samples, N _FP is the number of wrong positive samples;

TSS = TPrate - FPrate

Where: FPrate=1-TNrate.

$\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}}$

in:

N=N _TP +N _TP +N _TP +N _TP ,

$\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ,</span>$

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ;</span>$

The TPrate and TNrate are used to evaluate the accuracy of the flare forecast and the accuracy of the non-flare forecast respectively; the index TSS is not sensitive to the ratio of the number of flare samples and the number of non-flare samples; the HSS is used to reflect the forecast of the forecast model Ability increase over random guessing.

After the step B, further comprising:

C. Steps in evaluating the forecast model.

The solar flare prediction method based on the convolutional neural network provided by the present invention has the following advantages:

By using the newly established solar flare forecast model, the present invention no longer needs to artificially extract the physical parameters of the active area, but directly learns the feature expression of the active area from the original data, greatly reducing the influence of human factors on the forecast, reducing the It reduces the application difficulty of the forecast model and improves the application range of the forecast model.

Description of drawings

Figure 1 is the modeling process of the existing solar flare forecast model;

Fig. 2 is a schematic diagram of the model modeling flow chart of the solar flare forecasting method based on the convolutional neural network model of the present invention;

3 is a schematic diagram of a convolutional neural network used in an embodiment of the present invention;

Fig. 4 is a schematic diagram of the performance test results of the solar flare prediction model of the present invention.

detailed description

The solar flare prediction method based on the convolutional neural network of the present invention will be further described in detail below in conjunction with the accompanying drawings and the embodiments of the present invention.

Fig. 2 is a schematic diagram of the model modeling process of the solar flare prediction method based on the convolutional neural network model of the present invention.

As shown in Figure 2, the solar flare prediction method based on the convolutional neural network model of the present invention directly uses the observed original data as the input of the model, and uses the powerful learning ability of the deep neural network to automatically extract the The predictors of solar flare forecast, and the establishment of corresponding forecast models. The new solar flare forecast model no longer needs to artificially extract the physical parameters of the active area, but directly learns the characteristic expression of the active area from the original data. The difficulty of application increases the scope of application of the forecast model.

The invention uses the convolutional neural network to automatically extract features from the observation data, and provides the result of predicting whether the solar flare activity will erupt in the active area in the next 48 hours. The invention uses the magnetic field information of the solar active area to predict the occurrence of solar flares.

The solar flare prediction method based on the convolutional neural network model of the present invention will be described in detail below with reference to FIGS. 1 to 4 . The method comprises the steps of:

Step 10: the step of extracting the parameters of the active area to characterize the characteristics of the active area. This step is a well-known technology, and will not be repeated here.

Step 11: The preparation step of the original observation data in the active area.

In the embodiment of the present invention, the raw data of the solar active regions come from the SOHO/MDI full-helix longitudinal magnetic map with a time resolution of 96 minutes. The SOHO/MDI full sun longitudinal magnetogram can be downloaded from ftp://soi-ftp.stanford.edu/pub/magnetograms/.

The present invention uses the information of a total of 1055 NOAA active areas that occurred within 30° of the sun's surface from May 1996 to June 2007. For the active area within 30° of the solar surface, we ignore its projection effect.

Steps to get a solar flare sample. The solar flare samples in the embodiment of the present invention are obtained from National Geophysical Data Center (NGDC) ftp://ftp.ngdc.noaa.gov/STP/SOLAR_DATA/SOLAR_FLARES/FLARES_XRAY/.

Determines solar flare intensity. The solar flare intensity is determined by a weighted sum of solar flares occurring within a specified time period. Its expression is:

I _tot =∑c+10∑m+100∑x

Among them: c, m and x represent the coefficients of C-class, M-class and X-class flares respectively.

For example, if 3 flares (C4, M3 and X2) occur in a given time period, the total intensity of solar flares is 234 (ie 4+10×3+100×2).

In a given forecast period, if the total intensity of solar flares is greater than a given threshold, the sample is considered a flare sample, otherwise, the sample is considered a non-flare sample. In this embodiment, the forecast period is 48 hours, and the threshold of Itot is set to 10, which is equivalent to generating an M1.0 level flare.

In order to objectively evaluate the performance of the forecasting model, all data from 1996 to 2007 are divided into training data and test data according to time. The data from 1996 to 2006 is used as training data, and the data from 2007 is used as test data. The training data is used to establish the prediction model, and the test data is used to evaluate the performance of the prediction model.

In order to meet the requirements of the modeling method, the observation data of all active areas are converted into images of the same size (100×100 pixels) with values between 0 and 1.

Since the number of non-flare samples in the training samples is much larger than the number of flare samples, if the original data is directly used to train the prediction model, the model will usually be biased towards the category with a large number of samples. Therefore, the present invention retains all the flare samples in the training data, and randomly selects non-flare samples with the same number as the flare samples from the non-flare samples to form a new two-type balanced training data set. The test data remains unchanged.

Step 12: The step of establishing the depth prediction model.

The invention uses a convolutional neural network to extract features from observation data, and predicts whether solar flares are generated in the active area. FIG. 3 is a schematic diagram of a convolutional neural network used in an embodiment of the present invention. As shown in Figure 3, the convolutional neural network is composed of 6 layers, specifically:

The first layer is the input layer, and the input layer is 100×100 photosphere magnetic field observation data.

The second layer is a convolutional layer, which includes 100 filters in total, with a filter size of 7 and a step size of 5. Therefore, the output of the convolutional layer is 100 sets of 19×19 feature maps.

The third layer is the pooling layer, the filter size of the pooling layer is 3, the step size is 2, and the pooling method is to take the maximum value in the filter. The output of the pooling layer is 100 sets of 9×9 images.

The fourth layer is the first fully connected layer with 200 nodes.

The fifth layer is the second fully connected layer with 20 nodes.

The sixth layer is the output layer, and the number of nodes is 2, corresponding to the two output states of the model. That is, there will be solar flares and no solar flares in the future.

In the above model training process, the learning rate is set to 0.01, the momentum is set to 0.9, and the maximum number of cycles is set to 45000.

The invention gives the forecast of whether the active area produces solar flares greater than a certain threshold, which is a typical binary forecast problem. For a binary forecasting system, there are four possible results in its forecasting results, as shown in Table 1.

A sample that is itself a positive class and is correctly predicted as a positive class is called a correct positive, a sample that is itself a negative class and is correctly predicted as a negative class is called a correct negative, and a sample that is itself a positive class and is incorrectly predicted as a positive class A sample of the negative class is called a false negative, and a sample that is itself a negative class but is incorrectly predicted as a positive class is called a false positive.

Table 1: Four possible outcomes of binary forecast results

predicted positive class predicted negative class true class True Positive (TP) false negative (FN) real negative class false positive (FP) True Negative (TN)

In solar flare forecasting, we regard flare samples as positive samples, and non-flare samples as negative samples. Based on the four types of output shown in Table 1, the following four indicators are defined to describe the performance of the forecasting model:

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}$

Among them: N _TP is the number of correct positive samples, and N _FN is the number of wrong negative samples.

$\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}$

Among them: N _TN is the number of correct negative samples, and N _FP is the number of wrong positive samples.

TSS = TPrate - FPrate

Where: FPrate=1-TNrate.

$\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}}$

Among them: N=N _TP +N _TP +N _TP +N _TP ,

$\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ,</span>$

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; .</span>$

Among the above four evaluation indicators, TPrate and TNrate are used to evaluate the accuracy of flare forecast and the accuracy of non-flare forecast respectively. In order to give an overall evaluation of the entire forecasting model, we also need to use two evaluation indicators, trueskillstatistic (TSS) and Heidkeskillscore (HSS). The TSS is not sensitive to the ratio of the number of flare samples to the number of non-flare samples, while the HSS reflects the increase in the forecasting ability of the forecast model compared to random guessing.

Step 13: The step of evaluating the forecasting model.

The embodiments of the present invention use the active area and flare data in 2007 as test data. The test data contains 1172 flare samples and 8828 non-flare samples. The convolutional neural network in the present invention is trained 45,000 times, and the performance of the model is tested every 1,000 times. The test results are shown in FIG. 4 .

It can be seen from Figure 4 that in the first 5000 training sessions, the forecast model did not learn useful information from the data;

From the 6000th time, the convolutional neural network learns useful features from the observation data, and the model begins to have the ability to predict.

The performance of the 6000th and 8000th solar flare prediction is shown in Table 3 and Table 5. During the training process of the forecast model from 6000 to 8000 times, more attention is given to the flare samples, and at the same time, the forecast of the non-flare samples is accurate The rate is reduced, which requires us to choose a forecast model that meets the task requirements according to the needs of different tasks. The model tends to be stable after training for 10000 times, as shown in Figure 4.

Table 2: Output results of the flare forecast model at the 6000th step

predicted positive class predicted negative class true class Correct affirmation(707) False Negatives (465) real negative class False Positives (1129) correct negation (7699)

Table 3: Performance test results of the flare forecast model at the 6000th step

performance evaluation index Test Results TP rate 60% TN rate 87% ACC 84% HSS 0.38 TSS 0.48

Table 4: Output results of the flare forecast model at the 8000th step

predicted positive class predicted negative class true class Correct Affirmation(753) False Negatives (419) real negative class False Positives (1884) correct negation (6944)

Table 5: Performance test results of the flare forecast model at the 8000th step

performance evaluation index Test Results TP rate 64% TN rate 79% ACC 77% HSS 0.28 TSS 0.43

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention.

Claims (8)

1. A solar flare forecasting method based on a convolutional neural network model, characterized in that the forecasting method comprises: A. The preparation steps of the original observation data in the active area; B. Establish a deep prediction model, use convolutional neural network to extract features from observation data, and predict whether solar flares will occur in the active area. 2. The solar flare forecasting method based on the convolutional neural network model according to claim 1, characterized in that, before step A, further includes the step of extracting the parameters of the active area to characterize the characteristics of the active area. 3. The solar flare forecasting method based on the convolutional neural network model according to claim 1, wherein said step A specifically comprises: A1. Obtain the original data of the solar active region, that is, SOHO/MDI full sun longitudinal magnetic map; A2. Steps for obtaining solar flare samples; A3. Steps to determine the intensity of solar flares; A4. Divide the data into training data and test data according to time, and convert the observation data in the active area into images of the same size between 0 and 1; A5. Keep all the flare samples in the training data, and randomly select non-flare samples with the same number as the flare samples from the non-flare samples to form a new two-type balanced training data set. 4. the solar flare forecasting method based on convolutional neural network model according to claim 3, is characterized in that, described determining solar flare intensity, is specifically: A31. The intensity of solar flares is determined by the weighted sum of solar flares occurring within a specified time period, and its expression is: I _tot =∑c+10∑m+100∑x Among them: c, m and x represent the coefficients of C-class, M-class and X-class flares respectively. 5. the solar flare forecasting method based on the convolutional neural network model according to claim 1, is characterized in that, the convolutional neural network described in step B is made up of 6 layers, specifically: The first layer is the input layer, and the input layer is 100×100 photosphere magnetic field observation data; The second layer is the convolutional layer. The convolutional layer includes a total of 100 filters, the filter size is 7, and the step size is 5; the output of the convolutional layer is 100 sets of 19×19 feature maps; The third layer is the pooling layer, the filter size of the pooling layer is 3, the step size is 2, and the pooling method is to take the maximum value in the filter; the output of the pooling layer is 100 groups of 9×9 images; The fourth layer is the first fully connected layer, and the number of nodes is 200; The fifth layer is the second fully connected layer, and the number of nodes is 20; The sixth layer is the output layer, the number of nodes is 2, corresponding to the two output states of the model respectively, the two output states are: solar flares will be generated in the future and solar flares will not be generated; In the above model training process, the learning rate is set to 0.01, the momentum is set to 0.9, and the maximum number of cycles is set to 45000. 6. The method for predicting solar flares based on the convolutional neural network model according to claim 5, wherein the feature is extracted from the observation data, and whether solar flares are produced in the active area is predicted, specifically: The prediction of whether the active area produces solar flares greater than a certain threshold is a typical binary prediction problem. For a binary prediction system, the prediction results are the following four possible results: A sample that is itself a positive class and is correctly predicted as a positive class is called a correct positive; a sample that is itself a negative class and is correctly predicted as a negative class is called a correct negative; a sample that is itself a positive class and is incorrectly predicted as a positive class A sample of the negative class is called a false negative; a sample that is itself a negative class but is incorrectly predicted as a positive class is called a false positive. 7. The solar flare forecast method based on the convolutional neural network model according to claim 6, wherein, in the solar flare forecast, the flare sample is used as a positive sample, and the non-flare sample is used as a negative sample; according to the forecast result Four types of output, define the following four indicators to describe the performance of the forecasting model: $\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; PR</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}$ Among them: N _TP is the number of correct positive samples, N _FN is the number of wrong negative samples; $\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}}$ Among them: N _TN is the number of correct negative samples, N _FP is the number of wrong positive samples; TSS = TPrate - FPrate Where: FPrate = 1 - TNrate. $\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}}$ Among them: N=N _TP +N _TP +N _TP +N _TP , $\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}}{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; ,</span>$ $\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ;</span>$ The TPrate and TNrate are used to evaluate the accuracy of the flare forecast and the accuracy of the non-flare forecast respectively; the index TSS is not sensitive to the ratio of the number of flare samples and the number of non-flare samples; the HSS is used to reflect the forecast of the forecast model Ability increase over random guessing. 8. the solar flare forecasting method based on convolutional neural network model according to claim 1, is characterized in that, after described step B, further comprises: C. Steps in evaluating the forecast model.