KR102562273B1 - Apparatus for forecasting 24-hour global total electron content using deep learning based on conditional generative adversarial neural network and method thereof - Google Patents

Abstract

An apparatus and method for predicting global total electron count after 24 hours using a conditional adversarial neural network deep learning technique are disclosed.
This device is a device for predicting a total electron content image after 24 hours using a current total electron content image, and includes a learning unit and a prediction unit. The learning unit trains the prediction model by performing machine learning-based training on a prediction model that predicts a total electron number content image 24 hours after a specific point in time by using a total electron number content image of the past. The prediction unit applies the current total electron number content image to the prediction model to predict the total electron number content image 24 hours from now.

Description

Apparatus for forecasting 24-hour global total electron content using deep learning based on conditional generative adversarial neural network and method thereof}

The present invention relates to an apparatus and method for predicting total global electron content after 24 hours using a conditional adversarial neural network deep learning technique.

Total Electron Content (TEC) means the total number of electrons along a path between a wireless transmitter and a receiver, and the unit (TECU) is defined as 10 ¹⁶ electrons/m ² .

As TEC is one of the key parameters of space weather and is used as an indicator of ionospheric disturbances, accurate measurements of TEC provide useful information for satellite communications, navigation, defense and aviation.

These TEC forecasting methods include empirical models using calculation formulas based on past observed values, physics-based models using motion, energy, and gas equations related to the thermosphere and ionosphere, and models based on observational data and physics. It can be divided into this combined assimilation model.

First, the empirical model using observed statistical values has been used as a universal forecasting method since the past, but errors appear during periods of frequent solar and geomagnetic activity and anomalies in the equatorial region.

In addition, in the case of physics-based models and assimilation models, the more precise the formulas and grids are, the more time required for computer calculations, and, like the empirical models, errors for anomalies appear.

Since 1998, the International GNSS service (IGS) has been providing maps of global total electron content. This data is produced through the steps of calculation, analysis, and verification with GNSS observation data, and is the data with the highest accuracy. These data are provided in two forms, Rapid TEC and Final TEC. Final TEC is known to have higher accuracy, and is used as data to verify the value of the total electron content prediction model.

However, existing IGS TECs have a problem in that they cannot predict TEC images 24 hours later without additional data such as solar and geomagnetic activity indices.

The problem to be solved by the present invention is the global total after 24 hours using a conditional adversarial neural network deep learning technique that can well predict TEC images after 24 hours using only previous TEC images without additional data such as solar and geomagnetic activity indices. It is to provide an electron number content predicting device and method therefor.

In order to achieve the objects of the present invention as described above and realize the characteristic effects of the present invention described later, the characteristic configuration of the present invention is as follows.

According to one aspect of the present invention, a device for predicting the total number of electrons is provided, which includes:

A device for predicting a total electron content image after a specific time using a current total electron content image, wherein the total electron content after a specific time from a specific point in time using a past total electron content image A learning unit for learning the prediction model by performing machine learning-based training on a prediction model that predicts an image, and applying the current total number of electrons content image to the prediction model to obtain the total number of electrons after the specific time from the present A prediction unit for predicting an image is included.

According to another aspect of the present invention, a method for predicting the total number of electrons is provided, the method comprising:

A method of predicting a total electron number content image after a specific time using a current total electron number content image, wherein the total electron number content image after the specific time is predicted from a specific time point using a past total electron number content image. Learning the predictive model by performing machine learning-based training on the model, between the current total electron number content image, the total electron number content image before the specific time from the present, and the current total electron number content image A step of inputting a difference map representing a difference into the predictive model, and acquiring an image generated by the predictive model and outputting it as a total electron count content image after the specific time from the present.

According to the present invention, deep learning models can be successfully applied to the prediction of global TEC images.

also. TEC images after 24 hours can be well predicted using only previous TEC images without additional data such as solar and geomagnetic activity indices.

It also outperforms the 1-day CODE prediction model during solar maximum and solar minimum when considering all three metrics (RMSE, BIAS and STD).

1 is a schematic block diagram of an apparatus for predicting total global electron count content according to an embodiment of the present invention.
FIG. 2 is a diagram showing a specific configuration of the predictive model shown in FIG. 1, based on an operation used during training.
3 (a) is an example of a current IGS TEC image, (b) is an example of a difference map, (c) is an image generated by a generator, and (d) shows an example of an actual image.
4 is a diagram showing an example of a test of the global total electron count prediction device according to an embodiment of the present invention.
5 is an average of RMSE, BIAS, and STD between a TEC image and an IGS TEC image predicted by a global total electron count prediction device according to an embodiment of the present invention and an average between a 1-day CODE predicted TEC image and an IGS TEC image. It is a diagram showing RMSE, BIAS and STD.
6 shows monthly average RMSE, BIAS, and STD between output TEC images and IGS TEC images predicted by the global total electron count prediction device according to an embodiment of the present invention, and between 1-day CODE predicted TEC images and IGS TEC images. shows the monthly average of
7 is a schematic flowchart of a method for predicting the global total number of electrons according to an embodiment of the present invention.

Hereinafter, with reference to the accompanying drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily carry out the present invention. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description have been omitted, and similar reference numerals have been attached to similar parts throughout the specification.

Throughout the specification, when a certain component is said to "include", it means that it may further include other components without excluding other components unless otherwise stated. In addition, terms such as “… unit”, “… unit”, and “module” described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software. there is.

Devices described in the present invention are composed of hardware including at least one processor, memory device, communication device, and the like, and a program to be executed in combination with the hardware is stored in a designated place. The hardware has the configuration and capability to implement the method of the present invention. The program includes instructions implementing the operating method of the present invention described with reference to the drawings, and implements the present invention in combination with hardware such as a processor and a memory device.

Hereinafter, an apparatus for predicting total global electron count content after 24 hours using a conditional adversarial neural network deep learning technique according to an embodiment of the present invention will be described.

1 is a schematic block diagram of an apparatus for predicting total global electron count content according to an embodiment of the present invention.

As shown in FIG. 1 , the apparatus 100 for predicting the global total number of electrons according to an embodiment of the present invention includes an input unit 110, a difference map generator 120, a storage unit 130, and a learning unit 140. ), a prediction model 150, a prediction unit 160, and an output unit 170. At this time, since the global total electron count prediction device 100 shown in FIG. 1 is only one embodiment of the present invention, the present invention is not limitedly interpreted through FIG. 1, and various embodiments of the present invention Accordingly, it may be implemented differently from FIG. 1 .

The input unit 110 receives TEC data from the outside. For example, the input unit 110 receives IGS TEC data provided from IGS. These IGS TEC data will be obtained through the National Aeronautics and Space Administration (NASA) Crustal Dynamics Data Information System (CDDIS) website ( ftp://cddis.nasa.gov/gnss/products/ionex/ ). can The acquired IGS TEC data has (73 × 71) resolution in geographic longitude and latitude of 5° × 2.5°.

The input unit 110 may perform data preprocessing in order to use input IGS TEC data in the prediction model 1150 . Such pre-processing includes, for example, processing to create a 72 × 71 image by removing overlapping values from the timeline in an image corresponding to IGS TEC data (hereinafter referred to as 'IGS TEC image'), appropriate padding data ( processing to make a 128 × 128 image by adding the same data on the left and right sides and a value of 0 on the top and bottom). This preprocessing may be changed corresponding to the image used in the predictive model 160 .

The input unit 110 stores input IGS TEC data in the storage unit 130 together with a corresponding date and time.

The difference map generation unit 110 determines the difference between a specific IGS TEC image corresponding to specific IGS TEC data stored in the storage unit 130 and an IGS TEC image corresponding to IGS TEC data from 1 day ago, that is, 24 hours ago, from 1 day ago. Create a difference map representing

The storage unit 130 stores IGS TEC data transmitted through the input unit 110 .

Also, past IGS TEC data may be previously stored in the storage 130 for training and testing of the predictive model 150 . In an embodiment of the present invention, IGS TEC data from at least 2003 to the present may be stored.

The learning unit 140 trains the predictive model 150 using past IGS TEC data stored in the storage unit 130 through the input unit 110, and uses the input IGS TEC data and the difference map to generate 24 Learning to predict the TEC image after time is performed.

The prediction model 150 is a prediction model learned through deep learning. For example, a conditional adversarial network (cGAN) based model, which is a deep learning-based image conversion model, can be used. However, it is not limited to this. A detailed configuration of the predictive model 150 will be described later.

The prediction unit 160 obtains a TEC image after 24 hours, which is generated by applying the current IGS TEC image input through the input unit 110 and the difference map to the predictive model 150, so that the current IGS TEC image after 24 hours predict TEC.

The prediction unit 160 outputs the TEC image predicted by the prediction model 150 after 24 hours to the output unit 170 .

The TEC image predicted by the prediction unit 160 after 24 hours may be stored in the storage unit 130 .

The output unit 170 may externally display the TEC image predicted by the prediction unit 160 after 24 hours or transmit it to an external device.

FIG. 2 is a diagram showing a specific configuration of the predictive model 150 shown in FIG. 1, based on an operation used during training.

Referring to FIG. 2, the prediction model 150 according to an embodiment of the present invention is a conditional generative adversarial network (cGAN) model, and includes two networks, for example, a generator 151 and a discriminator 153. .

The generator 151 receives the current IGS TEC image (CI) and the difference map (DI) from the learning unit 140 and generates a fake image (GI) similar to the current IGS TEC image (CI). Here, the current IGS TEC image C1 is an image corresponding to past IGS TEC data stored in the storage unit 130, and for convenience of description, the term 'current' is used for a specific point in time in the past.

The generator 151 transfers the generated image GI, the current IGS TEC image CI input from the learning unit 140, and the difference map DI to the discriminator 153. An example of the current IGS TEC image (CI) input from the learning unit 140 is shown in (a) of FIG. 3, and an example of the difference map (DI) is shown in (b) of FIG. 3, and the generator ( 151) is shown in (c) of FIG.

The discriminator 153 also discriminates between the real image RI, which is the target image input from the learning unit 140, and the fake image generated by the generator 141, that is, the generated image GI, and the discrimination result is generated by the generator 151 ) as feedback. At this time, the determination result is 'fake (0)', which means that the real image (RI) and the generated image (GI) are determined to be different, or 'real', which means that the real image and the generated image are determined to be the same. (1) is one of two things. The actual image used in the discriminator 153 is an IGS TEC image 24 hours after the current IGS TEC image input from the learning unit 140 to the generator 151 (the present in the above). An example of such an actual image is It is shown in (d) of FIG.

The discriminator 153 receives the image GI generated by the generator 151, the current IGS TEC image CI and the difference map DI input from the input unit 120, and input from the learning unit 140. 24 hours later real image RI of the current IGS TEC image CI input for training, current IGS TEC image CI input from the input unit 120 and difference map DI It receives and discriminates between the generated image (GI) and the actual image (RI). Here, the image (GI) generated by the generator 151 and the current IGS TEC image (CI) and difference map (DI) input from the input unit 120 are referred to as a fake pair, and the real image (RI) and The current IGS TEC image (CI) and the difference map (DI) input from the learning unit 140 are referred to as a real pair.

The generator 151 reflects the discrimination result fed back from the discriminator 153 and continues learning to generate a fake image GI using the current IGS TEC image CI. Specifically, when the discrimination result fed back from the discriminator 153 is fake (0), the generator 151 generates fake images (GI) until the discrimination result fed back from the discriminator 153 becomes real (1). continue learning to generate

As described above, the prediction model 150 trained and learned by the learning unit 140 is input through the input unit 110 when the current IGS TEC image to be predicted and the corresponding difference map are input through the prediction unit 160. That is, when it is desired to predict the TEC 24 hours from now, the image generated using the current IGS TEC image and the difference map corresponding to the current IGS TEC image by the learned generator 151 is the predicted TEC image 24 hours later, the prediction unit ( 160) is output.

On the other hand, to train the aforementioned predictive model, three loss (target) functions are used.

First, the L ₁ function used to minimize the error between the actual image RI, which is the target image, and the image GI generated by the generator 141 may be expressed as in [Equation 1] below.

where G denotes the generator 151, x and G(x) denote the input by the generator 151, i.e., the current IGS TEC image (CI) and the output, i.e., the generated image (GI), and y is Indicates the actual image RI, which is the target image.

Next, the loss function for cGAN can be expressed as the following [Equation 2].

where D denotes discriminator 153, D(x, y) denotes discriminator 153's estimate of the probability that the actual pair is real (or genuine), and D(x, G(x)) is It represents discriminator 153's estimate of the probability that a fake pair is real.

In cGAN's predictive model, D (discriminator) calculates the probability that an image pair belongs to 0 (fake pair) or 1 (real pair), and D (discriminator) tries to minimize it while G (generator) tries to minimize it. try to maximize As a result, D (the discriminator) and G (the generator) play a min-max game. At this time, the final target (G*) can be expressed as the following [Equation 3].

[Equation 3]

Here, λ represents the relative weight of L _cGAN loss and L ₁ loss.

The aforementioned relative weight can be, for example, 100, and these weights are the convolution layer and the convolution-transpose layer using a normal distribution with a mean of 0.0 and a standard deviation of 0.02. , can be initialized as a batch-normalization layer using a normal distribution with a mean of 1.0 and a standard deviation of 0.02. In addition, an ADAM solver having a learning rate of 2×10 ^-4 , a momentum β ₁ of 0.5, and a momentum β ₂ of 0.999 may be used as an optimizer.

The learning unit 140 trains the prediction model 150 for 22 epochs, which is approximately 1,000,000 iterations for the training as described above, and consists of a generator (D) and a discriminator (G) every 20,000 iterations The predictive model 150 to be stored is stored in the storage unit 130 or other storage devices.

In an embodiment of the present invention, in order to train the prediction model 150 of the global total electron content prediction apparatus 100 according to the embodiment of the present invention described with reference to FIG. 1, actual data, for example, 2003 From 2012 to 2012, we used 2-hour resolution TEC images (approximately 43,800 images) covering solar maxima and minima. That is, in a state in which 2-hour resolution TEC images (approximately 43,800 images) covering solar maxima and minima from 2003 to 2012 are stored in the storage unit 130 as training images, the learning unit 140 inputs the input unit 110 Each of the training images delivered through ) was input to the prediction model 150 as a current IGS TEC image and trained.

In this way, in order to test the prediction model 150 of the trained global total electron content prediction device 100, TEC images (approximately 8,736 images) of solar maximum from 2013 to 2014 and 2017 to 2018 Until now, we have used two datasets consisting of TEC images of solar minimums (approximately 8,736 images). That is, in a state in which the first data set, which is TEC images of solar maximum from 2013 to 2014, is stored in the storage unit 110 as the first test image, each of the first test images is used as the current IGS TEC image, and the learning unit 140 ) was entered into the prediction model 150 to test the prediction model 150 for the solar maximum period.

In addition, in a state where a second data set, which is a TEC image of the solar minimum from 2017 to 2018, is stored in the storage unit 110 as a second test image, each of the second test images is a learning unit as a current IGS TEC image ( 140) into the prediction model 150 to test the prediction model 150 against the solar minimum.

An example of the above test will be described with reference to the drawings.

4 is a diagram showing an example of a test of the global total electron count prediction apparatus 100 according to an embodiment of the present invention.

Referring to FIG. 4, the global TEC after 24 hours was predicted using the input image (current IGS TEC image) shown in (a) observed at 2 am on October 1, 2013, and shown in (c) The output image (generated image) is the predicted image, (b) is the actual image, that is, the image after 24 hours from the time of the input image shown in (a), and (d) is the actual image (b) and the predicted image. is a difference map between the image (c).

When the aforementioned input image (a) is captured, the Kp index is 0, and when the target image (b) is captured, the Kp index is 5, which corresponds to a geomagnetic storm. The predicted output image (c) agrees well with the IGS TEC after 24 hours, which is the target image (b). In particular, the prediction model 150 of the apparatus 100 for predicting the global total number of electrons according to an embodiment of the present invention successfully generates a peak structure in the equatorial region. The root mean square error (RMSE), bias (BIAS), and standard deviation (STD) between the output image (c) and the target image (b) are 2.92 TECU, -0.52 TECU, and 2.87 TECU, respectively.

On the other hand, the result predicted by the global total electron count prediction device 100 according to an embodiment of the present invention can be compared with the result of a well-known CODE prediction model to confirm the prediction accuracy. CODE is one of the IAACs and uses GNSS data to perform global TEC predictions.

5 is an average of RMSE, BIAS, and STD between a TEC image and an IGS TEC image predicted by a global total electron count prediction device according to an embodiment of the present invention and an average between a 1-day CODE predicted TEC image and an IGS TEC image. It is a diagram showing RMSE, BIAS and STD.

The comparison in Figure 5 considers three test data sets: All test data (2013-2014 and 2017-2018), Solar maximum (2013-2014), and Solar minimum (2017-2018).

Referring to FIG. 5 , it can be seen that the prediction results of the apparatus 100 for predicting the global total number of electrons according to an embodiment of the present invention are superior to the results of the 1-day CODE prediction model for all test data. The predicted results of the global total electron count prediction device 100 according to an embodiment of the present invention are also better than those of the daily CODE prediction model both during solar maximum and solar minimum.

On the other hand, the TEC shows regular changes in daily, seasonal, annual, solar rotation cycle (27 days), and solar cycle (about 11 years) changes. 6 shows monthly average RMSE, BIAS, and STD between output TEC images and IGS TEC images predicted by the global total electron count prediction apparatus 100 according to an embodiment of the present invention, and 1-day CODE predicted TEC images and IGS Shows the monthly average between TEC images.

Referring to FIG. 6, except for the two data points collected in March and April 2014, all prediction results from the global total electron count prediction device 100 according to an embodiment of the present invention are CODE 1 day indicates that it is superior to the results of the predictive model. These results may be related to the duration of the training data. Solar activity in March and April 2014 is most active during the 24 solar cycle. Since the training data in the global total electron count prediction device 100 according to the embodiment of the present invention has been included since 2003, it is not well trained for the most active case.

According to an embodiment of the present invention, a deep learning model can be successfully applied to the prediction of global TEC images. Second, TEC images after 24 hours can be well predicted using only previous TEC images without additional data such as solar and geomagnetic activity indices, and also, given the three metrics (RMSE, BIAS and STD), solar maximum and solar minimum It is better than the 1-day CODE prediction model during the period.

On the other hand, the apparatus 100 for predicting global total electron count according to an embodiment of the present invention shows that a deep learning model based on an image conversion method is effective in predicting a future image using a previous image. Those skilled in the art will easily understand that this method can be applied to image prediction in other scientific fields.

Hereinafter, a global total electron count prediction method according to an embodiment of the present invention will be described with reference to the drawings.

7 is a schematic flowchart of a method for predicting the global total number of electrons according to an embodiment of the present invention. The global total electron count prediction method shown in FIG. 7 may be performed by the total electron count prediction device 100 described with reference to FIGS. 1 to 6 .

Referring to FIG. 7 , first, the prediction model 150 is learned using past IGS TEC data stored in the storage unit 130 (S100). Specifically, the learning unit 140 receives an IGS TEC image (current IGS TEC image) at a specific time in the past from the storage unit 130 through the input unit 110, an IGS TEC image at a specific time in the past, and a specific time in the past. A difference map between IGS TEC images before 24 hours and an IGS TEC image (actual image) 24 hours later from a specific point in the past are input to the prediction model 150 to predict a TEC image 24 hours later for the current IGS TEC image. The predictive model 150 may be learned by repeatedly performing learning on a plurality of past data.

Next, for actual prediction, the current IGS TEC image is input to the prediction model 150 (S110). Specifically, the input unit 110 connects the current IGS TEC image, the current IGS TEC image generated by the difference map generator 120, and the IGS TEC stored in the storage unit 130 24 hours before the present. Input the difference map of to the predictive model 150.

After that, the TEC image predicted by the prediction model 150 after 24 hours is acquired (S120), and the TEC image predicted from the prediction model 150 after 24 hours is output (S130). That is, the predicted TEC image 24 hours later output from the generator 151 of the predictive model 150 is externally displayed through the output unit 170 or transferred to an external device.

The embodiments of the present invention described above are not implemented only through devices and methods, and may be implemented through programs that realize functions corresponding to the configuration of the embodiments of the present invention or a recording medium on which the programs are recorded.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements made by those skilled in the art using the basic concept of the present invention defined in the following claims are also included in the scope of the present invention. that fall within the scope of the right.

Claims (12)

A device for predicting a total electron content image after a specific time using a current total electron content image,
A learning unit for learning the prediction model by performing deep learning-based training on a prediction model that predicts a total electron number content image after a specific time from a specific point in time using a total electron number content image of the past, and
A prediction unit for applying the current total electron number content image to the prediction model to predict the total electron number content image after the specific time from the present
Including,
The learning unit, among the total electron number content images of the past, the total electron number content image at the specific time point (hereinafter referred to as 'learning image'), the total electron number content image before the specific time from the specific time point, and the learning image A difference map representing the difference between the specific time and a total electron content image (hereinafter referred to as 'actual image') after the specific time from the specific time point is input to perform conditional generative adversarial network (cGAN)-based learning to perform learning at the specific time Create a predictive model that predicts the total electron number content image after
The prediction unit applies a difference map representing a difference between the current total number of electrons content image and the total number of electrons content image before the specific time from the present and the current total number of electrons content image to the prediction model to determine the current total number of electrons. Predicting the total electron number content image after the specific time,
A device for predicting the total number of electrons. delete According to claim 1,
The prediction model is a conditional generative adversarial network (cGAN)-based prediction model,
A generator receiving the training image and the difference map as inputs and generating a fake image corresponding to the training image; and
A discriminator for determining authenticity by comparing the fake image generated by the generator with the real image, and feeding back the determination result to the generator.
Including, total electron number prediction device. According to claim 3,
The learning unit continues to learn the learning image until a discrimination result fed back from the discriminator to the generator indicates that the fake image is real.
A device for predicting the total number of electrons. According to claim 4,
The prediction unit,
After inputting the current total electron number content image into the generator of the prediction model, the image generated by the generator is used as the total electron number content image after the specific time predicted for the current total electron number content ,
A device for predicting the total number of electrons. According to claim 3,
The loss function L ₁ used in training the predictive model is the following relational expression

where G represents the generator, x and G(x) represent input and output by the generator, respectively, and y represents the real image
is expressed as
The loss function L _cGAN for cGAN is the relation

where D represents the discriminator, D(x, y) represents the discriminator's estimate of the probability that the real image is genuine, and D(x, G(x)) represents the probability that the fake image is real Indicates the discriminator's estimate for
is expressed as
The final target G* of the cGAN is the following relational expression

where λ represents the relative weight of L _cGAN loss and L ₁ loss
Expressed as, Total electron number prediction device. According to claim 1,
The learning unit performs learning on the prediction model by dividing the total number of electrons content data of the date belonging to the maximum period of the sun and the total number of electrons content data belonging to the minimum period,
A device for predicting the total number of electrons. According to claim 1,
The specific time is 24 hours,
A device for predicting the total number of electrons. A method in which a total electron number content predicting device predicting a total electron number content after a specific time using a current total electron number content image predicts a total electron number content,
Performing deep learning-based training on a predictive model that predicts a total electron number content image after a specific time from a specific point in time using a total electron number content image in the past to learn the predictive model;
inputting a current total electron number content image and a difference map representing a difference between a total electron number content image from the present to the specific time ago and the current total electron number content image into the prediction model; and
Acquiring an image generated by the predictive model and outputting it as a total electron number content image after the specific time from the present
Including,
In the learning step, among the total electron number content images of the past, the total electron number content image at the specific time point (hereinafter referred to as 'learning image'), the total electron number content image before the specific time from the specific time point, and the By performing conditional generative adversarial network (cGAN)-based learning with a difference map representing the difference between training images and an image of the total number of electrons (hereinafter referred to as 'actual image') after the specific time from the specific point in time as inputs, Creating a predictive model that predicts the total electron number content image after a specific time,
Method for predicting the total number of electrons. delete According to claim 9,
The prediction model is a conditional generative adversarial network (cGAN)-based prediction model, including a generator and a discriminator,
The generator receives the training image and the difference map as inputs and generates a fake image corresponding to the training image,
The discriminator compares the fake image generated by the generator with the real image to determine authenticity, and feeds back the determination result to the generator.
Method for predicting the total number of electrons. According to claim 11,
The learning step is repeated until the discrimination result fed back from the discriminator to the generator indicates that the fake image is real.
Method for predicting the total number of electrons.