JP2020066365A - Prediction device, prediction method, and prediction program - Google Patents

Abstract

To provide a prediction device, a prediction method, and a prediction program capable of accurately predicting an orbit of a moving body.SOLUTION: A prediction device includes: an acquisition unit for acquiring environment information indicating an environment of a space where an object moving body moves; and a prediction unit for predicting a future orbit of the object moving body based on an output result of a model acquired by inputting the environment information acquired by the acquisition unit to the model learned so as to output orbit information indicating a future orbit of a certain moving body when the environment information indicating the environment of the space where the certain moving body has moved is input.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a prediction device, a prediction method, and a prediction program.

  As the use of outer space is expanding, it is becoming indispensable to accurately predict the orbit of a moving body (hereinafter referred to as a spacecraft) that moves in outer space such as satellites and space debris. is there. In order to realize high-precision orbit prediction, it is essential to elucidate the physical phenomena that affect the transition of the orbit. However, since there is little observation information in the altitude range where spacecraft are active, orbit prediction is performed through a model constructed from a small amount of observations. Under such circumstances, techniques aimed at improving the accuracy of orbit prediction include, for example, predicting the spacecraft orbit by learning the deviation between the prediction of the tracking antenna angle provided on the ground and the actual measurement. There are known techniques for predicting the orbit of a spacecraft based on observation information detected by a sensor mounted on the spacecraft (see, for example, Patent Documents 1 and 2).

JP, 2016-223781, A JP, 2009-220622, A

  In order to predict the orbit of a spacecraft with high accuracy, it is necessary to know at least several pieces of information on the cross-sectional area and mass ratio of the spacecraft in the flight direction, the resistance coefficient of the spacecraft, and the atmospheric density of the surrounding atmosphere. However, it is difficult to know the above information with high accuracy when the spacecraft is orbiting. Therefore, in the conventional technology, the mass ratio is assumed to be a constant value, the resistance coefficient is artificially given, and the atmospheric density uses the value obtained from the model, even though the cross-sectional area changes due to the solar paddle. . However, in the conventional technology, since uncertain parameters such as the cross-sectional area and mass ratio of the spacecraft, the resistance coefficient of the spacecraft, and the atmospheric density of the surrounding atmosphere are explicitly treated, the orbit of the spacecraft may not be predicted accurately. there were. Further, the problem of accurately predicting such an orbit is not limited to spacecraft, and is common to all mobile bodies that move in an environment where resistance is received.

  The present invention has been made in consideration of such circumstances, and an object thereof is to provide a prediction device, a prediction method, and a prediction program that can accurately predict the trajectory of a moving body.

  According to one aspect of the present invention, an acquisition unit that acquires environment information indicating an environment of a space in which a target moving body moves, and an environment unit indicating an environment of a space in which a certain moving body has moved are input. The model learned to output the trajectory information indicating the future trajectory of the, based on the output result of the model obtained by inputting the environment information acquired by the acquisition unit, the target And a prediction unit that predicts a future trajectory of the moving body.

  According to one aspect of the present invention, it is possible to accurately predict the trajectory of a moving body.

It is a figure which shows an example of a structure of the prediction apparatus 100 of embodiment. It is a figure which shows an example of the time history of a track radius change rate Adot. It is a figure which shows an example of the time history of a solar activity index SF. It is a figure which shows an example of the time history of the geomagnetic activity index Ap. It is a figure which shows an example of the time history of the global average value TEC of the total number of electrons in the ionosphere. It is a flow chart which shows an example of a series of processings performed by control part 110 at the time of learning. It is a figure which shows an example of the prediction result of each model. It is a figure which shows an example of the sum of the absolute error of the predicted value and observed value of each model. It is a flow chart which shows an example of a series of processings performed by control part 110 at the time of operation.

  Hereinafter, embodiments of a prediction device, a prediction method, and a prediction program according to the present invention will be described with reference to the drawings. The prediction device according to the present embodiment is a device that predicts a future orbit of a spacecraft such as an artificial satellite, a space probe, a space station, or space debris by using machine learning. Further, the prediction device is not limited to the spacecraft, and may predict a future trajectory (movement trajectory) of another moving body moving on the ground, underwater, or in the sky (for example, the stratosphere).

  FIG. 1 is a diagram illustrating an example of the configuration of the prediction device 100 according to the embodiment. As illustrated, the prediction device 100 includes, for example, a communication unit 102, a control unit 110, and a storage unit 130. The prediction device 100 may be a single device or a system in which a plurality of devices connected via a network such as a WAN (Wide Area Network) or a LAN (Local Area Network) operate in cooperation with each other. It may be. That is, the prediction device 100 may be realized by a plurality of computers (processors) included in a system using distributed computing or cloud computing.

  The communication unit 102 includes, for example, a communication interface such as a NIC (Network Interface Card) and a DMA (Direct Memory Access) controller. The communication unit 102 communicates with a predetermined terminal device or the like via a network such as WAN or LAN. The predetermined terminal device is, for example, a computer that can be used by a ground operator who monitors the orbit of the spacecraft.

  The control unit 110 includes, for example, a learning unit 112, an acquisition unit 114, a prediction unit 116, and an output control unit 118. The components of the control unit 110 are realized, for example, by a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) executing a program stored in the storage unit 130. Further, some or all of the constituent elements of the control unit 110 may be realized by hardware such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), or an FPGA (Field-Programmable Gate Array). , May be realized by the cooperation of software and hardware.

  The storage unit 130 is realized by, for example, a HDD (Hard Disc Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a ROM (Read Only Memory), a RAM (Random Access Memory), and the like. The storage unit 130 stores learning model information 132, a trajectory transition database 134, an environment database 136, and the like, in addition to various programs such as firmware and application programs.

  The learning model information 132 is information (program or data structure) defining a learning model used when predicting a future trajectory of a mobile body (for example, a spacecraft). The learning model includes, for example, a model of a part or all of LASSO (Least Absolute Shrinkage and Selection Operator) which is one of regularization regression, random forest, Gaussian process regression, and neural network. The learning model may include a model based on another regression method in addition to or in place of the above four types of models.

The orbit transition database 134 is a database in which orbit information indicating the temporal transition of the spacecraft orbit is recorded. For example, when the spacecraft is navigating in the upper atmosphere at an altitude of 100 [km] to 1000 [km], the spacecraft receives a resistance force from the atmosphere due to the influence of the slightly existing atmosphere. As a result, acceleration (acceleration for decelerating the body) a _drag is generated in the spacecraft, and the spacecraft gradually approaches (falls) from the orbit to the ground surface. In the orbit transition database 134, the orbit radius change rate Adot, which indicates how much the spacecraft that can generate the above acceleration a _drag falls from the orbit per day, is the time history (time transition) of the spacecraft's orbit. Is stored as The orbit transition database 134 may be updated each time the orbit of the spacecraft (orbit radius change rate Adot) is newly observed.

  FIG. 2 is a diagram showing an example of a time history of the trajectory radius change rate Adot recorded in the trajectory transition database 134. In the illustrated example, the orbit radius change rate Adot observed every day in the period from approximately 2012 to 2018 is recorded in the orbit transition database 134. For example, the orbital radius change rate Adot may be acquired from orbital information TLE (Two Line Elemnt) of a spacecraft released on the Internet by the United States, orbital information of a spacecraft held by a space agency in each country.

  The environment database 136 is a database in which environment information indicating the environment of the space in which the spacecraft moves along the orbit (that is, the upper atmosphere) is recorded.

  The environmental information of the upper atmosphere includes, for example, a solar activity index SF (Solar Flux), a geomagnetic activity index Ap, and a global average value (or also a daily average value) of ionization layer total electron content TEC (Total electron content). And some or all of the information is included.

  The solar activity index SF is a radio wave flux value per frequency when the radio wave wavelength is 10.7 [cm]. The solar activity index SF is also called solar flux or solar radio flux. Further, the solar activity index SF may be a radio wave density intensity based on a radio wave wavelength of 10.7 [cm]. In this case, the solar activity index SF is also called F10.7.

  The geomagnetic activity index Ap is an average of ap indexes showing the activity of geomagnetism based on observations at geomagnetic observatories in 12 countries around the world, and observed eight times a day.

  The global average value of the total number of electrons in the ionosphere is calculated from the amount of delay in the line-of-sight direction of a signal obtained by using positioning signal data acquired by GNSS (Global Navigation Satellite System) receiving stations around the world.

  The solar activity index SF and the geomagnetic activity index Ap are widely known as factors related to the atmospheric density that strongly affect the orbital transition of the spacecraft. On the other hand, the global average of the total number of electrons in the ionosphere is not known to what extent it affects the orbital transition of the spacecraft. The environment information of outer space may include other information in addition to or in place of the above-mentioned three information. The environmental database 136 may be updated each time the solar activity index SF, the geomagnetic activity index Ap, the global average of the total number of electrons in the ionosphere, and the like are observed.

  FIG. 3 is a diagram showing an example of the time history of the solar activity index SF, FIG. 4 is a diagram showing an example of the time history of the geomagnetic activity index Ap, and FIG. 5 is a global average value of the total number of electrons in the ionosphere. It is a figure which shows an example of the time history of. It is assumed that these pieces of environmental information are observed in the same period as the orbit radius change rate Adot or in a period longer than that.

[Processing flow during learning (training)]
Hereinafter, the flow of a series of processing at the time of learning of the control unit 110 will be described with reference to the flowchart. Learning is a state in which a learning model used in operation is learned. FIG. 6 is a flowchart showing an example of a series of processes executed by the control unit 110 during learning. The process of this flowchart may be repeated, for example, at a predetermined cycle. Further, when the prediction device 100 is realized by a plurality of computers included in a system using distributed computing or cloud computing, some or all of the processing of this flowchart may be processed in parallel by a plurality of computers. .

  First, the learning unit 112 selects, from a plurality of spacecraft, a spacecraft to be model-learned (hereinafter referred to as a spacecraft of interest) (step S100).

Next, the learning unit 112, from the orbit transition database 134 stored in the storage unit 130, from the reference date and time (hereinafter, reference date and time) t, the spacecraft of interest observed at the future date and time t + N N days ahead. The orbital radius change rate Adot _{t + N} is extracted as orbital information, and the solar activity index SF, the geomagnetic activity index Ap, and the geomagnetic activity index Ap observed from the environmental database 136 in the period T from the reference date t to the date t-T before T days, and The global average value of the total number of electrons in the ionosphere TEC is extracted as the environmental information _{Xt-T | t} (step S102). The Nth day is an example of the “second predetermined day”, and the Tth day is an example of the “first predetermined day”.

For example, when N = 1 and T = 10, the learning unit 112 orbits the orbit radius change rate Adot _{t + 1} of the spacecraft of interest observed at the time t + 1, which is one day after the reference date t, from the orbit transition database 134. The solar activity index SF, the geomagnetic activity index Ap, and the global average value of the total number of electrons in the ionosphere observed for 10 days before the reference date t are extracted as information from the environmental database 136, and are set as environmental information _{Xt-10 | t.} Extract. It should be noted that N may be arbitrarily changed depending on how far into the future it is predicted, and N may be 1 in the future one day later and N may be 7 in the future one week later. May be Further, T may be changed, for example, according to the rotation cycle of the sun, and is preferably about 83 days, which is close to three times the rotation cycle of the sun.

Next, the learning unit 112 selects an index value as an explanatory variable of the learning model from a plurality of index values included in the environment information _{Xt-10 | t} extracted from the environment database 136 (step S104).

For example, the learning model is represented by a function model having the trajectory radius change rate Adot _{t + N} as an objective variable when the environment information X _{t−T | t} is an explanatory variable, as shown in Expression (1). This learning model uses the environmental information Xt _{-T |} t from the reference date and time t to T days before to obtain the orbit radius change rate Adot _{t + N} after N days from the reference date and time t. Further, the mathematical expression (2) is N = 1 and T = 83, and the environmental information X _{t-83 | t} is the global average value of the solar activity index SF, the geomagnetic activity index Ap, and the ionospheric total electron number. It is a mathematical expression showing a learning model when three elements are included.

The environmental information _{Xt-T | t} is a part of the solar activity index SF, the geomagnetic activity index Ap, or the global average value of the total number of electrons in the ionosphere from the reference date t to the date t-T before T days, or All are multidimensional information (eg, vectors) included as elements. For example, solar activity index SF until 83 days ago, the global mean value of the geomagnetic activity index Ap, and ionospheric total electron number of environmental information X _{t-T |} if extracted as _t, the environmental information X _{t-T | t} is , 84 × 3, for a total of 252 elements.

Each of the plurality of elements included in the environment information _{Xt-T | t} is treated as an explanatory variable of the learning model. However, among the plurality of elements, elements having correlation with each other are included or the objective variable is included. There may be elements that have a low ability to explain.

Therefore, the learning unit 112 removes, from the plurality of elements included in the environment information X _{t−T | t} , the elements having a correlation with each other and the elements having a low ability to explain the objective variable, to thereby explain the explanatory variables of the learning model. Reduce the number of. In other words, the learning unit 112 compresses the dimension of the environment information _{Xt-T | t} .

Specifically, the learning unit 112 may reduce the number of explanatory variables of the learning model by using the selection result of the explanatory variables by LASSO, or among the characteristic amounts of the explanatory variables obtained by the random forest. , Even if the number of explanatory variables of the learning model is reduced by leaving the explanatory variables having the characteristic amount up to the number [%] of the maximum characteristic amount and removing the explanatory variables having the other characteristic amount. Good. For example, the learning unit 112 may leave the explanatory variables having the characteristic amount up to about 5% of the maximum characteristic amount among the characteristic amounts of the plurality of explanatory variables obtained by the random forest. Note that this numerical value is merely an example, and may be an arbitrary numerical value such as 1 [%] or 2 [%]. In this way, by compressing the dimension of the environment information X _{t−T | t} as the explanatory variable of the learning model, multicollinearity can be avoided and only the explanatory variables that are more likely to affect the objective variable. Can be extracted, and can be modeled with a simpler structure. As a result, over-learning can be suppressed.

Next, the learning unit 112, based on the elements of the environment information Xt _{-T | t} selected to suppress overlearning and the trajectory radius change rate Adot _{t + N} extracted as trajectory information from the trajectory transition database 134, The learning model of the space machine of interest is learned (step S106).

  For example, the learning unit 112 learns a learning model based on each learning method of LASSO, random forest, Gaussian process regression, and neural network, and generates four types of learning models. Expression (3) is an expression showing an example of a learning model based on LASSO.

In Formula (3), w ₁ represents a weighting coefficient vector by which the solar activity index SF is multiplied, w ₂ represents a weighting coefficient vector by which the geomagnetic activity index Ap is multiplied, and w ₃ represents a total ionospheric electron number. It represents a weighting coefficient vector by which the global average value is multiplied. For example, the learning unit 112 uses a learning model based on LASSO as exemplified in Expression (3), and the solar activity index SF _{t-83 | t} and the geomagnetic activity index Ap _{t-83 |} from the reference date t to 83 days before. and _t, global ionospheric total electron mean _{TEC t-83 |} substituting and _t, predicting the orbital radius change rate Adot _{t + 1} after one day of the reference date t. Then, the learning unit 112, orbital radius change rate predicted based on the learning model Adot _{t + 1} (i.e. predicted value), orbital radius change rate extracted as a track information after 1 day of the reference date t from the track transition database 134 Adot _{t + 1} The learning model is learned by deriving a difference from (i.e., an observed value) and determining the weighting coefficient vector w, which is a parameter of the learning model, and the bias component so that the derived difference becomes small. In addition to the weighting coefficient vector w and the bias component, the parameters to be learned may include hyperparameters that the designer can arbitrarily determine.

  Next, the learning unit 112 associates the learned learning model with the spacecraft of interest and stores this in the storage unit 130 as learning model information 132 (step S108). This completes the processing of this flowchart.

  In this way, the learning unit 112 voluntarily repeats the processing of the above-described flowcharts even if there is no particular instruction from an external device or the like (such as a prediction request to be described later). The learning model learned by using the past trajectory information and the past environment information of is re-learned (trained) many times. Thereby, the accuracy of the learning model can be improved every day.

  FIG. 7 is a diagram showing an example of the transition between the trajectory radius change rate Adot predicted by each learning model and the actually observed trajectory radius change rate Adot. Each model in the figure is a learning model that is obtained by dividing each of the databases described above into two and using one of the divided data. The orbital radius change rate Adot predicted by each of the four learning models changes with the same tendency as the temporal transition of the orbital radius change rate Adot, which is an observed value. In addition, all learning models can predict the orbit radius change rate Adot with an accuracy of an error of about 1.0 at the maximum with respect to the actual observed value, which shows that it is sufficiently practical.

  FIG. 8 is a diagram showing an example of the sum of absolute errors between the predicted value and the observed value of each model. In the illustrated example, the sum of the absolute values of the errors between the predicted value Adot of the learning model based on LASSO and the observed value Adot is 5.1 [m / day], and learning based on random forest (RF in the figure) The sum of the absolute values of the errors between the predicted value Adot of the model and the observed value Adot is 4.0 [m / day], and the predicted value Adot of the learning model based on Gaussian process regression (GP in the figure) and the observed value The sum of the absolute values of the error with the value Adot is 2.9 [m / day], which is the absolute value of the error between the predicted value Adot of the learning model based on the neural network (NN in the figure) and the observed value Adot. The sum represents 5.6 [m / day]. From such a result, in one example, the learning model based on Gaussian process regression has the highest prediction accuracy among the four learning models.

[Processing flow during operation]
Hereinafter, the flow of a series of processing when the control unit 110 is in operation will be described with reference to the flowchart. The operating state is a state in which the future orbit of the spacecraft is predicted by using the learning model learned during the learning. FIG. 9 is a flowchart showing an example of a series of processes executed by the control unit 110 during operation. The process of this flowchart may be repeated, for example, at a predetermined cycle. When the prediction device 100 is realized by a plurality of computers included in a system using distributed computing or cloud computing, part or all of the processing of this flowchart is performed by a plurality of computers as in the learning. May be processed in parallel by.

  First, the acquisition unit 114 requests the communication unit 102 to predict a future orbit of a spacecraft to be monitored from a predetermined terminal device used by a ground operator or the like (hereinafter, referred to as a prediction request). Is received (step S200). The prediction request includes, for example, the identification information of the spacecraft designated as the monitoring target and the current date and time.

When the communication unit 102 receives the prediction request, the acquisition unit 114, from the environment database 136 stored in the storage unit 130, observes solar activity during a period T from the current date and time τ to a date and time τ-T that is T days ago. Environmental information _{Xτ-T | τ} including the index SF, the geomagnetic activity index Ap, and the global average value of the total number of electrons in the ionosphere is acquired (step S202). For example, at the time of learning, when the learning is performed using the environment information X _{t-83 |} t from the reference date t to 83 days before as the explanatory variable of the learning model, the acquisition unit 114 determines from the environment database 136 from the current date τ. Environmental information X _{τ-83 | τ} up to 83 days ago is acquired.

  Next, the prediction unit 116 refers to the identification information of the spacecraft included in the prediction request, and selects the learning model associated with the spacecraft to be monitored from among the plurality of learning models included in the learning model information 132. A selection is made (step S204). As described above, when the learning model based on each of the LASSO, the random forest, the Gaussian process regression, and the neural network is associated with the spacecraft, the predicting unit 116 selects from among the four learning models, The learning model with the smallest prediction error (highest prediction accuracy) may be selected. For example, the prediction unit 116 selects the learning model having the smallest sum of the absolute values of the errors between the prediction value Adot and the observed value Adot as the learning model having the smallest prediction error. In the example of FIG. 8, a learning model based on Gaussian process regression is selected.

  In addition, when the prediction unit 116 selects the learning model of the spacecraft to be monitored specified by the prediction request from the plurality of learning models included in the learning model information 132, the spacecraft whose identification information does not necessarily match. Does not have to be selected. For example, the prediction unit 116 selects a learning model of the same type of spacecraft that has the same shape as the spacecraft to be monitored specified by the prediction request from the plurality of learning models included in the learning model information 132. You can do it. In addition, the prediction unit 116 may select a learning model of a spacecraft having the same navigation altitude as the spacecraft to be monitored specified by the prediction request from the plurality of learning models included in the learning model information 132. Good. As a result, even if a spacecraft that has not generated a learning model is selected as a spacecraft to be monitored, the learning model of a similar spacecraft can be used to determine the future trajectory of the spacecraft to be monitored. Can be predicted.

  Next, the prediction unit 116 uses the learning model of the selected spacecraft to be monitored to predict the future trajectory of the spacecraft to be monitored (step S206).

Specifically, the prediction unit 116 sets the environment information X _{τ-T | τ} from the current date and time τ acquired by the acquisition unit 114 to T days before as an explanatory variable for the selected learning model of the spacecraft to be monitored. By inputting as, the learning model is made to predict the orbital radius change rate Adot of the spacecraft to be monitored. For example, at the time of learning, when learning is performed using the orbital radius change rate Adot _{t + 1} of the spacecraft of interest observed at a future time t + 1 one day ahead of the reference time t as the objective variable of the learning model, the learning model is The orbit radius change rate Adot _{τ + 1} of the spacecraft that will be observed at a future date / time τ + 1 one day after the current date / time τ will be predicted.

  Next, the output control unit 118 controls the communication unit 102 so that the information indicating the future orbit of the spacecraft to be monitored, which is the prediction result of the prediction unit 116, is transmitted to the predetermined terminal device that is the transmission source of the prediction request. (Step S208). This completes the processing of this flowchart.

According to the embodiment described above, the environmental information indicating the environment of the upper atmosphere in which the spacecraft to be monitored is navigating is acquired, and the environmental information _{Xt-T of the} upper atmosphere in the period from the reference date t to T days ago. _{When | t} is input, the acquired environment information is input to the learning model learned to output the orbital radius change rate Adot _{t + N} of the spacecraft N days ahead from the reference time t, and the environment information is input. Since the future orbit of the spacecraft to be monitored is predicted based on the result output by the learning model that has input, the orbit of the spacecraft can be accurately predicted.

Generally, from the attitude, shape, mass, resistance coefficient of the spacecraft, the projected area of the spacecraft in the traveling direction, the relative velocity of the spacecraft to the surrounding atmosphere, the atmospheric density of the surrounding atmosphere, etc., theoretically, the resistance of the upper atmosphere It is known that the acceleration a _drag of the spacecraft caused by the force can be derived, and it is conceivable to predict the future trajectory transition of the spacecraft based on the acceleration a _drag derived by such a theoretical formula.

However, in the upper atmosphere where spacecraft are navigating, energy inflow from the lower atmosphere and outer space is intense, and it constantly fluctuates, resulting in a temperature difference of several hundreds of degrees and an atmospheric density of 1 billion minutes on the ground surface. Of the above various parameters, the atmospheric density of the surrounding atmosphere, the resistance coefficient of the spacecraft depending on the temperature of the surrounding atmosphere, and the attitude information of the spacecraft are uncertain parameters. Can be. Therefore, when the acceleration a _drag is derived with the parameter having such uncertainty as the positive, the predicted orbit of the spacecraft may be uncertain.

  On the other hand, in the present embodiment, when predicting the orbit, the parameters such as the atmospheric density, the resistance coefficient, and the attitude information having uncertainty are not explicitly treated (without being an explanatory variable), and the already-observed orbit is Since the model is learned based on the information and the environment information and the model predicts the orbit of the spacecraft, it is possible to accurately predict the orbit of the spacecraft.

Further, according to the above-described embodiment, at the time of model learning, elements having low correlation with each other and elements having low ability to explain the objective variable are removed from the plurality of elements included in the environment information X _{t−T | t.} Since the number of explanatory variables in the learning model is reduced, overlearning can be suppressed.

  In addition, according to the above-described embodiment, each model of LASSO, random forest, Gaussian process regression, and neural network is learned, and the future trajectory of the spacecraft to be monitored is calculated using the learning model with the smallest prediction error. Therefore, the orbit of the spacecraft can be predicted more accurately.

(Modification)
Hereinafter, modified examples of the above-described embodiment will be described. Although the learning model is described as predicting the future trajectory of the spacecraft in the above-described embodiment, the learning model is not limited to this. For example, the learning model may be trained to predict the future trajectory of a moving body such as a vehicle moving on the ground, or the future trajectory of a moving body such as a ship navigating over water or a submarine navigating underwater. May be learned, or may be learned to predict the future trajectory of a mobile such as an aircraft flying above. In such a case, the environment information may indicate the environment of the space in which each mobile body moves.

In the above-described embodiment, the learning model is described as predicting the trajectory radius change rate Adot, but the present invention is not limited to this. For example, the learning unit 112 includes environmental information _{Xt-T | that} includes, as elements, the solar activity index SF, the geomagnetic activity index Ap, and the global average of the total number of electrons in the ionosphere from the reference date t to the date t-T before T days. _{By using t} as an explanatory variable and learning the learning model with the acceleration a _drag of the spacecraft observed at a future time t + N N days ahead from the reference date t as the objective variable, the learning model is given a spacecraft N days ahead. The acceleration a _drag may be predicted.

  As described above, the embodiments for carrying out the present invention have been described using the embodiments, but the present invention is not limited to such embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

100 ... Prediction device, 102 ... Communication part, 110 ... Control part, 112 ... Learning part, 114 ... Acquisition part, 116 ... Prediction part, 118 ... Output control part, 130 ... Storage part

Claims (10)

An acquisition unit that acquires environment information indicating the environment of the space in which the target moving body moves,
When the environment information indicating the environment of the space where a certain moving body has moved is input, for the model learned to output the trajectory information indicating the future trajectory of the certain moving body, it is acquired by the acquisition unit. Based on the output result of the model obtained by inputting the environmental information, a prediction unit that predicts a future trajectory of the target moving body,
Prediction device. The moving body is a spacecraft,
The space is the upper atmosphere,
The model outputs the change rate of the orbital radius of the spacecraft as orbital information indicating a future orbit of the spacecraft,
The prediction device according to claim 1. The environment further includes a learning unit that learns the model such that environment information indicating the environment of a space in which a moving body has moved in the past is an explanatory variable, and trajectory information indicating a past trajectory of a certain moving body is an objective variable.
The prediction device according to claim 1. The learning unit uses the environment information indicating the environment of the space observed at a date and time that is a first predetermined day before the reference date and time as the explanatory variable, and the movement that is observed at a date and time that is a second predetermined date ahead from the reference date and time. Trajectory information indicating the trajectory of the body is used as the objective variable to learn the model,
The prediction device according to claim 3. The learning unit learns some or all models of regularization regression, random forest, Gaussian process regression, or neural network,
The predicting unit predicts a future trajectory of the target moving body based on an output result of at least one of the one or more models learned by the learning unit,
The prediction device according to claim 3 or 4. The prediction unit predicts a future trajectory of the target moving body based on the output result of the model with the smallest prediction error among the one or more models learned by the learning unit.
The prediction device according to claim 5. The learning unit reduces the number of environment information used as the explanatory variables when learning the model,
The prediction device according to any one of claims 3 to 6. The environmental information includes at least one of a solar activity index, a geomagnetic activity index, and the total number of electrons in the ionosphere,
The prediction device according to any one of claims 1 to 7. Computer
Acquire environment information indicating the environment of the space where the target moving body moves,
When the environmental information indicating the environment of the space where a certain moving body has moved is input, the acquired environmental information is stored with respect to the model learned to output the trajectory information indicating the future trajectory of the certain moving body. Based on the output result of the model obtained by input, to predict the future trajectory of the moving body of the object,
Prediction method. On the computer,
A process of acquiring environment information indicating the environment of the space in which the target moving body moves,
When the environmental information indicating the environment of the space where a certain moving body has moved is input, the acquired environmental information is stored with respect to the model learned to output the trajectory information indicating the future trajectory of the certain moving body. Based on the output result of the model obtained by inputting, a process of predicting a future trajectory of the target moving body,
Prediction program to execute.

Images