JP7199075B2 - Forecasting systems and methods - Google Patents

Description

The present technology relates to a prediction system and prediction method for predicting phenomena occurring in outer space.

It is not easy to analyze and predict various phenomena that occur in outer space, partly due to the limited information that can be observed. As one of the various phenomena that occur in outer space, observation and prediction of solar activity, especially explosion phenomena that occur on the sun (generally referred to as "solar flares" or simply "flares") have hitherto been carried out. being continued. Large-scale solar flares produce high-energy particle emissions, flare X-ray radiation, and solar wind disturbances. As a result, social impacts such as satellite failures, satellite positioning failures, shortwave communication failures, aircraft route changes, and power transmission facility troubles can occur. Therefore, a system is in place to predict the occurrence of solar flares and prepare for unforeseen circumstances.

As a method for predicting the occurrence of such a solar flare, for example, in Non-Patent Document 1, an observation image of the sun is input to a CNN (Convolutional Neural Network), which is one of the neural networks, and the probability of occurrence of a solar flare is calculated. is disclosed for each class (scale).

Xin Huang, Huaning Wang, Long Xu, Jinfu Liu, Rong Li, and Xinghua Dai, "Deep Learning Based Solar Flare Forecasting Model. I. Results for Line-of-sight Magnetograms," The American Astronomical Society, The Astrophysical Journal, 856 :7(11pp), March 20, 2018 N. Nishizuka, K. Sugiura, Y. Kubo, M. Den, S. Watari, M. Ishii, "Solar Flare Prediction Model with Three Machine-learning Algorithms using Ultraviolet Brightening and Vector Magnetograms," The Astrophysical Journal, 835:156 (10pp), February 1, 2017

According to the method disclosed in Non-Patent Document 1, a prediction model is constructed by providing a CNN with labeled observed images. Even if a prediction model built by such an image-based method shows good prediction accuracy, what kind of phenomena appearing in observation images should be focused on in order to predict the occurrence of solar flares? It is difficult to obtain scientific knowledge. In other words, a prediction model that not only predicts the probability of solar flare occurrence, but also provides some kind of feedback for clarifying the mechanism of solar flare occurrence is desirable.

The purpose of this technology is to provide a prediction system using a prediction model that can be used to elucidate the mechanism of occurrence of phenomena occurring in outer space.

According to one aspect of the present invention, a prediction system for predicting phenomena occurring in outer space is provided. The prediction system includes an acquisition unit that acquires time-series observation data derived from outer space, a feature amount calculation unit that calculates a plurality of feature amounts based on data elements before the prediction timing in the observation data, and a feature amount an estimating unit that calculates the probability of occurrence of each of the one or more events within a predetermined length of prediction period following the prediction timing by applying the plurality of feature values calculated by the calculating unit to the prediction model. .

The plurality of feature amounts may include feature amounts calculated based on data elements observed within a predetermined period immediately before the prediction timing.

The estimation unit may calculate a first probability that the first event will occur within the prediction period and a second probability that the first event will not occur within the prediction period.

The estimation unit may calculate the probability of occurrence of each of the plurality of events, and output the event with the highest calculated probability as the prediction result.

The prediction model includes a plurality of linearly-connected layers connected in series, an input of a first linearly-connected layer among the plurality of linearly-connected layers, and an output of a second linearly-connected layer different from the first linearly-connected layer. may include a skip path connecting the .

The one or more events may include the probability that a solar flare of a given class or higher will occur.

The feature amount calculation unit includes means for detecting and temporally tracking the active region of the sun, means for acquiring a class of solar flare that occurred in the tracked active area, and a plurality of feature amounts for each predetermined calculation cycle. is calculated, and a label indicating the largest class of solar flare that occurred within a prediction period following the calculation timing of the plurality of feature values is assigned to the calculated plurality of feature values to generate a training sample; may be included. The prediction system may further include a learning unit that optimizes the network parameters defining the prediction model by supervised learning using training samples.

The prediction system may further include an influence calculator that calculates the influence of each feature on the prediction accuracy of the prediction model.

According to another aspect of the present invention, a prediction method for predicting phenomena occurring in outer space is provided. The prediction method includes a step of acquiring time-series observation data derived from outer space, a step of calculating a plurality of feature values based on data elements of the observation data prior to the prediction timing, and a step of calculating a plurality of feature values. and calculating the probability of each of the one or more events occurring within a predetermined length of prediction period following the prediction timing by providing the quantities to the prediction model.

According to yet another aspect of the invention, a trained model is provided for operating a computer to estimate the probability of a solar flare occurring. The trained model includes the steps of detecting and temporally tracking an active region of the sun, obtaining classes of solar flares that occurred in the tracked active region, and calculating a plurality of feature values for each predetermined calculation cycle. a step of generating a training sample by calculating and assigning to the calculated plurality of feature quantities a label indicating the largest class of solar flare that occurred within a prediction period following the calculation timing of the plurality of feature quantities; and optimizing the network parameters that define the prediction model by supervised learning using training samples.

According to this technology, it is possible to realize a prediction system using a prediction model that can be used to elucidate the mechanism of occurrence of phenomena that occur in outer space.

It is a mimetic diagram showing a schematic structure of a prediction system according to this embodiment. FIG. 4 is a diagram for explaining the temporal relationship for calculating the solar flare occurrence probability by the prediction system according to the present embodiment; FIG. 4 is a diagram showing an example of observed images and detected active regions used in the prediction system according to the present embodiment; FIG. 10 is a diagram for explaining teacher label assignment and training sample generation processing by the prediction system according to the present embodiment; 1 is a schematic diagram showing an example of a network of prediction models adopted by a prediction system according to the present embodiment; FIG. FIG. 6 is a schematic diagram showing an example of the neural network shown in FIG. 5; FIG. 1 is a schematic diagram showing an example of a hardware configuration for realizing a prediction device according to an embodiment; FIG. 4 is a flow chart showing a processing procedure relating to a learning phase of the prediction system according to the present embodiment; It is a flow chart which shows the processing procedure concerning the operation phase of the prediction system according to this embodiment. FIG. 4 is a diagram showing an output example of prediction results provided by the prediction system according to the present embodiment; FIG. 4 is a schematic diagram for explaining a procedure for evaluating the importance of feature quantities in the prediction system according to the present embodiment;

Embodiments of the present invention will be described in detail with reference to the drawings. The same or corresponding parts in the drawings are denoted by the same reference numerals, and the description thereof will not be repeated.

[A. Application example of prediction system]
Prediction system 1 according to the present embodiment predicts one or more phenomena occurring in outer space based on time-series observation data derived from outer space. Predicting one or more phenomena includes calculating the probability of occurrence of each phenomenon.

In the following description, as an example of such a phenomenon occurring in outer space, an application for predicting the probability of occurrence of a solar flare will be exemplified. More specifically, the one or more events predicted by the prediction system 1 include the probability that a solar flare of a predetermined class or higher will occur.

FIG. 1 is a schematic diagram showing a schematic configuration of a prediction system 1 according to this embodiment. Referring to FIG. 1, a prediction system 1 predicts in advance the probability of occurrence of a solar flare, which is an explosion phenomenon occurring on the sun 2 .

The prediction system 1 has a function of acquiring time-series observation data derived from outer space. More specifically, the prediction system 1 can use a ground observation facility 4 as a configuration for observing the sun 2 . The ground observation facility 4 includes an optical telescope 6 and/or a radio telescope 7 such as a radio heliograph, a solar radio intensity polarimeter, a millimeter wave radio telescope, a millimeter wave interferometer, and an observation device 8 .

The observation device 8 outputs time-series observation data based on signals observed by the optical telescope 6 and/or the radio telescope 7 . Observation data may be collected from multiple ground observation facilities 4 , or multiple optical telescopes 6 and/or radio telescopes 7 may be arranged at a single ground observation facility 4 .

The prediction system 1 may additionally have available observation satellites 10 and ground reception facilities 12 . The observation satellite 10 observes various electromagnetic waves and particles emitted from the sun 2 in outer space, and transmits the observation results to the ground reception facility 12 . Terrestrial reception facility 12 includes communication antenna 14 and communication device 16 . The communication device 16 performs signal processing on radio waves received by the communication antenna 14 and outputs observation data.

Observation data output from the observation device 8 and/or the communication device 16 is accumulated in the observation data collection device 18 . The observation data collection device 18 outputs the accumulated observation data upon request.

The function of acquiring time-series observation data derived from outer space in the prediction system 1 does not need to actually have observation facilities and observation devices, and time-series observation data can be obtained from available observation facilities and observation devices. including any means for obtaining (typically, such as downloading)

The prediction system 1 has a prediction device 20 including a feature quantity calculator 30 and an estimator 40 .

Based on the observation data from the observation data collection device 18, the feature amount calculator 30 calculates a plurality of different feature amounts used for calculating the probability of occurrence of solar flares. At the time of prediction, the feature amount calculation unit 30 calculates a plurality of feature amounts based on data elements of observation data that precede the prediction timing.

The estimator 40 has a prediction model using a DNN (Deep Neural Network) as will be described later. is input, the probability of occurrence of a solar flare is calculated. More specifically, the estimator 40 applies the plurality of feature quantities calculated by the feature quantity calculation unit 30 to the prediction model, so that one or more events occur within a prediction period of a predetermined length following the prediction timing. (solar flares of a predetermined class or higher) are calculated.

FIG. 2 is a diagram for explaining temporal relationships for calculating the probability of occurrence of solar flares by prediction system 1 according to the present embodiment. With reference to FIG. 2, assume that the estimation process is performed at an arbitrary time. The probability that a solar flare will occur during a predetermined period (prediction period) from the timing at which this estimation process is performed is calculated. Following the operation of existing solar flare observation, prediction system 1 according to the present embodiment calculates the occurrence probability for each scale (class) of solar flare.

For example, when X-ray magnitude is used as an index of the scale of a solar flare, a total of five classes, A, B, C, M, and X, are defined according to the maximum intensity of emitted X-rays. . In practice, the probability of occurrence of X-, M-, and C-class solar flares is important, and attention is often paid to M-class and C-class, which have relatively high occurrence frequencies.

The occurrence probabilities output by prediction system 1 according to the present embodiment may be, for example, the probability that a solar flare of class M or higher will occur and the probability that a solar flare of class C or higher will occur. Furthermore, as an index of the scale (class) of solar flares, the Hα magnitude or the R scale may be used instead of the X-ray magnitude.

In the estimation process, a plurality of feature quantities calculated from observation data immediately before the time at which the estimation process was performed are used. The feature amounts calculated by the feature amount calculation unit 30 (FIG. 1) include feature amounts calculated based on data elements observed within a predetermined period immediately before the prediction timing.

The observation period of the observation data used to calculate each feature amount is determined arbitrarily. In order to improve prediction accuracy, it is preferable to diversify the types of observation data used to calculate the feature amount, the content of the calculated feature amount, the period of the observation data used to calculate the feature amount, and the like.

[B. Feature value]
Unlike the prior art described above, the prediction model according to the present embodiment uses a plurality of feature values calculated from observation data to predict the probability of occurrence of a solar flare. First, a method for calculating the feature amount and the like will be described.

(b1: Overview)
Observation of solar flares has a long history, and observation facilities are in place around the world. The inventors of the present application are SDO (Solar Dynamics Observatory), a solar observation satellite managed by the National Aeronautics and Space Administration (NASA), and the National Oceanic and Atmospheric Administration (NOAA: National Oceanic and Atmospheric Administration) Observation data collected by GOES (Geostationary Operational Environmental Satellite), which is a meteorological satellite under management, was used.

Various observation data are collected, but in this embodiment, (1) line-of-sight magnetogram, (2) vector magnetogram, (3) 1600 Å filter (4) EUV hot coronal brightening image with 131 Å filtering; (5) the light curves of the soft X-ray emission), etc.

The line-of-sight magnetic field image is observed by an HMI (Helioseismic and Magnetic Imager) mounted on the SDO. Also, the UV continuum at the top of the photosphere is obtained as a result of 1600 Å filtering of the AIA (Atmospheric Imaging Assembly) mounted on the SDO. An EUV hot coronal emission image of the region in which the solar flare is occurring is obtained by 131 Å filtering results with SDO. A global integral value of X-ray emission over the range 1-8 Å is observed by GOES.

The observation period of the line-of-sight magnetic field image is 45 seconds, the observation period of the vector magnetic field image is 12 minutes, the observation period of both the 1600 Å filtered image and the 131 Å filtered image is 12 seconds, and the observation period of GOES is less than 1 minute. Considering the data size generated in such an observation period, the prediction system 1 according to the present embodiment is configured to be able to calculate and predict the feature amount in one hour period.

In the prediction system 1, active regions (hereinafter also referred to as “AR” or “ARs”) are detected from full-disk images of the line-of-sight magnetic field, and each active region is tracked along.

For each active region, a feature value is calculated from the observation results of multiple wavelengths, and a solar flare of any class X, M, or C has occurred within 24 hours from the time the target image was observed. In this case, the calculated feature amount is assigned a label indicating the class of the solar flare that occurred and stored in the database.

DNN-based supervised learning is performed using the feature quantities labeled in this way as training samples. That is, the prediction model used in prediction system 1 according to the present embodiment corresponds to a trained model generated by supervised learning.

In supervised learning in the present embodiment, the task of predicting the largest class of solar flares that can occur within the next 24 hours is learned every hour. Due to the observation cycle of the observation data used this time, the feature amount is calculated every hour, and the cycle of supervised learning is determined accordingly.

In each training sample, if the corresponding observation data is missing, the past observation data that exists within the last 30 minutes is searched, and if it is also missing, the learning of the cycle Processing is skipped.

The inventors of the present application used observation data obtained from June 2010 to December 2015 by SDO (operation started in February 2010). During this period, 26 X-class solar flares, 383 M-class solar flares, and 4054 C-class solar flares were observed. These observed solar flares on the sun's surface represent about 90% of all solar flares that occurred during that time period.

(b2: detection of active area)
First, an active area (AR) for calculating a feature amount is detected from an acquired observed image. That is, the feature amount calculation unit 30 detects and temporally tracks the active region of the sun. Typically, active regions may be detected from a full-view image of the line-of-sight magnetic field. The line-of-sight magnetic field image has less noise than the vector magnetic field image, and the active region detection process can be performed more accurately.

Detection of active regions can be realized, for example, by searching and extracting regions above a predetermined threshold value (eg, 140 G (Gauss)) in each line-of-sight magnetic field image. The coordinate positions in the frame of the detected active regions are then applied to other images (ie, images observed at different wavelengths). An active region detected on the outer edge of a celestial body in a photospheric image may be regarded as invalid and excluded from the feature amount calculation targets. A unique identification number may be assigned to track the detected active area. See Non-Patent Document 2 and the like for a more specific method of detecting an active region.

FIG. 3 is a diagram showing an example of observed images and detected active regions used in prediction system 1 according to the present embodiment. FIG. 3(a) shows an example of a full image of white light intensity while light intensity observed by the HMI mounted on the SDO. In FIG. 3(a), four active regions are detected. FIG. 3(b) shows an example of a full image of the line-of-sight magnetic field observed by the HMI mounted on the SDO. FIG. 3(c) shows an example of a 1600 Å filtered full-surface UV continuum image observed by the AIA onboard the SDO. FIG. 3(d) shows an example of a 131 Å filtered full-surface image of the hot coronal emission observed by the AIA onboard the SDO.

In addition, in the active area AR1 set in each image of FIGS. A solar flare is occurring.

(b3: Feature value calculation for active region)
In the prediction system 1 according to the present embodiment, the required feature amount is calculated.

In addition, a 131 Å filtered EUV high temperature coronal emission image (SDO) showing line emission of 20-valent iron ions (Fe _XX ) and 23-valent iron ions (Fe _XXIII ) emitted at temperatures above 10 ⁷ K. (observed by the AIA installed in the )) may be used to calculate the feature amount.

Assuming a practical prediction, the EUV high-temperature coronal emission image with 131 Å filter processing and the image 1 hour and 2 hours before the X-ray data collected by GOES may also be adopted as feature quantities. good.

Also, the maximum intensity in each active region in the EUV high-temperature corona emission image subjected to 131 Å filter processing may be employed as a feature amount.

More specifically, the 79-dimensional (79 types) feature amount attached as an appendix at the end of this document is used. Each feature amount is calculated for each detected active area (AR).

In prediction system 1 according to the present embodiment, a feature quantity vector having each feature quantity as an element is generated and input to a prediction model, thereby estimating the occurrence probability of a solar flare in a target active region.

It should be noted that it is not always necessary to generate a feature amount vector including all of the feature amounts described above, and a part thereof and its time series may be used.

(b4: Labeling of features and generation of training samples)
Next, the assignment of teacher labels to the feature values (feature value vectors) calculated at predetermined intervals and the generation of training samples will be described.

FIG. 4 is a diagram for explaining teacher label assignment and training sample generation processing by prediction system 1 according to the present embodiment. Referring to FIG. 4, a plurality of feature amounts as described above are calculated at predetermined calculation intervals (for example, one hour). The largest class among the solar flares occurring within a predetermined period (for example, 24 hours) (corresponding to the prediction period) following the timing of calculating the feature amount is set as the teacher label of the feature amount.

Therefore, for example, when the feature amount group (feature amount vector) is calculated at time _tn , the largest class among the solar flares occurring within the prediction period following time _tn is determined as the teacher label.

As a result, {AR _i (t _{n )} : teacher label|feature ₁ , feature 2, feature 3, . . For each active region, training samples are generated sequentially at predetermined intervals over a trackable time period. In this way, the feature quantity calculator 30 acquires the class of solar flare that occurred in the tracked active region. Then, the feature amount calculation unit 30 calculates a plurality of feature amounts for each predetermined calculation cycle, and adds the calculated feature amounts to the sun that occurred within the prediction period following the calculation timing of the plurality of feature amounts. Generate training samples labeled with the largest class of flares.

These training samples will be used to learn a predictive model. That is, prediction system 1 according to the present embodiment optimizes network parameters that define a prediction model by supervised learning using training samples.

[C. prediction model]
Next, prediction model 400 adopted by prediction system 1 according to the present embodiment will be described.

FIG. 5 is a schematic diagram showing an example of a network of prediction models 400 adopted by prediction system 1 according to the present embodiment. The prediction model 400 shown in FIG. 5 is a DNN-classified network and is implemented within the estimator 40 . The prediction model 400 has a plurality of layers, and a feature amount vector calculated from observation data (in this embodiment, a 79-dimensional normalized feature amount) is given as an input, and each class , the probability of occurrence of a solar flare is output.

FIG. 5 shows a map from inputs to outputs in predictive model 400 . The predictive model 400 includes multiple layers of neural networks, including linear combinations and activation functions (usually non-linear). More specifically, predictive model 400 includes linear combination layers 411 , 412 , 413 , 414 , 415 , 416 and nonlinear combination layer 417 .

FIG. 6 is a schematic diagram showing an example of the neural network shown in FIG. FIG. 6(a) shows an example of a network of linear combination layers 411, 412, 413, 414, 415, and 416, and FIG. 6(b) shows an example of a network of nonlinear combination layer 417. FIG.

FIG. 6A shows a configuration example in which one hidden layer 443 is arranged between an input layer 442 and an output layer 444 as the simplest network. An input vector 441 representing a feature amount is given to the input layer 442, and an output vector 445, which is the result of linear combination, is output from the output layer 444. FIG.

Basically, each node included in the input layer 442 and each node included in the hidden layer 443 are connected to each other. A weighting factor vector 446, which is a set of weighting factors given to the connections between each node of the input layer 442 and each node of the hidden layer 443, is determined by learning. Similarly, each node contained in hidden layer 443 and each node contained in output layer 444 are interconnected. A weighting factor vector 447, which is a set of weighting factors given to the connection between each node of the hidden layer 443 and each node of the output layer 444, is determined by learning.

A bias vector 448, which is a set of biases applied to each node of the hidden layer 443, and a bias vector 449, which is a set of biases applied to each node of the output layer 444, are set as necessary.

By using a network as shown in FIG. 6(a), the input vector 441 to the output vector 445 becomes a first-order linear combination defined by the weighting coefficient vectors 446, 447 and the bias vectors 448, 449. . That is, the network shown in FIG. 6(a) is a set for linearly transforming the input data group, and between the input vector x and the output vector y, the weighting factor vector W (the weighting factor vector shown in FIG. 6(a) W1 and W2 combined) and a bias vector b (combined bias vectors b1 and b2 shown in FIG. 6A) can be expressed in the form of the following affine transformation.

y=f(x;W,b)=f(Wx+b)
In addition to the network shown in FIG. 6(a), FIG. shows a network with additional additions. The ReLU function 451 transfers the feature quantity input to each node to the subsequent stage according to a nonlinear activation function. More specifically, the ReLU function 451 outputs the following values for the input vector x.

ReLU(x)=log(1+exp(x))≈max(x,0)
The Softmax function 452 normalizes the probabilities for the feature vectors output from the ReLU function 451 to estimate probabilities for each class.

The networks shown in FIGS. 6A and 6B are examples, and in each of the linear coupling layers 411, 412, 413, 414, 415, and 416 and the nonlinear coupling layer 417, the number of input layer nodes, The number of nodes in the output layer, the number of layers in the hidden layer, and the like may be determined as appropriate. Also, the activation function is not limited to the ReLU function and the Softmax function, and any activation function can be used.

Referring to FIG. 5 again, the prediction model 400 outputs the solar flare occurrence probability p(y). In this embodiment, the estimator 40 uses the prediction model 400 to obtain a first probability that the first event will occur within the prediction period and a second probability that the first event will not occur within the prediction period. At least two kinds of probabilities are calculated.

As an example, attention is paid to the occurrence probability of solar flares of M class or higher (or C class or higher). Here, y is a vector indicating the solar flare class and is defined as y=(y ₁ , y ₂ ). When (y ₁ , y ₂ )=(0, 1), it indicates an event that a solar flare of M class or higher (or C class or higher) occurs, and (y ₁ , y ₂ )=(1, 0) ) indicates an event that a solar flare less than M-class (or less than C-class) occurs or that a solar flare does not occur.

In the prediction performance evaluation described later, the probability of occurrence of an event that a solar flare of M class or higher (or C class or higher) will occur and a solar flare of less than M class (or C class or higher) will occur, or We estimated the occurrence probability of an event that a solar flare would not occur, and adopted the larger one of the estimated occurrence probabilities as the final result.

In the prediction model 400, batch normalization processing (Batch Normalization: also denoted as "BN") 421, 422, 423, 424, 425, 426 is performed between each layer. Batch normalization processing stabilizes learning and increases prediction accuracy by normalizing the input parameters of each layer.

_{The output vector y={y 1 , y 2 , .} (1B) can be expressed as in formula.

where μ _B denotes the mini-batch mean and σ _B denotes the mini-batch variance. A mini-batch corresponds to the number of training samples in one forward/backward pass. The mini-batches are actually normalized by the batch normalization process. γ and β in the equations are optimized in a similar manner as the weighting factors. ε indicates an arbitrary value to prevent division by zero from occurring, and the subscript i indicates the hidden layer node index.

Furthermore, in the predictive model 400 shown in FIG. 5, relatively simple skip connections are introduced to increase the number of layers. That is, the prediction model 400 includes a plurality of linearly-connected layers connected in series, an input of a first linearly-connected layer among the plurality of linearly-connected layers, and a second linearly-connected layer different from the first linearly-connected layer. and skip paths that combine the outputs of This skip connection increases the prediction accuracy of the prediction model and functions to prevent the disappearance or divergence of the gradient due to the increase in the number of layers.

More specifically, predictive model 400 has two skip connections. A first skip connection consists of a skip path 431 leading from the input side of the linear combination layer 411 to the output side of the linear combination layer 412 and an adder 432 for coupling the skip path 431 to the output side of the linear combination layer 412. Become. A second skip connection is from a skip path 433 leading from the input side of linear combination layer 413 to the output side of linear combination layer 415 and an adder 434 for coupling skip path 433 to the output side of linear combination layer 415. Become.

In general, model accuracy decreases when the number of layers increases too much. To solve this problem, algorithms have been proposed to optimize network parameters by learning the residual function of each layer using skip connections. By adopting such skip connections, it is possible to improve model accuracy and stabilize learning even when the number of layers increases.

The estimator 40 calculates the probability of occurrence of each of a plurality of events, and outputs the event with the highest calculated probability as a prediction result. That is, the maximum estimated occurrence probability p(y _k ) for each of the k labels y _k ^* (the label with the maximum occurrence probability) is output as the prediction result. Specifically, the prediction result is determined in the form shown in the following equation (2).

[D. Learning prediction model]
Next, learning processing for constructing the prediction model 400 shown in FIG. 5 will be described. To maximize the prediction accuracy, in learning the prediction model 400 shown in FIG. 5, the network parameters are optimized to minimize the cost function J as shown below.

In general identification problems, network parameters are optimized to minimize cross-entropy. However, in the event handled by prediction model 400 according to the present embodiment, the probabilities of occurrence of solar flares are not balanced. We are hiring.

In equation (3), _wk is a weighting factor for each class k, and the reciprocal of the occurrence probability of each class is adopted. y _nk ^* denotes the teacher label of the nth training sample in class k, and p(y _nk ) denotes the estimated probability of occurrence for the nth training sample in class k.

In prediction system 1 according to the present embodiment, prediction model 400 is set according to the parameters shown in Table 1 below, and network parameters defining prediction model 400 are optimized using an arbitrary optimization method. The number of nodes in each layer of the prediction model 400 and the batch size in the batch normalization process may be set to values in the range of 50-200. Also, the number of linear coupling layers is preferably about 4 to 9, and if necessary, skip connections may be appropriately added.

As an optimization method, for example, the Adam (Adaptive moment estimation) method can be used.

[E. Hardware configuration of prediction device 20]
Next, an example of the hardware configuration of prediction device 20 according to the present embodiment will be described. FIG. 7 is a schematic diagram showing an example of a hardware configuration for realizing prediction device 20 according to the present embodiment. Predictor 20 is typically implemented using a general-purpose computer.

7, the prediction device 20 includes, as main hardware components, a processor 100, a main memory 102, a display 104, an input device 106, a network interface (I/F) 108, an optical It includes a drive 110 , an input interface (I/F) 114 , an output interface (I/F) 116 and a secondary storage device 120 . These components are connected to each other via internal bus 118 .

Processor 100 is a computing entity that executes processing necessary for realizing prediction system 1 according to the present embodiment by executing various programs as described later. For example, one or more CPUs (Central Processing Unit) and a GPU (Graphics Processing Unit). Any CPU or GPU with multiple cores may be used.

The main memory 102 is a storage area that temporarily stores program code, work memory, etc. when the processor 100 executes a program. It consists of a volatile memory device, etc.

The display 104 is a display unit that outputs a user interface related to processing, processing results, and the like, and is configured by, for example, an LCD (Liquid Crystal Display) or an organic EL (Electroluminescence) display.

The input device 106 is a device that receives commands, operations, and the like from the user, and includes, for example, a keyboard, mouse, touch panel, pen, and the like. Also, the input device 106 may include a microphone for collecting sounds necessary for machine learning, or an interface for connecting to a sound collecting device that collects sounds necessary for machine learning. good too.

A network interface 108 exchanges data with any information processing device or the like on the Internet or intranet. As the network interface 108, for example, any communication method such as Ethernet (registered trademark), wireless LAN (Local Area Network), Bluetooth (registered trademark), or the like can be adopted.

The optical drive 110 reads information stored in an optical disc 112 such as a CD-ROM (Compact Disc Read Only Memory) or a DVD (Digital Versatile Disc) and outputs it to other components via an internal bus 118 . The optical disk 112 is an example of a non-transitory recording medium, and is distributed in a state in which arbitrary programs are stored in a non-volatile manner. The optical drive 110 reads the program from the optical disc 112 and installs it in the secondary storage device 120 or the like, so that the function of the prediction system 1 can be provided by the computer. Therefore, the subject of the present invention can be a program itself installed in secondary storage device 120 or the like, or a recording medium such as optical disc 112 storing a program for realizing functions and processes according to the present embodiment. .

FIG. 7 shows an optical recording medium such as an optical disc 112 as an example of a non-transitory recording medium, but is not limited to this, semiconductor recording media such as flash memory, magnetic recording media such as hard disks or storage tapes. , and MO (Magneto-Optical disk) may be used.

The input interface 114 is connected to an external observation device or the like, and takes in necessary observation data or the like. The output interface 116 is connected to a display device or the like, and outputs prediction results (estimation results) such as the probability of occurrence of solar flares. The input interface 114 and the output interface 116 can use general-purpose communication interfaces such as USB (Universal Serial Bus).

The secondary storage device 120 is a component that stores a program executed by the processor 100, a training data set for learning a neural network as described later, and network parameters that define the neural network. It consists of a non-volatile storage device such as a hard disk or SSD (Solid State Drive).

More specifically, the secondary storage device 120 includes an OS (Operating System) (not shown), a feature amount calculation program 122 for realizing the feature amount calculation unit 30, and a learning program 124 for constructing a prediction model. , an estimation program 126 for calculating the probability of occurrence of a solar flare using the prediction model, and network parameters 130 that define the neural network that is the substance of the prediction model. A training data set 132 may also be stored in the secondary storage device 120 .

Some of the libraries and functional modules required for executing these programs on the processor 100 may be replaced with libraries or functional modules provided as standard by the OS. In this case, each program alone does not include all the program modules necessary to implement the corresponding function, but by installing it under the execution environment of the OS, the necessary function can be implemented. . Even a program that does not include some of such libraries or functional modules can be included in the technical scope of the present invention.

Moreover, these programs may be distributed by being stored in any recording medium as described above and not only being distributed, but also being downloaded from a server device or the like via the Internet or an intranet.

FIG. 7 shows an example in which a single computer configures the prediction system 1, but not limited to this, a plurality of computers connected via a computer network explicitly or implicitly cooperate to form a prediction system 1 may be implemented. When a plurality of computers work together, some of the computers may be unspecified computers on the network, so-called cloud computers.

All or part of the functions realized by the computer (processor 100) executing the program may be realized using a hard-wired circuit such as an integrated circuit. For example, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or the like may be used.

A person skilled in the art will be able to realize the prediction system according to the present embodiment by appropriately using the technology suitable for the era in which the present invention is implemented.

[F. Processing procedure]
Next, an example of a processing procedure in prediction system 1 according to the present embodiment will be described.

(f1: learning phase)
First, a processing procedure relating to a learning phase for constructing prediction model 400 in prediction system 1 according to the present embodiment will be described.

FIG. 8 is a flow chart showing a processing procedure relating to the learning phase of prediction system 1 according to the present embodiment. Each step shown in FIG. 8 is typically implemented by the processor 100 of the prediction device 20 executing the feature amount calculation program 122 and the learning program 124 (see FIG. 7 for both).

Referring to FIG. 8, first, prediction device 20 collects observation data necessary for learning prediction model 400 (step S100).

Predictor 20 then performs active region detection and tracking. More specifically, the prediction device 20 detects an active region from the entire image of the line-of-sight magnetic field at each time (step S102) among the collected observation data, and detects an active region that is positionally and temporally continuous. are tracked as the same active area (step S104). Then, the prediction device 20 determines the coordinate position at each time for the time series of the tracked active region (step S106).

Subsequently, training sample generation processing is executed. That is, the prediction device 20 selects the time series of one active region from the tracked active regions (step S108), and determines the occurrence time and occurrence time of the solar flare that occurred during tracking of the selected active region. A class is acquired (step S110). Furthermore, the prediction device 20 calculates a plurality of predetermined feature amounts for each predetermined calculation cycle for the tracking period of the selected active region to generate a feature amount vector (step S112). Acquire a teacher label based on the class of solar flare that occurred in the prediction period following the reference timing for generating the quantity vector (step S114), assign the acquired teacher label to the generated feature quantity vector, and generate a training sample. Generate (step S116).

Prediction device 20 determines whether generation of training samples has been completed for all tracking periods of the selected active region (step S118). If generation of training samples has not been completed for all tracking periods of the selected active region (NO in step S118), steps S112 to S116 are repeated.

On the other hand, if training samples have been generated for all of the selected active region tracking periods (YES in step S118), prediction device 20 determines whether all of the tracked active regions have been processed. It is determined whether or not (step S122). If there are unprocessed active areas among the tracked active areas (NO in step S122), prediction device 20 selects one of the unprocessed active areas (step S124), and performs steps S110 and after. process again.

On the other hand, if all the tracked active regions have been processed (YES in step S122), the training sample generation process is complete.

Subsequently, the prediction device 20 executes processing for constructing the prediction model 400 . That is, the prediction device 20 initializes (randomly determines) each network parameter of the target prediction model 400 (step S126), and sets the feature amount vector included in each of the plurality of training samples generated by the above-described processing to Input to the prediction model 400 (step S128), compare the estimation result output from the prediction model 400 and the teacher label assigned to the input training sample, and update the network parameters that define the prediction model 400 (Step S130).

Prediction device 20 determines whether or not the termination condition of the learning process for prediction model 400 is satisfied (step S132). Repeat process. Here, the conditions for terminating the learning process are the completion of learning using all the generated training samples and/or the amount of update of the network parameters when new training samples are input, which is a predetermined threshold value. is less than, and the like.

If the termination condition of the learning process for prediction model 400 is satisfied (YES in step S132), prediction device 20 outputs the network parameters defining current prediction model 400 as the learning result (step S134). Then the process ends.

(f2: operation phase)
Next, a processing procedure relating to the operation phase for calculating the probability of occurrence of solar flare using prediction system 1 according to the present embodiment will be described.

FIG. 9 is a flow chart showing a processing procedure relating to the operation phase of prediction system 1 according to the present embodiment. Each step shown in FIG. 9 is typically implemented by the processor 100 of the prediction device 20 executing the feature quantity calculation program 122 and the estimation program 126 (see FIG. 7 for both).

With reference to FIG. 9, prediction device 20 determines whether or not a predetermined prediction timing has arrived (step S200). If the predetermined prediction timing has not arrived (NO in step S200), the process of step S200 is repeated.

If the predetermined prediction timing has arrived (YES in step S200), prediction device 20 determines whether or not one or more active regions can be detected (step S202). If no active region can be detected (NO in step S202), prediction device 20 outputs as a result that there is no active region to be predicted (step S204). Then the process ends.

On the other hand, if one or more active regions can be detected (YES in step S202), prediction device 20 selects one of the detected active regions (step S206). Then, the prediction device 20 calculates a plurality of predetermined feature amounts and generates feature amount vectors (step S208). Subsequently, the prediction device 20 inputs the generated feature quantity vector to the prediction model 400, and calculates the occurrence probability for each label (step S210). Then, the prediction device 20 determines the maximum occurrence probability among the calculated occurrence probabilities for each label as the prediction result of the selected active region (step S212).

Subsequently, the prediction device 20 determines whether all detected active regions have been processed (step S214). If there is an unprocessed active area among the detected active areas (NO in step S214), prediction device 20 selects one of the unprocessed active areas (step S216), and performs steps S208 and after. process again.

On the other hand, if all of the detected active regions have been processed (YES in step S214), prediction device 20 outputs the solar flare occurrence probability for each active region (step S218), and the process ends.

(f3: other forms)
For convenience of explanation, an example is shown in which the prediction device 20 executes both construction of the prediction model 400 (learning phase) and calculation of the probability of occurrence of solar flare using the prediction model 400 (operation phase), but this is not the only case. Alternatively, the entity that constructs the prediction model 400 and the entity that executes the prediction using the prediction model 400 may be separate.

For example, an embodiment is also envisioned in which a specific institution constructs the prediction model 400 and provides the constructed prediction model 400 to other institutions. In such cases, the essential part of the present invention is the constructed prediction model 400 itself.

[G. Output form of prediction result]
The prediction results output by prediction system 1 according to the present embodiment may be used in any form. For example, in current practice, experts predict the probability of occurrence of solar flares based on observational data, and may be used as an indicator of such predictions.

In such a case, a user interface may be provided so that the probability of occurrence of solar flares for each active region can be grasped at a glance.

FIG. 10 is a diagram showing an output example of prediction results provided by prediction system 1 according to the present embodiment. In FIG. 10, the occurrence probability of each class of solar flare is graphically displayed for each active region detected from the line-of-sight magnetic field image. More specifically, for each active region, the probability that a solar flare of class M or higher will occur (≧M), the probability that a solar flare of class M or higher will not occur (<M), and the solar flare of class C or higher will occur. A bar graph is displayed showing the probability (>C) and the probability (<C) that a solar flare of class C or greater will not occur.

By visually displaying the calculated probability of occurrence on the actually observed image in this way, it is possible to grasp at a glance which active region has the highest probability of occurrence of a solar flare.

[H. Results of prediction performance evaluation]
Next, results of prediction performance evaluation based on prediction results by prediction system 1 according to the present embodiment will be described. In the prediction performance evaluation below, we estimated the probability of a solar flare occurring within 24 hours, as in the existing operation.

In the prediction performance evaluation below, the observation data from 2010 to 2015 were divided into two groups. Specifically, observation data from 2010 to 2014 was used as the group used for training and validation, and observation data for 2015 was used as the group used for testing (ie predictive performance evaluation).

By separating the data into two groups along the time series in this way, it is possible to more accurately evaluate the performance for the prediction task of predicting future (unknown) phenomena based on past observation data.

A large number of iterations of the optimization process were performed on the predictive model 400 using mini-batches of the training data set in order to compute the predictive performance evaluations presented below. The mini-batches corresponding to training samples corresponded to the number of samples in one forward/backward pass of predictive model 400 . For mini-batches, we randomly selected from all training datasets to avoid duplication. In learning, updating of the network parameters of the predictive model 400 was performed for each epoch, ie for each forward/backward path. We also used mini-batches to speed up convergence.

In the prediction performance evaluation, two categories of prediction were performed. Specifically, (1) either an event that a solar flare of M class or more occurs, or an event that a solar flare of less than M class occurs, or no solar flare occurs, and (2 ) Either a C-class or higher solar flare will occur, or a less than C-class solar flare will occur, or no solar flare will occur. Since the number of occurrences of X-class solar flares was insufficient in the target observation data, predictions are made to include them in the M-class rather than the X-class alone.

In the prediction performance evaluation, the constructed prediction model 400 was evaluated for each epoch using test data. Based on these evaluations, the prediction model 400 with the highest TSS (true skill score) as the skill score was selected as the best model.

Evaluation results of predictive model 400 according to the present embodiment are shown in Tables 2-1 and 2-2 below.

Tables 2-1 and 2-2 show four types of outcome events: TP (True Positive), FP (False Positive), TN (True Negative), and FN (False Negative).

TP indicates the number of events in which solar flares have occurred even in the actual observation results in response to predictions that solar flares will occur. FP indicates the number of events in which solar flares did not occur in the actual observation results in contrast to predictions that solar flares would occur. TN indicates the number of events in which a solar flare did not occur even in the actual observation results for the prediction that no solar flare would occur. FN indicates the number of events in which solar flares occurred in the actual observation results, as opposed to predictions that no solar flares would occur.

Note that the number of TN events is relatively large because solar flares do not occur in many prediction cycles. In addition, since the occurrence of a solar flare of C class or higher includes the occurrence of a solar flare of M class or higher, the TP for C class or higher is relatively large.

In the task of predicting the occurrence of solar flares, the following six types of typical skill scores are known.

POD (probability of detection) means recall and can be calculated according to equation (4A). CSI (criteria success of index) can be calculated according to equation (4B). FAR (false alarm ratio) can be calculated according to equation (4C). HSS (Heidke skill score) can be calculated according to formula (4D). TSS (true skill score) can be calculated according to equation (4E). Accuracy, which means accuracy, can be calculated according to formula (4F). Note that (') in the formula means inversion (NOT).

Table 3 below shows the results of evaluating the prediction results of the prediction model 400 according to the present embodiment with respect to each skill score.

Table 3 above also shows the skill scores of the SVM, kNN, and ERT methods disclosed in Non-Patent Document 1 as baselines.

As shown in Table 3, according to the prediction model 400 according to the present embodiment, under the same conditions as in general operation, the TSS for the probability of occurrence of M-class or higher solar flares is "0.80", and C The TSS for the probability of occurrence of a solar flare of class or higher is "0.63". These results are superior to the TSS for probability of occurrence based on expert predictions.

[I. Importance evaluation]
Next, prediction model 400 used in prediction system 1 according to the present embodiment employs a configuration that predicts the probability of occurrence of a solar flare based on a plurality of feature quantities. can also be evaluated.

FIG. 11 is a schematic diagram for explaining the procedure for evaluating the importance of feature amounts in prediction system 1 according to the present embodiment. Referring to FIG. 11, first, a prediction model 400 is constructed by learning processing using a feature amount vector 460 including all available feature amounts. This constructed prediction model 400 is used as a reference prediction model, and a skill score is calculated for a prediction result obtained when test data 470 is given to this reference prediction model. This calculated skill score is the reference skill score.

Subsequently, for example, a feature amount vector 461 is generated by deleting the feature amount 1 among all available feature amounts, and the prediction model 401 is constructed by learning processing using the feature amount vector 461 . A prediction model 401 is a learning result in which the feature quantity 1 is deleted. Furthermore, a skill score is also calculated for a prediction result obtained when test data 471 (the component corresponding to feature value 1 is deleted) is given to the constructed prediction model 401 . This calculated skill score is the performance when the feature amount 1 is deleted.

Finally, the difference between the reference skill score and the skill score calculated when the feature 1 is deleted (that is, the degree of change in the skill score) is calculated as the influence of the deleted feature 1 on the skill score (that is, importance).

According to the same method as in FIG. 11, the degree of influence (importance) on the skill score for each feature amount can be calculated. Finally, by ranking the features based on their importance, we can estimate what kind of features are important for predicting the occurrence of solar flares. It can be a clue to clarify the generation mechanism of

As described above, prediction system 1 according to the present embodiment has an influence degree calculation function for calculating the degree of influence of each feature amount on the prediction accuracy of prediction model 400, and according to the processing procedure shown in FIG. , a numerical value that can objectively evaluate the importance of each feature quantity may be output.

[J. summary]
According to the prediction system 1 according to the present embodiment, a prediction model that uses, as inputs, a plurality of feature amounts obtained by capturing solar activity from various viewpoints is adopted instead of an image-based prediction model. By using multiple feature values, the number of dimensions of information can be reduced compared to using images as they are, so it is possible to construct a fully connected network, and prediction with high prediction accuracy with a relatively small number of layers. model can be realized.

Further, according to the prediction system 1 according to the present embodiment, since a prediction model with a plurality of feature values as input is adopted, based on the prediction accuracy of the prediction model constructed as a result of learning, each feature value for prediction accuracy can calculate the importance of As a result, we can obtain knowledge about what kind of feature values affect predictions, and can elucidate the mechanism by which solar flares occur.

The embodiments disclosed this time should be considered as examples and not restrictive in all respects. The scope of the present invention is indicated by the scope of the claims rather than the description of the above-described embodiments, and is intended to include all modifications within the scope and meaning equivalent to the scope of the claims.

[K. appendix]

1 Prediction System 2 Sun 4 Ground Observation Facility 6 Optical Telescope 7 Radio Telescope 8 Observation Device 10 Observation Satellite 12 Ground Reception Facility 14 Communication Antenna 16 Communication Device 18 Observation Data Collection Device 20 Prediction device, 30 feature quantity calculator, 40 estimator, 100 processor, 102 main memory, 104 display, 106 input device, 108 network interface, 110 optical drive, 112 optical disc, 114 input interface, 116 output interface, 118 internal bus , 120 secondary storage device, 122 feature amount calculation program, 124 learning program, 126 estimation program, 130 network parameters, 132 training data set, 400, 401 prediction model, 411, 412, 413, 414, 415, 416 linear combination layer , 417 nonlinear connection layer, 431, 433 skip path, 432, 434 adder, 441 input vector, 442 input layer, 443 hidden layer, 444 output layer, 445 output vector, 446, 447 weight coefficient vector, 448, 449 bias vector , 451 normalized linear function (ReLU function), 452 Softmax function, 460, 461 feature vector, 470, 471 test data, AR active region, J cost function.

Claims (6)

A prediction system for predicting phenomena occurring in outer space,
an acquisition unit that acquires time-series observation data derived from outer space;
A feature amount calculation unit that calculates a plurality of feature amounts based on data elements of the observation data before the prediction timing;
an estimating unit that calculates the probability that a solar flare will occur within a prediction period of a predetermined length following the prediction timing by giving a plurality of feature values calculated by the feature value calculation unit to a prediction model ;
The observation data includes data from an observation satellite that observes the sun in outer space,
The feature amount calculation unit calculates the plurality of feature amounts based on respective data elements in a plurality of mutually different periods immediately before the prediction timing,
The plurality of feature quantities are
a maximum intensity in X-ray data observed within a predetermined period immediately before the prediction timing;
and a maximum intensity in an EUV hot corona luminescence image or X-ray data observed a predetermined time before the prediction timing . 2. The prediction system according to claim 1, wherein said plurality of feature quantities further include histories of occurrence of solar flares according to a plurality of classes within a predetermined period immediately before said prediction timing. 3. The estimating unit according to claim 1, wherein the estimating unit calculates a probability of occurrence for each class of solar flare, and outputs a class corresponding to a maximum probability among the calculated probabilities as a prediction result. prediction system. The prediction system according to any one of claims 1 to 3, wherein the estimator calculates a probability that a solar flare of a predetermined class or higher will occur. 5. The prediction system according to any one of claims 1 to 4, further comprising an influence calculator that calculates the influence of each feature quantity on the prediction accuracy of the prediction model. A prediction method for predicting phenomena occurring in outer space,
a step of acquiring time-series observation data derived from the outer space;
a step of calculating a plurality of feature amounts based on data elements of the observation data before the prediction timing;
and calculating the probability that a solar flare will occur within a prediction period of a predetermined length following the prediction timing by applying the calculated plurality of feature values to a prediction model ;
The observation data includes data from an observation satellite that observes the sun in outer space,
The step of calculating the feature amount includes calculating the plurality of feature amounts based on each data element in a plurality of mutually different periods immediately before the prediction timing,
The plurality of feature quantities are
a maximum intensity in X-ray data observed within a predetermined period immediately before the prediction timing;
and a maximum intensity in an EUV hot corona luminescence image or X-ray data observed a predetermined time before the prediction timing .

Images