Traffic prediction device, traffic management device, traffic management system, learning device, traffic prediction method, learning method, traffic prediction program, and learning program

Abstract

To obtain an operation prediction device capable of predicting an operation state of a train in the future with high accuracy.SOLUTION: An operation prediction device 20 includes: a data acquisition part 251 for acquiring, from a train 10, prediction object data including operation data showing an operation state of the train 10 and human flow data showing a state of human flow of users using the train 10; and a prediction part for predicting at least one of an operation state and a state of human flow in the future from the prediction object data by using a prediction model which has learned similarity data similar to recent data in past data and the recent data, in which data for learning including the operation data and the human flow data is divided into the past data and the recent data acquired later than the past data. The similarity data is data determined to be similar to the recent data in the past data by using a similarity determination model which has learned the past data and the recent data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an operation prediction device that predicts train operation conditions, an operation management device, an operation management system, a learning device, an operation prediction method, a learning method, an operation prediction program, and a learning program. [Background technology]

[0002] In the field of railways, future train operation conditions are predicted and train operations are adjusted based on the predicted operation conditions to alleviate congestion for train passengers or to enable efficient train operation.

[0003] Patent Document 1 discloses a system for predicting train operation, which predicts the arrival and departure times of currently operating trains at future stations and creates an operation prediction distribution, which is a distribution of train delay times and the probability of delays at those delay times. The system disclosed in Patent Document 1 extracts operation performance data when the operation conditions are similar to those of currently operating trains from past operation performance data, and creates an operation prediction distribution based on the extracted operation performance data. The system disclosed in Patent Document 1 determines whether the current operation conditions are similar to past operation conditions by individually determining whether multiple conditions are similar. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-98425 Summary of the Invention [Problem to be solved by the invention]

[0005] To enable highly accurate prediction of train operation status, it is desirable to set many conditions that can affect the operation status as operation conditions. The system disclosed in Patent Document 1 individually determines whether multiple conditions, which are operation conditions, are similar. Therefore, the greater the number of conditions, the more difficult it is to determine whether current operation conditions are similar to past operation conditions. Furthermore, since the degree of influence on operation status varies for each condition, the greater the number of conditions, the more difficult it is to reflect the degree of influence of each condition in determining whether they are similar. For this reason, the technology disclosed in Patent Document 1 has a problem in that it is difficult to extract operation history data when operation conditions are similar to those of currently operating trains, making it difficult to accurately predict future train operation status.

[0006] The present disclosure has been made in consideration of the above, and aims to provide an operation prediction device that can predict future train operation conditions with high accuracy. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems and achieve the object, the operation prediction device according to the present disclosure includes a data acquisition unit that acquires, from a train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the flow status of passengers using the train, and a prediction unit that divides learning data including the operation data and people flow data into past data and most recent data acquired after the past data, and predicts at least one of the future operation status and people flow status from the prediction target data using a prediction model that has been trained with similar data from the past data that is similar to the most recent data. The similar data is data from the past data that has been determined to be similar to the most recent data using a similarity determination model that has been trained with the past data and the most recent data. [Effects of the Invention]

[0008] The operation prediction device according to the present disclosure has the effect of being able to predict future train operation conditions with high accuracy. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating a configuration example of a traffic management system according to a first embodiment. [Figure 2] FIG. 1 is a first diagram illustrating an example of a table in a feature database held by an operation prediction device according to a first embodiment; [Figure 3] FIG. 1 is a conceptual diagram showing the relationship between past data, most recent data, and similar data in the first embodiment. [Figure 4] FIG. 1 is a first diagram for explaining details of generation of a similarity determination model and determination of similar data in an operation prediction device according to a first embodiment. [Figure 5] FIG. 2 is a second diagram for explaining details of generation of a similarity determination model and determination of similar data in the operation prediction device according to the first embodiment. [Figure 6] FIG. 6 is a diagram illustrating a threshold value used to determine similar data in the example shown in FIG. 5; [Figure 7] FIG. 3 is a third diagram for explaining details of generation of a similarity determination model and determination of similar data in the operation prediction device according to the first embodiment. [Figure 8] FIG. 2 is a second diagram illustrating an example of a table in a feature database held by the operation prediction device according to the first embodiment; [Figure 9] 1 is a flowchart showing an example of an operation procedure during learning by the operation prediction device according to the first embodiment; [Figure 10] A flowchart showing an example of a detailed procedure for determining similar data in step S4 shown in FIG. 9. [Figure 11] 1 is a flowchart illustrating an example of a procedure of an operation performed by the operation prediction device according to the first embodiment at the time of prediction. [Figure 12] FIG. 1 is a diagram illustrating a configuration example of a traffic management system according to a first modified example of the first embodiment. [Figure 13] FIG. 10 is a diagram illustrating a configuration example of a traffic management system according to a second modification of the first embodiment. [Figure 14]FIG. 10 is a diagram illustrating a configuration example of a traffic control system according to a second embodiment. [Figure 15] FIG. 10 is a diagram for explaining conversion from a first feature quantity to a second feature quantity in a feature quantity conversion unit of an operation prediction device according to a second embodiment. [Figure 16] FIG. 10 is a diagram for explaining a first example of a second feature amount used to determine similar data in the operation prediction device according to the second embodiment; [Figure 17] FIG. 10 is a first diagram illustrating a second example of a second feature amount used to determine similar data in the operation prediction device according to the second embodiment; [Figure 18] FIG. 2 is a second diagram illustrating a second example of a second feature amount used to determine similar data in the operation prediction device according to the second embodiment; [Figure 19] 10 is a flowchart showing an example of an operation procedure when similar data is determined by the operation prediction device according to the second embodiment. [Figure 20] A flowchart showing an example of a detailed procedure for converting feature quantities in step S32 shown in FIG. 19. [Figure 21] FIG. 1 is a diagram illustrating an example of a hardware configuration of an operation prediction device constituting an operation management system according to a first or second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] An operation prediction device, an operation management device, an operation management system, a learning device, an operation prediction method, a learning method, an operation prediction program, and a learning program according to embodiments will be described in detail below with reference to the accompanying drawings.

[0011] Embodiment 1 1 is a diagram illustrating an example of the configuration of a traffic control system 100 according to a first embodiment. The traffic control system 100 is a system that manages the operation of trains 10 on a railway. The traffic control system 100 includes a plurality of trains 10, a traffic control device 30 that manages the operation of each train 10, and a diagram creation device 40 that creates a train diagram for each train 10.

[0012] Each train 10 consists of one or more cars. Each train 10 is equipped with a communication unit 11 that communicates with the traffic management device 30 and on-board equipment that acquires equipment data. The on-board equipment is not shown in the figure. The equipment data is data that indicates the status of the train 10. Examples of the equipment data include the acceleration or speed of the train 10, door opening / closing information regarding the opening and closing of doors that are used by passengers to board and alight, running position information of the train 10, the occupancy rate of each car, the number of passengers boarding and alighting in each car, running time between stations, and stop time at each station. Note that the equipment data is not limited to these. The communication unit 11 transmits the equipment data acquired by the on-board equipment to the traffic management device 30.

[0013] The operation management device 30 includes an operation prediction device 20 that predicts the operation status of the train 10, a communication unit 31, a data collection device 32, a countermeasure processing unit 33, and a memory unit 34a that stores a train schedule database 34.

[0014] The communication unit 31 communicates with devices external to the traffic management device 30. The communication unit 31 communicates with the communication unit 11 of each train 10 and with the diagram creation device 40. The communication unit 31 and the communication unit 11 communicate with each other via wireless communication. There are no particular limitations on the communication method of wireless communication.

[0015] The communication unit 31 receives equipment data output from the communication unit 11 of each train 10 and outputs the equipment data to the data collection device 32. The data collection device 32 collects the equipment data received by the communication unit 31 and outputs the collected equipment data to the operation prediction device 20. The data collection device 32 collects the equipment data by acquiring the equipment data at regular intervals. The interval at which the data collection device 32 acquires the equipment data can be set arbitrarily between several tens of milliseconds and one second, for example.

[0016] The operation prediction device 20 includes a data extraction unit 21, a feature extraction unit 22, a storage unit 23a that holds a feature database 23, a learning device 24, an inference device 25, and a model storage unit .

[0017] The data extraction unit 21 extracts operation data and people flow data from the equipment data collected by the data collection device 32. The operation data and people flow data are data necessary for extracting feature quantities, which will be described later. The operation data is data indicating the operation status of the train 10. The people flow data is data indicating the flow of people using the train 10. Examples of the operation data include the departure time of the train 10, which is the time when the train 10 departs from a station, the running time between stations, and the stop time at each station. Examples of the people flow data include the occupancy rate of each car and the number of people getting on and off each car. The data extraction unit 21 can determine whether the equipment data is data while the train 10 is running or data while the train is stopped at a station by using the acceleration, speed, door opening / closing information, running position information, etc. of the train 10 included in the equipment data. Note that the operation data and people flow data are not limited to the data exemplified in the first embodiment. The data extraction unit 21 outputs the extracted operation data and the extracted people flow data to the feature extraction unit 22.

[0018] The feature extraction unit 22 extracts features, which are data between stations, from the time-series data, which are the operation data and people flow data, extracted by the data extraction unit 21. The feature extraction unit 22 extracts features by converting the time-series data into data for each unit, using the period from when the train 10 departs from a first station to when the train 10 arrives at a second station, which is the next stop, and departs from the second station, or the period from when the train 10 arrives at a first station to when the train 10 departs from the first station and arrives at the second station, which is the next stop, as a unit. In this way, the feature extraction unit 22 generates features from the time-series data. In machine learning, features are also called explanatory variables.

[0019] The feature extraction unit 22 may generate feature values ​​based on the time series data and the information on the bus schedule stored in the bus schedule database 34. An example of a feature value generated based on the information on the bus schedule and the time series data is the delay time of the train 10. The difference between the departure time indicated in the bus schedule and the departure time indicated in the operation data corresponds to the delay time.

[0020] The features extracted by the feature extraction unit 22 are stored in the feature database 23. Date data is associated with each row of the table in the feature database 23. The date data is data on the date and time when operation data or people flow data is acquired. The date and time when operation data or people flow data is acquired is the date and time when the equipment data that is the source of the operation data or people flow data is acquired for each train 10.

[0021] The processing by the operation prediction device 20 is divided into a learning phase and an inference phase. The above description of the operation prediction device 20 does not distinguish between the learning phase and the inference phase. In the learning phase, the operation prediction device 20 generates a prediction model for at least one of a future operation status and a future pedestrian flow status. In the inference phase, the operation prediction device 20 uses the prediction model to predict at least one of a future operation status and a future pedestrian flow status.

[0022] In the learning phase, the data extraction unit 21 extracts learning data including operation data and people flow data from the equipment data collected by the data collection device 32. The feature extraction unit 22 extracts features, which are learning data for each station, from the learning data, which is time-series data. The features extracted by the feature extraction unit 22 are stored in the feature database 23. In the feature database 23, the learning data is associated with date and time data indicating the date and time when the learning data was acquired.

[0023] The learning device 24 performs learning in the learning phase. The learning device 24 includes a data classification unit 241, a similarity determination model generation unit 242, a similar data determination unit 243, a data acquisition unit 244, and a prediction model generation unit 245. The data classification unit 241 reads learning data from the feature database 23 and classifies the learning data into past data and most recent data acquired after the past data. The data classification unit 241 outputs the classification results. For example, the data classification unit 241 outputs a label indicating whether the data is the most recent data as the classification result. Each row of the table in the feature database 23 is associated with a label output by the data classification unit 241.

[0024] The similarity determination model generation unit 242 acquires the past data and the most recent data by reading them from the feature database 23. The similarity determination model generation unit 242 generates a similarity determination model by learning the past data and the most recent data. The similarity determination model is a model that determines similar data from the past data that is similar to the most recent data. In this way, the similarity determination model generation unit 242 divides the learning data into the past data and the most recent data, and generates a similarity determination model that determines similar data from the past data that is similar to the most recent data by learning the past data and the most recent data. The similarity determination model generation unit 242 builds a similarity determination model by learning the relationship between the feature and a label indicating whether the data is the most recent data, according to a supervised learning algorithm.

[0025] The similar data determination unit 243 acquires past data by reading the past data from the feature database 23. The similar data determination unit 243 acquires the similarity determination model generated by the similarity determination model generation unit 242. The similar data determination unit 243 determines similar data from the past data based on the similarity determination model. The similar data determination unit 243 outputs the result of the determination. For example, the similar data determination unit 243 outputs a label indicating whether the data is similar or not as the result of the determination. The similar data is data from the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data.

[0026] For learning in the prediction model generation unit 245, learning data consisting of the most recent data and similar data is used. That is, the prediction model generation unit 245 generates a prediction model by learning using learning data from which past data that does not correspond to similar data has been excluded. The data acquisition unit 244 acquires learning data consisting of the most recent data and similar data. The learning data consisting of the most recent data and similar data is input from the data acquisition unit 244 to the prediction model generation unit 245. Hereinafter, the learning data consisting of the most recent data and similar data will be referred to as learning input data.

[0027] Each row of the table in the feature database 23 is associated with a label indicating whether or not it is input data for learning. The label output by the similar data determination unit 243 can also be said to be a label indicating whether or not past data is input data for learning. Each row of past data in the table in the feature database 23 is associated with a label indicating whether or not it is similar data, as a label indicating whether or not it is input data for learning. Note that in each row of the most recent data in the table in the feature database 23, the label indicating whether or not it is input data for learning is set to a value indicating that it is input data for learning.

[0028] The data acquisition unit 244 references the labels indicating whether the data is input data for learning, and reads out the most recent data and similar data from the feature database 23. The data acquisition unit 244 acquires input data for learning by reading out the most recent data and similar data from the feature database 23. The data acquisition unit 244 outputs the acquired input data for learning to the prediction model generation unit 245.

[0029] The prediction model generation unit 245 generates a prediction model by learning the training input data, i.e., by learning the most recent data and similar data. The prediction model is a model that predicts at least one of the future operation status of the train 10 and the future pedestrian flow status of users using the train 10. In the first embodiment, the future refers to, for example, the time when the train 10 will travel several stations ahead while the train 10 is currently operating, or several minutes or hours ahead while the train 10 is currently operating. The future may also refer to the next day or several days ahead. The operation status predicted by the prediction model generation unit 245 includes delay time, stop time, or power consumption when traveling between stations. The pedestrian flow status predicted by the prediction model generation unit 245 includes, for example, the occupancy rate or the number of passengers getting on and off. The prediction model generation unit 245 outputs the generated prediction model. The model storage unit 26 stores the prediction model output from the prediction model generation unit 245.

[0030] In the inference phase, the data extraction unit 21 extracts prediction target data including operation data and people flow data from the equipment data collected by the data collection device 32. The feature extraction unit 22 extracts features, which are prediction target data on a station-to-station basis, from the prediction target data, which is time-series data. The features extracted by the feature extraction unit 22 are stored in a feature database 23.

[0031] In the inference phase, the inference device 25 predicts at least one of future operation conditions and people flow conditions. The inference device 25 includes a data acquisition unit 251 and an inference unit 252, which is a prediction unit. The data acquisition unit 251 acquires prediction target data, which is data for inference, by reading prediction target data from the feature database 23. In this way, the data acquisition unit 251 acquires prediction target data, including operation data and people flow data, from each train 10 via the communication unit 31, the data collection device 32, the data extraction unit 21, the feature extraction unit 22, and the feature database 23. The data acquisition unit 251 outputs the prediction target data to the inference unit 252.

[0032] The inference unit 252 reads a prediction model from the model storage unit 26. Based on the prediction target data and the prediction model, the inference unit 252 infers at least one of the future operation status of the train 10 and the future pedestrian flow status of users using the train 10. The inference unit 252 inputs the prediction target data acquired by the data acquisition unit 251 into the prediction model, and outputs a prediction result for at least one of the future operation status and the future pedestrian flow status. In this way, the prediction model generation unit 245 separates the learning data, including operation data and pedestrian flow data, into past data and most recent data, and predicts at least one of the future operation status and the pedestrian flow status from the prediction target data using a prediction model trained with similar data from the past data that is similar to the most recent data. The inference unit 252 outputs the prediction result of at least one of the future operation status and the pedestrian flow status to the countermeasure processing unit 33.

[0033] The countermeasure processing unit 33 formulates countermeasures according to the prediction result output by the inference unit 252. For example, when the inference unit 252 outputs a prediction result indicating a delay of the train 10, the countermeasure processing unit 33 formulates countermeasures that can mitigate the delay or prevent the delay. The countermeasure processing unit 33 formulates countermeasures such as changing the arrival and departure times of each train 10 or adjusting the number of trains 10 in operation. When a delay is predicted for a train 10 used for transferring at a station, the countermeasure processing unit 33 may also adjust the departure time of the transfer destination train 10.

[0034] The countermeasure processing unit 33 outputs instructions for executing the formulated countermeasures to the communication unit 31. The communication unit 31 transmits instructions for executing the formulated countermeasures to each train 10. Each train 10 adjusts its operation in accordance with the instructions output by the countermeasure processing unit 33. Alternatively, the communication unit 31 transmits instructions for executing the formulated countermeasures to the diagram creation device 40.

[0035] The bus schedule creation unit 41 of the bus schedule creation device 40 creates a bus schedule for each train 10. The bus schedule information is stored in the bus schedule database 34. The bus schedule creation unit 41 also adjusts the bus schedule in accordance with instructions output by the countermeasures processing unit 33. The bus schedule creation unit 41 not only adjusts the bus schedule for the day on which the prediction was made in accordance with the instructions, but may also create a bus schedule for a day after the day on which the prediction was made in accordance with the instructions.

[0036] The traffic management device 30 is a device capable of communicating with each of the multiple trains 10 and the timetable creation device 40. The traffic management device 30 may be a server device that is a server built in a cloud environment. The cloud environment includes computer resources provided in a cloud service platform. Since the server device is built in the cloud environment, it is also called a cloud server.

[0037] The traffic management device 30 may output the prediction results output by the inference device 25 to announce the predicted traffic status or the predicted pedestrian flow status to users. The traffic management system 100 may use the prediction results by the inference device 25 to present information about the predicted traffic status or the predicted pedestrian flow status on station information boards, a website, or an application.

[0038] Next, a description will be given of data stored in the feature database 23. Fig. 2 is a first diagram illustrating an example of a table in the feature database 23 held by the operation prediction device 20 according to the first embodiment.

[0039] In the example shown in FIG. 2, the columns of "Departure Time," "Delay Time," "Number of Passengers Boarding and Alighting," and "Future Delay Time" represent items of operation data or items of people flow data. Feature quantities are groups of data for each of multiple items and are shown in each row of the table. In FIG. 2, "Departure Time" is expressed in hours, minutes, and seconds. The units of "Delay Time" and "Future Delay Time" are both seconds. "Number of Passengers Boarding and Alighting" is expressed as a numerical value representing the number of people. "Future Delay Time" is the delay time that occurs, for example, n stations away (n is a positive integer) from a certain station. In the example shown in FIG. 2, the delay time at the next stop is defined as the future delay time. In the training data, feature quantities are linked to at least one of items related to future operation conditions and items related to future people flow conditions. In the example shown in FIG. 2, "Future Delay Time," an item related to future operation conditions, is linked to the feature quantity. The item related to future operation conditions is not limited to delay time, but may also be the stop time at a station, etc. Items regarding future passenger flow conditions include the occupancy rate or the number of people getting on and off.

[0040] In the example shown in FIG. 2, the three columns "year," "month," and "day" represent date data. Each row of the table is linked to date data that indicates the date and time when operation data or people flow data was acquired. Note that the "day" in the date data does not represent a period separated by midnight, but rather represents a period from the start of operation of the first train 10 to the end of operation of the last train 10.

[0041] The "Most Recent Data" column indicates a label indicating whether the data is the most recent data. In the example shown in Figure 2, the label "0" indicates that the data is not the most recent data, and the label "1" indicates that the data is the most recent data. In the example shown in Figure 2, operation data and people flow data acquired before June 3, 2022 are classified as past data, and operation data and people flow data acquired after June 4, 2022 are classified as the most recent data. The "Learning Input Data" column indicates a label indicating whether the data is learning input data. The table shown in Figure 2 is in a state before the value of the label indicating whether the data is learning input data is stored.

[0042] The feature amounts are explanatory variables used in learning by the similarity determination model generation unit 242. The label indicating whether the data is the most recent data is a target variable used in learning by the similarity determination model generation unit 242. That is, in each row of the table in the feature amount database 23, an explanatory variable and a target variable are associated with each other.

[0043] The data included in the feature quantities are not limited to the data of the items exemplified in FIG. 2 . The feature quantities can include data of various items regarding the operation status of the train 10 and data of various items regarding the pedestrian flow status. For example, the item regarding the operation status of the train 10 may be the travel time between stations or the stop time at each station. The item regarding the pedestrian flow status may be the occupancy rate while the train 10 is traveling, the fluctuation in occupancy rate while passengers are getting on and off at stations, or the number of passengers getting on and off at stations. The item regarding the pedestrian flow status may include not only an item indicating the status of passengers on the train 10, but also an item indicating the congestion status at the platform where the train 10 departs or arrives, the station premises, or the ticket gates. Two or more of the item indicating the congestion status of the platform, the item indicating the congestion status within the premises, and the item indicating the congestion status at the ticket gates may be combined and used as an item indicating the congestion status.

[0044] If the occupancy rate of each vehicle is acquired, the feature may include the occupancy rate of each vehicle. If the occupancy rate during travel is acquired by measuring the weight of the vehicle, the measured value may fluctuate, so the feature may use the average occupancy rate during travel. If the occupancy rate of the entire train 10 is acquired, the feature may include the occupancy rate of the entire train 10.

[0045] Occupancy rate fluctuation is an index that represents the degree of change in occupancy rate as passengers get on and off. Including occupancy rate fluctuation in the feature is useful for predicting delay times due to the frequency of passengers getting on and off, or predicting occupancy rates. Occupancy rate fluctuation is expressed by the average occupancy rate during boarding and alighting, the variance of occupancy rate during boarding and alighting, or the standard deviation of occupancy rate during boarding and alighting. If occupancy rates are acquired for each vehicle, the feature can include occupancy rate fluctuation for each vehicle. Note that, since the boarding and alighting status of the vehicle with the highest number of passengers getting on and alighting within train 10 has the greatest impact on the time required to board and alight from train 10, the occupancy rate fluctuation of the vehicle with the largest standard deviation of occupancy rate may be included in the feature.

[0046] The number of passengers boarding and alighting is the sum of the number of passengers alighting from the train 10 at a certain station and the number of passengers boarding the train 10 at that station. When the number of passengers boarding and alighting for each car is acquired, the feature can include the number of passengers boarding and alighting for each car. Note that the number of passengers alighting may be calculated based on the occupancy rate when the doors open upon arrival at the station and the minimum occupancy rate while the doors are open, and the calculated number of passengers alighting may be included in the feature. Alternatively, the number of passengers boarding may be calculated based on the occupancy rate when the doors close for departure from the station and the minimum occupancy rate while the doors are open, and the calculated number of passengers may be included in the feature.

[0047] The feature may include a station ID (identifier), weather, vehicle type, type of train 10, destination of train 10, direction of travel, running speed of train 10, amount of power consumed or regenerated when train 10 travels between stations, type of equipment such as a motor mounted on train 10, distance between train 10 and a preceding train traveling in front of train 10, time interval between train 10 and preceding train, etc. The type of vehicle may be, for example, whether or not there is a restroom, or whether or not there is a women-only car. The type of train 10 may be, for example, a local train or an express train. Furthermore, the feature for a certain station may include operation data or people flow data for the station that stops just before the station, or operation data or people flow data for the stations that stop two or more before the station. Furthermore, the feature amount for a certain station may include the difference between the operation data or people flow data for that station and the operation data or people flow data for the station that stopped immediately before that station, or the difference between the operation data or people flow data for that station and the operation data or people flow data for the station that stopped two or more stations before that station. Furthermore, the feature amount for a certain train 10 may include the operation data or people flow data for a preceding train.

[0048] 3 is a conceptual diagram showing the relationship between past data, most recent data, and similar data in embodiment 1. The operation prediction device 20 inputs the prediction target data into the prediction model to predict at least one of the future operation status of the train 10 and the future pedestrian flow status of users using the train 10. The prediction target data is operation data and pedestrian flow data acquired after learning by the learning device 24.

[0049] The most recent data is operation data and people flow data acquired over a certain period of time prior to the date on which learning by the learning device 24 is performed. The past data is operation data and people flow data acquired before the date on which acquisition of the most recent operation data and people flow data began. FIG. 3 shows the arrangement of past data, most recent data, and prediction target data in chronological order, with the horizontal axis representing days. The operation prediction device 20 acquires learning data including the past data and most recent data.

[0050] The similarity determination model generation unit 242 generates a similarity determination model by learning from past data and most recent data. The similar data determination unit 243 determines similar data from past data based on the similarity determination model. The operation prediction device 20 obtains learning input data including most recent data and similar data by extracting similar data from past data. The prediction model generation unit 245 generates a prediction model by learning from the learning input data. The white arrows in FIG. 3 indicate that similar data is extracted from past data based on the determination by the similar data determination unit 243.

[0051] In the feature database 23, the learning data is linked to date data indicating the date and time when the learning data was acquired. The data sorting unit 241 sorts the learning data in the feature database 23 into past data and most recent data based on the date data linked to each row of the table in the feature database 23. A period that constitutes the range of the most recent data is set in advance in the data sorting unit 241. The data sorting unit 241 sorts the learning data that is linked to date data included in that period into the most recent data. For example, the user of the traffic management device 30 inputs information indicating the period that constitutes the range of the most recent data into the data sorting unit 241, thereby setting the period that constitutes the range of the most recent data. The period that constitutes the range of the most recent data is set, for example, to a period from the day of learning to several weeks before, or a period from the day of learning to several months before, etc.

[0052] The period of time that is the range of the most recent data can be set to any period. For example, if an event that may affect the flow of users is being held, or if an event that may affect the flow of users is being scheduled, the period during which an event similar to the currently held or scheduled event was held may be set as the period of time that is the range of the most recent data. Also, if an event that may affect the flow of users, such as an infectious disease outbreak, occurs, the period during which an event similar to the currently occurring event occurred may be set as the period of time that is the range of the most recent data. This allows the operation prediction device 20 to include past data acquired when the pedestrian flow situation was similar to the time of prediction in the similar data, thereby enabling the future operation status of the train 10 or the future pedestrian flow situation to be predicted with high accuracy.

[0053] Furthermore, if there are no events that may affect the flow of users and no situations that may affect the flow of people, the period that excludes the period when an event that may affect the flow of people is held and the period when an situation that may affect the flow of people occurs may be set as the period that is the range of the most recent data. In this case, the operation prediction device 20 can also include past data acquired when the flow of people is similar to the time of prediction in the similar data, and therefore can predict the operation status of the train 10 in the future or the flow of people in the future with high accuracy.

[0054] Next, a description will be given of details of the generation of a similarity determination model and the determination of similar data in the operation prediction device 20. Fig. 4 is a first diagram for describing details of the generation of a similarity determination model and the determination of similar data in the operation prediction device 20 according to the first embodiment.

[0055] In the first embodiment, the similarity determination model generation unit 242 and the similar data determination unit 243 use an adversarial validation method to determine similar data from past data. The similarity determination model generation unit 242 divides past data into training data and test data using a method such as cross validation, and performs supervised learning. Here, k-fold cross validation will be used as an example.

[0056] In the example shown in Fig. 4, past data in the feature database 23 is divided into k parts, one of the k parts is set as test data, and the remaining (k-1) parts of the past data are set as training data labeled "0". Furthermore, the most recent data is set as training data labeled "1". These labels are objective variables in learning in the similarity determination model generation unit 242, and are labels indicating whether the data is the most recent data or not. In the following description, divided data refers to each of the k parts obtained by dividing the past data.

[0057] The similarity determination model generation unit 242 generates a similarity determination model by learning from training data. The similar data determination unit 243 inputs test data into the similarity determination model and calculates the similarity of the test data with the most recent data. The similarity is calculated for each row of the table. The similar data determination unit 243 determines, among the test data, data whose calculated similarity is equal to or greater than a preset threshold, as similar data. The similarity determination model generation unit 242 and the similar data determination unit 243 repeat the generation of a similarity determination model and the calculation of similarity k times while sequentially changing the portion of the past data to be used as test data.

[0058] As described above, in the example shown in FIG. 4, the similarity determination model generation unit 242 generates a similarity determination model by learning the first portion, which is a part of the past data, and the most recent data. The similar data determination unit 243 inputs the second portion, which is the portion of the past data excluding the first portion, into the similarity determination model, and calculates the similarity between the second portion of the past data and the most recent data, and determines whether the second portion of the past data corresponds to similar data based on the similarity. The first portion refers to the portion of the past data that is designated as training data labeled "0." The second portion refers to the portion of the past data that is designated as test data. In the example shown in FIG. 4, the similar data is data from the second portion of the past data that is determined to be similar to the most recent data based on the similarity.

[0059] This allows the similarity determination model generation unit 242 and the similar data determination unit 243 to determine similar data from past data. Note that the similar data determination unit 243 may determine similar data by selecting a preset number of data items in descending order of the calculated similarity. The number of data items refers to the number of data items per row in the table of the feature database 23, i.e., the number of data items per station.

[0060] Fig. 5 is a second diagram for explaining details of generation of a similarity determination model and determination of similar data in the operation prediction device 20 according to the first embodiment. The example shown in Fig. 5 is an application example of the example shown in Fig. 4. In the example shown in Fig. 5, not only past data but also most recent data is divided into k pieces of partial data, and the partial data of the most recent data is also included in the test data.

[0061] In the example shown in Fig. 5, similar to the example shown in Fig. 4, past data is divided into k parts, one of the k parts is set as test data, and the remaining (k-1) parts of the past data are set as training data labeled "0." Also, most recent data is divided into k parts, one of the k parts is set as test data, and the remaining (k-1) parts of the most recent data are set as training data labeled "1."

[0062] The similarity determination model generation unit 242 generates a similarity determination model by learning from training data. The similar data determination unit 243 calculates the similarity of the test data with the most recent data by inputting test data into the similarity determination model. The similar data determination unit 243 determines, as similar data, data from divided data of past data included in the test data whose calculated similarity is equal to or greater than a threshold. The similarity determination model generation unit 242 and the similar data determination unit 243 repeat the calculation of the similarity determination model and the similarity k times while sequentially changing the part of the past data to be used as test data and the part of the most recent data to be used as test data.

[0063] 5, the similarity determination model generation unit 242 generates a similarity determination model by learning a first portion that is a part of the past data and a first portion that is a part of the most recent data. The first portion of the past data refers to a portion of the past data that is used as training data labeled "0". The first portion of the most recent data refers to a portion of the most recent data that is used as training data labeled "1".

[0064] In the example shown in FIG. 5 , the similar data determination unit 243 inputs a second portion of the past data excluding the first portion and a second portion of the most recent data excluding the first portion into the similarity determination model, and calculates the similarity between the second portion of the past data and the first portion of the most recent data for each of them, and determines whether the second portion of the past data corresponds to similar data by comparing the similarity calculated for the second portion of the past data with a threshold. The second portion of the past data refers to a portion of the past data that was used as test data. The second portion of the most recent data refers to a portion of the most recent data that was used as test data. In the example shown in FIG. 5 , the similar data is data that is determined to be similar to the most recent data in the second portion of the past data by comparing the similarity calculated for the second portion of the past data with a threshold.

[0065] 5, the similarity determination model generation unit 242 and the similar data determination unit 243 can also determine similar data from past data. Furthermore, in the example shown in Fig. 5, the similarity determination model generation unit 242 and the similar data determination unit 243 can set a threshold value for determining similar data based on the distribution of similarities calculated for a portion of the most recent data that has been set as test data.

[0066] FIG. 6 is a diagram for explaining the threshold value used to determine similar data in the example shown in FIG. 5. FIG. 6 shows an example of a graph plotting the distribution of similarity for a portion of past data that has been designated as test data and the distribution of similarity for a portion of most recent data that has been designated as test data. As shown in FIG. 6, the similar data determination unit 243 sets the threshold value to a similarity that is lower than the similarity calculated for the test data, which is the most recent data. Thus, in the example shown in FIG. 5, the threshold value is set based on the distribution of similarity calculated for the second portion of the most recent data.

[0067] The similarity determination model generation unit 242 generates k similarity determination models by generating similarity determination models while sequentially changing the portion of the past data to be used as test data and the portion of the most recent data to be used as test data. The similar data determination unit 243 sets a threshold for each of the k similarity determination models based on the distribution of similarities calculated for the test data, which is the most recent data.

[0068] When past data is divided into multiple partial data, differences may appear in the feature values ​​for each divided data set that is used as test data among the past data. In the example shown in FIG. 5, a threshold is set based on the distribution of similarities calculated for the test data, which is the most recent data, and the similar data determination unit 243 can determine similar data based on a threshold value according to the trend of the feature values ​​in the test data, which is the past data. Therefore, the similar data determination unit 243 can appropriately extract data from the past data that is similar to the most recent data as similar data. By being able to appropriately extract similar data, the operation prediction device 20 can improve prediction accuracy.

[0069] Fig. 7 is a third diagram for explaining details of the generation of a similarity determination model and the determination of similar data in the operation prediction device 20 according to the first embodiment. The example shown in Fig. 7 is an application example of the example shown in Fig. 4 and the example shown in Fig. 5. In the example shown in Fig. 7, the similar data determination unit 243 performs adjustment by deleting part of the past data whose calculated similarity is equal to or greater than a threshold value.

[0070] The right part of FIG. 7 shows a histogram of the most recent data stored in the feature database 23, i.e., the learning data for which the label indicating whether the data is the most recent is "1." The histogram shown in the right part of FIG. 7 is a bar graph showing the frequency of each calculated similarity value, obtained by calculating the similarity for each row of the most recent data in the feature database 23, i.e., for the data for each acquisition date and time of the most recent data. That is, the most recent data is classified based on the similarity calculated for the data for each acquisition date and time, and the histogram shown in the right part of FIG. 7 shows the frequency distribution of the classified data. Hereinafter, the frequency distribution represented by the histogram shown in the right part of FIG. 7 will be referred to as a first distribution.

[0071] The upper left part of FIG. 7 shows a histogram of past data stored in the feature database 23, i.e., learning data labeled "0" to indicate whether the data is the most recent data. The histogram shown in the upper left part of FIG. 7 shows a bar graph representing the frequency of each calculated similarity value, obtained by calculating the similarity for each row of past data in the feature database 23, i.e., for data acquired at each date and time of acquisition of the past data. In the histogram, solid lines represent the frequency of data whose similarity is equal to or greater than a threshold, and dashed lines represent the frequency of data whose similarity is equal to or less than the threshold. Hereinafter, the frequency distribution represented by the solid line bar graph in the histogram shown in the upper left part of FIG. 7 is referred to as the "second distribution." That is, past data determined to be similar because its similarity is equal to or greater than a threshold is classified based on the similarity of the data acquired at each date and time, and the frequency distribution of the classified data is the second distribution.

[0072] In the example shown in FIG. 7, the similar data determination unit 243 performs an adjustment to bring the second distribution closer to the first distribution by comparing the similarity with a threshold and deleting from the similar data a portion of the past data determined to be similar. The lower left of FIG. 7 shows a histogram after deleting from the similar data a portion of data whose similarity is equal to or greater than the threshold, i.e., a portion of the past data determined to be similar. In the histogram shown in the lower left of FIG. 7, the dashed line portion of the bar graph represents the portion of the past data that was deleted by the adjustment to bring the second distribution closer to the first distribution. The solid line portion of the bar graph represents the portion of the past data that remains after the adjustment. The portion represented by the solid line in the histogram shown in the lower left of FIG. 7 is the same as the histogram shown in the upper left of FIG. 7.

[0073] In the example shown in FIG. 7, the similar data determination unit 243 adjusts the number of data included in the past data so that the frequency distribution of data when the data included in the past data is classified based on similarity matches the frequency distribution of data when the most recent data is classified based on similarity.

[0074] If the frequency of data classified by similarity in the past data exceeds the frequency of data classified by similarity in the most recent data, it can be said that leaving much of that data from the past data in the learning input data will not lead to an improvement in the accuracy of the prediction model. In the example shown in Figure 7, the similar data determination unit 243 can adjust the second distribution to approach the first distribution, so that only data from the past data that is useful for improving the accuracy of the prediction model can be included in the learning input data.

[0075] Next, a description will be given of the learning input data used for learning in the prediction model generation unit 245. Fig. 8 is a second diagram illustrating an example of a table in the feature database 23 held by the operation prediction device 20 according to the first embodiment. The table illustrated in Fig. 8 is the table illustrated in Fig. 2 to which a label indicating whether or not the data is learning input data has been added.

[0076] In the "Learning Input Data" column shown in FIG. 8, the label "0" indicates that the data is not learning input data, and the label "1" indicates that the data is learning input data. In the example shown in FIG. 8, the most recent data, which is learning data acquired on or after June 4, 2022, is included in the learning input data. In addition, of the learning data acquired before June 3, 2022, the data taken at 7:10:00 on June 2, 2022 and the data taken at 7:14:00 on the same day are included in the learning input data as similar data.

[0077] In the example shown in Fig. 8, of the feature quantities, "future delay time" is set as the objective variable in learning by the prediction model generation unit 245. Furthermore, feature quantities other than "future delay time" are set as explanatory variables in learning by the prediction model generation unit 245. The prediction model generation unit 245 builds a delay time prediction model that represents future operation conditions by learning the relationship between the objective variable, which is training data, and the explanatory variables according to a supervised learning algorithm.

[0078] The prediction model generation unit 245 receives learning input data from which past data that does not correspond to similar data has been excluded, thereby enabling the prediction model generation unit 245 to omit learning about unnecessary data. This allows the prediction model generation unit 245 to shorten the learning time. Furthermore, the prediction model generation unit 245 can generate a prediction model that enables highly accurate predictions that match the most recent operation status and the most recent pedestrian flow status. The operation prediction device 20 can highly accurately predict future operation statuses or future pedestrian flow statuses. The operation management device 30 can adjust the operation of the train 10 so as to alleviate congestion among users or to enable efficient operation of the train 10.

[0079] The prediction model generation unit 245 may use only the items useful for predicting operation conditions or only the items useful for predicting people flow conditions, among the feature items used as explanatory variables in the learning by the similarity determination model generation unit 242, as explanatory variables in the learning by the prediction model generation unit 245. In other words, some of the feature items used as explanatory variables in the learning by the similarity determination model generation unit 242 may be excluded from the explanatory variables in the learning by the prediction model generation unit 245. By leaving only the items useful for predicting operation conditions or only the items useful for predicting people flow conditions as feature amounts, the operation prediction device 20 can shorten the learning time in the prediction model generation unit 245 and make highly accurate predictions of operation conditions or people flow conditions.

[0080] In the first embodiment, the similarity determination model generation unit 242 may use an unsupervised learning method instead of supervised learning to learn the most recent data and past data. The similar data determination unit 243 may perform clustering of the past data and determine that data among the past data that belongs to the same cluster as the most recent data is similar data. The similarity determination model generation unit 242 may reduce the number of dimensions of the feature amounts in the feature amount database 23 by a method such as principal component analysis, and perform supervised learning or clustering after reducing the number of dimensions.

[0081] The accuracy of a prediction model may gradually decrease over time due to changes in the characteristics of people flow, etc. Furthermore, the accuracy of a prediction model may also decrease if, after normal times when there are no irregular events that could affect people flow or the occurrence of situations that could affect people flow, an irregular event such as an event being held or an infectious disease outbreak occurs, causing a change in people flow. The accuracy of a prediction model may also decrease if, after an irregular event occurs, the irregular event disappears and normal times return.

[0082] The operation prediction device 20 may monitor the accuracy of the prediction model, and if it determines that the accuracy has decreased, update the learning input data by updating the most recent data and determining similar data based on the updated most recent data, and perform relearning based on the updated learning input data by the prediction model generation unit 245. This enables the operation prediction device 20 to maintain the accuracy of the prediction model used to predict operation conditions or pedestrian flow conditions, and to continue making highly accurate predictions of operation conditions or pedestrian flow conditions.

[0083] The operation prediction device 20, for example, determines whether the most recent data and the prediction target data are similar, and determines that the accuracy of the prediction model has decreased if the number of data determined to be similar among the prediction target data is less than a certain standard. The operation prediction device 20 may check the accuracy of the prediction model at regular intervals, and update the learning input data and re-learn if the accuracy has been determined to have decreased a preset number of times in succession.

[0084] The operation prediction device 20 may calculate an error, which is the difference between the predicted result obtained by using a prediction model and the actual result, for the operation status or pedestrian flow status, and may update the learning input data and perform re-learning if the error exceeds a standard. The operation prediction device 20 may accumulate error values ​​and calculate an average of the accumulated values, and may update the learning input data and perform re-learning if the average exceeds a standard. Furthermore, the operation prediction device 20 may update the learning input data and perform re-learning in advance if it is known in advance that an irregular phenomenon such as an event will occur.

[0085] The operation prediction device 20 may periodically update the most recent data, determine similar data, and perform relearning by the prediction model generation unit 245. The operation prediction device 20 may perform processing related to learning during times when the train 10 is not in operation.

[0086] Next, an operation procedure performed by the operation prediction device 20 according to the first embodiment will be described. An example of the operation procedure performed by the operation prediction device 20 will be described below, divided into a learning phase in which a prediction model is generated and a prediction phase in which an operation status is predicted using the prediction model. FIG. 9 is a flowchart showing an example of the operation procedure performed by the operation prediction device 20 according to the first embodiment during learning. The procedure shown in FIG. 9 is an example of the operation procedure in the learning phase.

[0087] In step S1, the operation prediction device 20 extracts learning data from the equipment data using the data extraction unit 21. The data extraction unit 21 extracts learning data from the equipment data collected from each train 10 by the data collection device 32. The data extraction unit 21 outputs the extracted learning data to the feature extraction unit 22.

[0088] In step S2, the operation prediction device 20 extracts features from the learning data extracted in step S1 using the feature extraction unit 22. The feature extraction unit 22 extracts features by converting the time-series learning data input from the data extraction unit 21 into learning data in units of inter-stations.

[0089] In step S3, the operation prediction device 20 stores the feature amounts extracted in step S2 in the feature amount database 23 by having the feature amount extraction unit 22 write the feature amounts to the feature amount database 23. The feature amount database 23 stores the feature amounts and date data linked to the feature amounts.

[0090] In step S4, the operation prediction device 20 generates a similarity determination model using the similarity determination model generation unit 242, and determines similar data from past data based on the similarity determination model using the similarity data determination unit 243.

[0091] In step S5, the operation prediction device 20 associates the value indicating whether or not the data corresponds to learning input data with the feature by having the feature extraction unit 22 write the value indicating whether or not the data corresponds to learning input data to the feature database 23. For the most recent data and similar data, the similar data determination unit 243 writes "1", which is a value indicating that the data is learning input data, in the "learning input data" column of the feature database 23. For past data other than similar data, the similar data determination unit 243 writes "0", which is a value indicating that the data is not learning input data, in the "learning input data" column of the feature database 23.

[0092] In step S6, the operation prediction device 20 extracts learning input data from the feature database 23 by having the data acquisition unit 244 read out the learning input data from the feature database 23. The data acquisition unit 244 extracts learning input data by reading out learning data whose label indicating whether it is learning input data is "1", that is, most recent data and similar data, from the feature database 23. The data acquisition unit 244 outputs the extracted learning input data to the prediction model generation unit 245.

[0093] In step S7, the prediction model generation unit 245 of the operation prediction device 20 generates a prediction model by learning the learning input data extracted in step S6. The model storage unit 26 stores the prediction model generated by the prediction model generation unit 245. With the above, the operation prediction device 20 ends the operation according to the procedure shown in FIG.

[0094] Fig. 10 is a flowchart showing an example of a detailed procedure for determining similar data in step S4 shown in Fig. 9. Here, the procedure for generating a similarity determination model and determining similar data will be described using the example shown in Fig. 4.

[0095] In step S11, the data classification unit 241 classifies the learning data, which are features stored in the feature database 23 in step S3 shown in Fig. 9, into past data and most recent data. Of the learning data in the feature database 23, the data classification unit 241 determines that the learning data whose date data falls within the period indicating the range of the most recent data is the most recent data. Of the learning data in the feature database 23, the data classification unit 241 determines that the learning data other than the learning data determined to be the most recent data is past data.

[0096] The data classification unit 241 associates the value indicating whether or not the data corresponds to the most recent data with the learning data by writing the value indicating whether or not the data corresponds to the most recent data to the feature database 23. For the learning data classified as the most recent data, the data classification unit 241 writes "1", which is a value indicating that the data is the most recent data, in the "most recent data" column of the feature database 23. For the learning data classified as the past data, the data classification unit 241 writes "0", which is a value indicating that the data is not the most recent data, in the "most recent data" column of the feature database 23.

[0097] In step S12, the similarity determination model generation unit 242 reads past data having a value of "0" in the "most recent data" column from the feature database 23, and divides the past data into k divided data. The similarity determination model generation unit 242 also reads most recent data having a value of "1" in the "most recent data" column from the feature database 23.

[0098] In step S13, the similarity determination model generation unit 242 generates a similarity determination model by using one divided data as test data and learning the most recent data and a portion of the past data excluding the test data. The similarity determination model generation unit 242 sets one of the k divided data as test data, and sets the remaining (k-1) divided data of the past data as training data labeled "0". The similarity determination model generation unit 242 also sets the most recent data read from the feature database 23 as training data labeled "1". The similarity determination model generation unit 242 generates a similarity determination model by learning the training data.

[0099] In step S14, the similar data determination unit 243 inputs the test data into the similarity determination model to determine the similarity of the test data with the most recent data. The similarity determination model generation unit 242 and the similar data determination unit 243 execute steps S13 and S14 when all divided data is used as test data. That is, the similarity determination model generation unit 242 and the similar data determination unit 243 repeat the generation of the similarity determination model and the calculation of the similarity k times while sequentially changing the part of the past data that is used as test data.

[0100] In step S15, the similar data determination unit 243 determines that the past data having a similarity equal to or greater than a threshold value is similar data. After completing step S15, the operation prediction device 20 proceeds to step S5 shown in FIG.

[0101] 11 is a flowchart showing an example of an operation procedure at the time of prediction by the operation prediction device 20 according to the first embodiment. The procedure shown in FIG. 11 is an example of an operation procedure in the inference phase.

[0102] In step S21, the operation prediction device 20 extracts prediction target data from the equipment data using the data extraction unit 21. The data extraction unit 21 outputs the extracted prediction target data to the feature extraction unit 22.

[0103] In step S22, the operation prediction device 20 extracts features from the prediction target data extracted in step S21 using the feature extraction unit 22. The feature extraction unit 22 extracts features by converting the time-series prediction target data input from the data extraction unit 21 into prediction target data in units of inter-stations.

[0104] In step S23, the operation prediction device 20 stores the feature amounts extracted in step S22 in the feature amount database 23 by causing the feature amount extraction unit 22 to write the feature amounts in the feature amount database 23.

[0105] The data acquisition unit 251 of the inference device 25 acquires prediction target data, which is data for inference, by reading prediction target data to be used for predicting at least one of the operation status and the pedestrian flow status from the feature database 23. The data acquisition unit 251 outputs the prediction target data to the inference unit 252. The inference unit 252 reads the prediction model from the model storage unit 26. In step S24, the operation prediction device 20 inputs the feature into the prediction model in the inference unit 252, thereby obtaining a prediction result of the future operation status or future pedestrian flow status by the inference unit 252. The prediction result output by the inference unit 252 is stored in the feature database 23. With this, the operation prediction device 20 ends the operation according to the procedure shown in FIG. 11 .

[0106] 1 shows an example in which the traffic management device 30 is provided with the traffic prediction device 20 and the data collection device 32, but the configuration of the traffic management system 100 is not limited to the configuration shown in Fig. 1. Modified examples of the traffic management system 100 will be described below.

[0107] 12 is a diagram illustrating an example of the configuration of a traffic management system 100A according to a first modified example of the first embodiment. The traffic management system 100A includes a plurality of trains 10, a traffic management device 30A that manages the operation of each train 10, a diagram creation device 40, and a server device 50. The traffic management device 30 includes an operation prediction device 20A that predicts the operation status of the train 10. The configuration of the operation prediction device 20A is similar to the configuration of the operation prediction device 20 shown in FIG. 1. The traffic management device 30A differs from the traffic management device 30 shown in FIG. 1 in that it does not include a data collection device 32.

[0108] The server device 50 is a server constructed in a cloud environment. The server device 50 includes a data collection device 51 and a communication unit 52 that communicates with devices external to the server device 50. The data collection device 51 collects equipment data, similar to the data collection device 32 shown in FIG. 1. The data collection device 51 outputs the collected equipment data to the operation prediction device 20A. The communication unit 52 communicates with the communication unit 11 of each train 10 and with the operation management device 30A.

[0109] The operation management device 30 of the operation management system 100 shown in Figure 1 acquires equipment data directly from each train 10, but the operation management device 30A of the operation management system 100A shown in Figure 12 acquires equipment data via a data collection device 51 provided in the server device 50.

[0110] 13 is a diagram illustrating a configuration example of a traffic control system 100B according to a second modified example of the embodiment 1. The traffic control system 100B includes a plurality of trains 10, a traffic control device 30B, a diagram creation device 40, and a server device 50B.

[0111] The server device 50B is a server constructed in a cloud environment. Like the server device 50 shown in FIG. 12, the server device 50B includes a data collection device 51 and a communication unit 52. The communication unit 52 communicates with the communication unit 11 of each train 10 and with the traffic management device 30B. The server device 50B further includes an operation prediction device 20B that predicts the operation status of the train 10. The operation prediction device 20B has the same configuration as the operation prediction device 20 shown in FIG. 1. The operation prediction device 20B also includes a storage unit 34a that stores a bus schedule database 34.

[0112] 1, the traffic management device 30B includes a communication unit 31 and a countermeasure processing unit 33. The traffic management device 30B differs from the traffic management device 30 shown in FIG. 1 in that the traffic management device 30B does not include the traffic prediction device 20, the data collection device 32, and the storage unit 34a.

[0113] The operation prediction device 20 of the operation management system 100 shown in Figure 1 is provided inside the operation management device 30, but the operation prediction device 20B of the operation management system 100B shown in Figure 13 is provided in a server device 50B together with a data collection device 51.

[0114] The inference unit 252 of the inference device 25 outputs the prediction result to the communication unit 52. The communication unit 52 transmits the prediction result to the communication unit 31 of the traffic management device 30B. The communication unit 31 outputs the received prediction result to the countermeasure processing unit 33. The countermeasure processing unit 33 outputs an instruction to the communication unit 31 to execute the formulated countermeasure.

[0115] In the traffic management system 100 shown in FIG. 1, the learning device 24 is a device built into the traffic prediction device 20. The learning device 24 may be a device external to the traffic prediction device 20 or a device external to the traffic management device 30. The learning device 24, which is a device external to the traffic management device 30, is a device that constitutes the traffic management system 100. The learning device 24 may be a device that can be connected to the traffic prediction device 20 via a network. The learning device 24 may be a device that exists on a cloud server. In the traffic management system 100A shown in FIG. 12, the learning device 24 may also be a device external to the traffic prediction device 20A or a device external to the traffic management device 30A. In the traffic management system 100B shown in FIG. 13, the learning device 24 may also be a device external to the traffic prediction device 20B or a device external to the server device 50B.

[0116] According to the first embodiment, the operation prediction devices 20, 20A, and 20B generate a similarity determination model by learning from past data and most recent data, and determine similar data from the past data based on the similarity determination model. By using the similarity determination model generated by learning from the past data and most recent data, the operation prediction devices 20, 20A, and 20B can extract similar data from past data even when the feature vectors contain many items that may affect the operation status. Furthermore, even when the feature vectors contain many items with different degrees of influence on the determination of similarity, the operation prediction devices 20, 20A, and 20B can reflect the degree of influence of each item in the determination of similarity. Furthermore, the operation prediction devices 20, 20A, and 20B can predict the operation status of the train 10 using data from past data that has similar operation status or pedestrian flow status, thereby enabling highly accurate prediction of the operation status. As a result, the operation prediction devices 20, 20A, 20B, the operation management devices 30, 30A, 30B, and the operation management systems 100, 100A, 100B have the advantage of being able to predict the future operation status of the train 10 with high accuracy.

[0117] Embodiment 2 In the first embodiment, the operation prediction devices 20, 20A, and 20B determine similar data based on feature amounts that are learning data for each station. In the second embodiment, an example will be described in which similar data is determined based on feature amounts for each arbitrary period.

[0118] FIG. 14 is a diagram showing an example of the configuration of a traffic control system 100C according to the second embodiment. In the second embodiment, the same components as those in the first embodiment are given the same reference numerals, and the configuration different from the first embodiment will be mainly described. The traffic control system 100C is a system that manages the operation of trains 10 on a railway. The traffic control system 100C includes a plurality of trains 10, a traffic control device 30C that manages the operation of each train 10, and a diagram creation device 40 that creates a train diagram for each train 10.

[0119] The traffic management device 30C differs from the traffic management device 30 shown in FIG. 1 in that it includes a traffic prediction device 20C having a different configuration from the traffic prediction device 20. The traffic prediction device 20C differs from the traffic prediction device 20 shown in FIG. 1 in that it includes a learning device 24C having a different configuration from the learning device 24. Furthermore, the traffic prediction device 20C includes a storage unit 27a that stores a first feature database 27 instead of the storage unit 23a that stores the feature database 23. The first feature database 27 is a feature database for operation prediction that stores first feature values ​​that are feature values ​​used in predicting the operation status. The first feature database 27 is a database similar to the feature database 23.

[0120] The learning device 24C includes a data classification unit 241, a similarity determination model generation unit 242C, a similar data determination unit 243C, a data acquisition unit 244, a prediction model generation unit 245, a feature conversion unit 246, and a memory unit 247a that holds a second feature database 247.

[0121] The feature conversion unit 246 acquires first feature amounts by reading out the first feature amounts, which are training data for each station, from the first feature database 27. The feature conversion unit 246 converts the acquired first feature amounts into second feature amounts that are based on a set period. In the second embodiment, the second feature amounts are based on days. The second feature amounts are not limited to day-based feature amounts, and may be week-based or month-based feature amounts. Any period can be set as the period that is used as the unit for the second feature amounts.

[0122] The second feature amount is stored in a second feature amount database 247. The second feature amount database 247 is a feature amount database for similarity determination that stores second feature amounts, which are feature amounts used in determining similar data. Each row of the table in the second feature amount database 247 is associated with date data and a label indicating whether the data is the most recent data. The second feature amount database 247 is similar to the feature amount database 23 shown in FIG. 1 except that the second feature amount database 247 stores second feature amounts, which are data on a daily basis.

[0123] The similarity determination model generation unit 242C acquires the past data that is the second feature amount and the most recent data that is the second feature amount by reading the past data that is the second feature amount and the most recent data that is the second feature amount from the second feature amount database 247. The similarity determination model generation unit 242C generates a similarity determination model by learning the past data that is the second feature amount and the most recent data that is the second feature amount.

[0124] The similar data determination unit 243C acquires past data that is the second feature by reading the past data that is the second feature from the second feature database 247. The similar data determination unit 243C determines similar data from the past data that is the second feature based on the similarity determination model. The similar data determination unit 243C outputs the determination result. For example, the similar data determination unit 243C outputs a label indicating whether or not the data is similar as the determination result. The determination result by the similar data determination unit 243C is reflected in a label in the first feature database 27 that indicates whether or not the data is learning input data.

[0125] Fig. 15 is a diagram for explaining conversion from a first feature to a second feature in the feature conversion unit 246 of the operation prediction device 20C according to the second embodiment. The left part of Fig. 15 shows an example of a part of date data in a table in the first feature database 27. The right part of Fig. 15 shows an example of a part of date data in a table in the second feature database 247. The hollow arrows in Fig. 15 indicate that a first feature is converted into a second feature.

[0126] The tables in the first feature database 27 are similar to those shown in FIGS. 2 and 8. In the tables in the first feature database 27, each first feature is shown in one row. In the second feature database 247, each first feature linked to the same date data in the first feature database 27 is grouped into a second feature, which is learning data for each day. In the tables in the second feature database 247, each second feature is shown in one row. Note that the "day" that is the unit of the second feature is not a period separated by midnight, but a period from the start of operation of the first train 10 to the end of operation of the last train 10.

[0127] In the second embodiment, the operation prediction device 20C determines similar data from past data based on a second feature value that is calculated for a period longer than the period between stations. Events such as the holding of an event or the spread of an infectious disease may affect the operation of the train 10 or the flow of passengers on a daily basis or for periods longer than one day. The operation prediction device 20C takes such effects into account to determine similar data from past data.

[0128] Next, examples of the second feature amount will be described. Here, two specific examples of the second feature amount will be described. FIG. 16 is a diagram for explaining a first example of the second feature amount used to determine similar data in the operation prediction device 20C according to the second embodiment. The first example is an example in which learning data on a station-to-station basis is converted into data groups for each station and for each time period, and the frequency of data in each data group is used as learning data on a daily basis.

[0129] In the first example shown in FIG. 16, passenger occupancy rate fluctuations are classified into three categories—“small,” “medium,” and “large”—based on the magnitude of their values, and the frequency of data classified into “small,” “medium,” and “large” for each station and time period is used as the second feature. In the first example shown in FIG. 16, data is classified into “small,” “medium,” and “large” for each station and time period. FIG. 16 shows a histogram representing the frequency of data for each station and time period using a bar graph. FIG. 16 shows an example of the distribution of frequencies calculated for passenger occupancy rate fluctuations at “Station A” between 8:00 a.m. and 9:00 a.m. The feature conversion unit 246 converts the first feature into the second feature by classifying the learning data for each day and calculating the frequency. Note that the frequency of data refers to the number of data items per row in the first feature database 27.

[0130] In the first example, the second feature is a feature that takes into account the number of users for each station and each time period, which allows the similar data determination unit 243C to determine similar data taking into account the number of users who practice staggered work hours, etc.

[0131] 17 is a first diagram for explaining a second example of the second feature used to determine similar data in the operation prediction device 20C according to the second embodiment. The second example is an example in which the learning data on a station-to-station basis is divided into a plurality of clusters based on the features of the values ​​of a plurality of items in the first feature, and the frequency of data for each cluster is used as the learning data on a daily basis.

[0132] In the second example shown in FIG. 17, two or more items, such as passenger load fluctuation and delay time, are each classified into "small," "medium," "large," etc., based on the classification that represents the characteristics of the values ​​of each item. Then, based on this classification that represents the characteristics of the values ​​of each item, the training data for each station is divided into multiple clusters, such as cluster number "1" for data with "large" passenger load fluctuation and "large" delay time, and cluster number "2" for data with "medium" passenger load fluctuation and "small" delay time. Furthermore, the frequency of data for each cluster is used as a second feature. In the second example shown in FIG. 17, the training data is divided into six clusters with cluster numbers "1" to "6." FIG. 17 shows a histogram that represents the frequency of data for each cluster using a bar graph.

[0133] Fig. 18 is a second diagram for explaining a second example of the second feature used to determine similar data in the operation prediction device 20C according to the second embodiment. Fig. 18 shows an example of a table in the second feature database 247. In the table in the second feature database 247, the frequency of each cluster, which is the second feature, is shown in each row of the table. In the second feature database 247, the second feature, which is learning data for each day, is organized in each row of the table.

[0134] The feature conversion unit 246 converts the first feature into the second feature by performing such clustering and frequency calculation on the training data for each day. Note that the feature conversion unit 246 may reduce the number of dimensions of the first feature using a technique such as principal component analysis, and then perform clustering. The feature conversion unit 246 may also convert the first feature into the second feature by combining the first example shown in FIG. 16 and the second example shown in FIG. 17. The feature conversion unit 246 may cluster the training data by time period or station. The feature conversion unit 246 can classify values ​​for any of the items in the training data and perform clustering.

[0135] Next, an operation of the operation prediction device 20C according to the second embodiment will be described. Here, an operation of determining similar data, which is one of the operations during learning to generate a prediction model, will be described. Fig. 19 is a flowchart showing an example of the procedure of the operation when determining similar data by the operation prediction device 20C according to the second embodiment.

[0136] In step S31, the data classification unit 241 classifies the learning data, which is the first feature stored in the first feature database 27, into past data and most recent data. The data classification unit 241 writes a value indicating whether or not the data corresponds to the most recent data to the first feature database 27, thereby linking the value indicating whether or not the data corresponds to the most recent data to the learning data.

[0137] The feature conversion unit 246 acquires learning data for each inter-station by reading the learning data for each inter-station from the first feature database 27. In step S32, the feature conversion unit 246 converts the feature for each inter-station into a feature for each day, and stores the second feature, which is a feature for each day, in the second feature database 247.

[0138] In step S33, the similarity determination model generation unit 242C reads past data having a value of "0" in the "most recent data" column from the second feature amount database 247, and divides the past data into k divided data. The similarity determination model generation unit 242C also reads most recent data having a value of "1" in the "most recent data" column from the second feature amount database 247.

[0139] In step S34, the similarity determination model generation unit 242C generates a similarity determination model by using one divided data as test data and learning a part of the past data excluding the test data and the most recent data. The similarity determination model generation unit 242C generates a similarity determination model in the same way as in step S13 shown in FIG.

[0140] In step S35, the similar data determination unit 243C inputs the test data into the similarity determination model to determine the similarity between the test data and the most recent data. The similar data determination unit 243C determines the similarity in the same way as in step S14 shown in FIG.

[0141] The similarity determination model generation unit 242C and the similar data determination unit 243C execute steps S34 and S35 when all divided data are used as test data. That is, the similarity determination model generation unit 242C and the similar data determination unit 243C repeat the generation of a similarity determination model and the calculation of similarity k times while sequentially changing the part of the past data that is used as test data.

[0142] In step S36, the similar data determination unit 243C determines that the past data having a similarity equal to or greater than a threshold value is similar data. After completing step S36, the operation prediction device 20C proceeds with the same procedures as those from step S5 onwards shown in FIG.

[0143] Fig. 20 is a flowchart showing an example of detailed procedures for converting feature amounts in step S32 shown in Fig. 19. Here, an operation will be described in which the feature amount converting unit 246 converts a first feature amount into a second feature amount according to the second example shown in Fig. 17.

[0144] In step S41, the feature conversion unit 246 clusters the features between stations. The feature conversion unit 246 divides the training data into multiple clusters based on the features of the values ​​of two or more items in the training data. In step S42, the feature conversion unit 246 calculates the frequency of each cluster in the training data for one day.

[0145] In step S43, the feature converter 246 generates daily features for the one day by obtaining a histogram representing the frequency for each cluster. In step S44, the feature converter 246 stores the generated daily features in the second feature database 247. The feature converter 246 executes steps S42 to S44 for all days linked to the past data.

[0146] In step S45, the feature conversion unit 246 links the value indicating whether the data corresponds to the most recent data to the daily feature stored in the second feature database 247. After completing step S45, the operation prediction device 20C proceeds to step S33 shown in FIG.

[0147] According to the second embodiment, the operation prediction device 20C generates a similarity determination model by learning from past data and the most recent data, and determines similar data from the past data based on the similarity determination model. The operation prediction device 20C can extract similar data from past data even when the feature includes many items that may affect the operation status. Furthermore, the operation prediction device 20C can arbitrarily set the period used as the unit of the second feature depending on the period of time during which the operation of the train 10 or the flow of passengers will be affected, and can determine similar data taking into account the impact on the operation of the train 10 or the flow of passengers. This allows the operation prediction device 20C to determine similar data with high accuracy. As a result, the operation prediction device 20C, the operation management device 30C, and the traffic management system 100C achieve the effect of being able to predict the future operation status of the train 10 with high accuracy.

[0148] Next, a hardware configuration of the operation prediction devices 20, 20A, 20B, and 20C according to the first or second embodiment will be described. FIG. 21 is a diagram illustrating an example of a hardware configuration of the operation prediction devices 20, 20A, 20B, and 20C constituting the operation control systems 100, 100A, 100B, and 100C according to the first or second embodiment. The operation prediction devices 20, 20A, 20B, and 20C are realized by a computer system including a processing circuit 80, an input unit 81, and an output unit 84. The processing circuit 80 includes a processor 82 and a memory 83. The processing circuit 80 is a circuit on which the processor 82 executes software.

[0149] The data extraction unit 21, feature extraction unit 22, learning device 24, 24C, and inference device 25 of the operation prediction devices 20, 20A, 20B, and 20C are each realized by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in memory 83. In the processing circuit 80, the processor 82 reads and executes the program stored in memory 83, thereby realizing the functions of each of the operation prediction devices 20, 20A, 20B, and 20C. That is, the processing circuit 80 includes a memory 83 for storing a program that results in the processing of the operation prediction devices 20, 20A, 20B, and 20C. The program stored in memory 83 is a traffic management program that causes a computer to execute the procedures and methods of the traffic management systems 100, 100A, 100B, and 100C.

[0150] The processor 82 is a CPU (Central Processing Unit), processing device, arithmetic device, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor). The memory 83 is, for example, a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), or EEPROM (Electrically Erasable Programmable Read Only Memory). The input unit 81 is a device for inputting information to the operation prediction devices 20, 20A, 20B, and 20C. The output unit 84 outputs information to the outside of the operation prediction devices 20, 20A, 20B, and 20C.

[0151] The data collection device 32 and the countermeasure processing unit 33 of the traffic management device 30 shown in Fig. 1, the countermeasure processing unit 33 of the traffic management device 30A shown in Fig. 12, and the countermeasure processing unit 33 of the traffic management device 30B shown in Fig. 13 are realized by a configuration similar to that of the processing circuit 80 shown in Fig. 21. The hardware configuration of the traffic management devices 30, 30A, 30B, and 30C includes a communication device for performing communication.

[0152] The learning device 24 may be realized by the processor 82 reading and executing a learning program stored in the memory 83. The learning program is a program that can be executed by a computer and is a program for executing the operations of the learning device 24. It can also be said that the learning program causes a computer to execute the procedures and methods of the learning device 24.

[0153] The server device 50 shown in Fig. 12 and the data collection device 51 of the server device 50B shown in Fig. 13 are realized by a configuration similar to that of the processing circuit 80 shown in Fig. 21. The hardware configuration of the server devices 50 and 50B includes a communication device for performing communication. The diagram creation device 40 of the traffic control system 100, 100A, 100B, 100C according to the first or second embodiment is realized by a hardware configuration similar to that shown in Fig. 21.

[0154] The configurations shown in the above embodiments are examples of the contents of the present disclosure. The configurations of each embodiment can be combined with other known technologies. The configurations of each embodiment can also be combined as appropriate. Part of the configuration of each embodiment can be omitted or modified without departing from the gist of the present disclosure.

[0155] Various aspects of the present disclosure are summarized below as appendices.

[0156] (Appendix 1) a data acquisition unit that acquires, from the train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the situation of people flow of users using the train; a prediction unit that divides learning data including the operation data and the people flow data into past data and most recent data acquired after the past data, and predicts at least one of the operation status and the people flow status in the future from the prediction target data using a prediction model that has been trained using similar data from the past data that is similar to the most recent data and the most recent data; Equipped with The similar data is data of the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data. An operation prediction device characterized by: (Appendix 2) the similarity determination model is generated by learning a first portion that is a part of the past data and the most recent data; a second portion of the past data excluding the first portion is input to the similarity determination model, thereby determining a similarity between the second portion of the past data and the most recent data; The similar data is data in the second portion of the past data that is determined to be similar to the most recent data based on the similarity. 2. An operation prediction device according to claim 1, (Appendix 3) the similarity determination model is generated by learning a first portion that is a part of the past data and a first portion that is a part of the most recent data; a second portion of the past data excluding the first portion and a second portion of the most recent data excluding the first portion are input to the similarity determination model, thereby determining a similarity between the second portion of the past data and the first portion of the most recent data and the most recent data; the similar data is data in the second portion of the past data that is determined to be similar to the most recent data by comparing the similarity calculated for the second portion of the past data with a threshold value; 2. An operation prediction device according to claim 1, (Appendix 4) the threshold is set based on the distribution of the similarity calculated for the second portion of the most recent data. 4. An operation prediction device according to claim 3. (Appendix 5) classifying the most recent data based on the similarity calculated for data for each acquired date and time, and defining a frequency distribution for the classified data as a first distribution; classifying the past data determined to be similar data based on the similarity calculated for data for each acquired date and time, and defining a frequency distribution for the classified data as a second distribution; and adjusting the second distribution to approach the first distribution by deleting a portion of the past data determined to be similar data from the similar data. 5. An operation prediction device according to any one of appendices 2 to 4. (Appendix 6) The learning data is linked to data on the date and time when the learning data was acquired, and the range of the most recent data is set by specifying a date and time range, and the past data and the most recent data are distinguished based on the date and time when the learning data was acquired. 6. An operation prediction device according to any one of appendices 1 to 5. (Appendix 7) a data extraction unit that extracts the prediction target data from equipment data that is data output from the train; a feature extraction unit that converts the prediction target data into data for each unit, using a period from when the train departs from a first station to when the train arrives at a second station that is the next stop and departs from the second station, or a period from when the train arrives at the first station to when the train departs from the first station and arrives at a second station that is the next stop, as a unit, to extract features that are the prediction target data for each station; Equipped with a target variable representing a result of separating the past data and the most recent data input to the similarity determination model is associated with the feature amount, for each of the past data and the most recent data input to the similarity determination model; 7. An operation prediction device according to any one of claims 1 to 6. (Appendix 8) a feature conversion unit that converts the first feature, which is the feature in units of stations, into a second feature in units of a set period; Equipped with the similarity determination model is generated by learning the past data, which is the second feature amount, and the most recent data, which is the second feature amount; The similar data is determined from the past data, which is the second feature amount. 8. An operation prediction device according to claim 7, (Appendix 9) An operation prediction device according to appendix 1; a countermeasure processing unit that formulates countermeasures according to the prediction results output by the operation prediction device; Equipped with the operation prediction device outputs the prediction result of at least one of the future operation status and the future pedestrian flow status to the countermeasure processing unit; An operation management device characterized by: (Appendix 10) a communication unit that receives equipment data output from the train; a data collection device that collects the received equipment data and outputs the collected equipment data to the operation prediction device; Equipped with the operation prediction device includes: a data extraction unit that extracts the prediction target data from the equipment data; 10. The operation management device according to claim 9, comprising: (Appendix 11) An operation management device according to Supplementary Note 9; a train schedule creation device that creates a train schedule; Equipped with the traffic management device outputs information on the measures according to the prediction result of at least one of the future traffic status and the future pedestrian flow status to the diagram creation device; the diagram creation device creates the bus diagram based on the information on the measures; A traffic management system characterized by: (Appendix 12) An operation management device according to Supplementary Note 9; a data collection device that collects equipment data output from the train and outputs the collected equipment data to the traffic management device; An operation management system comprising: (Appendix 13) An operation prediction device according to appendix 1; a data collection device that collects equipment data output from the train and outputs the collected equipment data to the operation prediction device; an operation management device that formulates measures according to the prediction results output by the operation prediction device; Equipped with the operation prediction device outputs the prediction result of at least one of the future operation status and the future pedestrian flow status to the operation management device; A traffic management system characterized by: (Appendix 14) a similarity determination model generation unit that divides learning data, which includes operation data indicating train operation status and people flow data indicating the people flow status of users using the train, into past data and most recent data acquired after the past data, and generates a similarity determination model that determines similar data in the past data that is similar to the most recent data by learning the past data and the most recent data; a similar data determination unit that determines the similar data from the past data based on the similarity determination model; a prediction model generation unit that generates a prediction model that predicts at least one of the operation status and the pedestrian flow status in the future by learning the most recent data and the similar data; A learning device comprising: (Appendix 15) acquiring, from the train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the situation of people flow of users using the train; a step of dividing learning data including the operation data and the people flow data into past data and most recent data acquired after the past data, and predicting at least one of the operation status and the people flow status in the future from the prediction target data using a prediction model that has been trained on similar data from the past data that is similar to the most recent data and the most recent data; Including, The similar data is data of the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data. An operation prediction method characterized by: (Appendix 16) A step of dividing learning data including operation data indicating the operation status of a train and people flow data indicating the situation of people flow of users using the train into past data and most recent data acquired after the past data; generating a similarity determination model that determines similar data of the past data that is similar to the most recent data by learning the past data and the most recent data; determining the similar data from the past data based on the similarity determination model; generating a prediction model that predicts at least one of the operation status and the pedestrian flow status in the future by learning from the most recent data and the similar data; A learning method comprising: (Appendix 17) acquiring, from the train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the situation of people flow of users using the train; a step of dividing learning data including the operation data and the people flow data into past data and most recent data acquired after the past data, and predicting at least one of the operation status and the people flow status in the future from the prediction target data using a prediction model that has been trained on similar data from the past data that is similar to the most recent data and the most recent data; An operation prediction program that causes a computer to execute The similar data is data of the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data. An operation prediction program characterized by: (Appendix 18) A step of dividing learning data including operation data indicating the operation status of a train and people flow data indicating the situation of people flow of users using the train into past data and most recent data acquired after the past data; generating a similarity determination model that determines similar data of the past data that is similar to the most recent data by learning the past data and the most recent data; determining the similar data from the past data based on the similarity determination model; generating a prediction model that predicts at least one of the operation status and the pedestrian flow status in the future by learning from the most recent data and the similar data; A learning program characterized by causing a computer to execute the above. [Explanation of symbols]

[0157] 10 Train, 11, 31, 52 Communication unit, 20, 20A, 20B, 20C Operation prediction device, 21 Data extraction unit, 22 Feature extraction unit, 23 Feature database, 23a, 27a, 34a, 247a Memory unit, 24, 24C Learning device, 25 Inference device, 26 Model memory unit, 27 First feature database, 30, 30A, 30B, 30C Operation management device, 32, 51 Data collection device, 33 Countermeasure processing unit, 34 Operation diagram database, 40 Operation diagram creation device, 41 Operation diagram creation unit, 50, 50B Server device, 80 Processing circuit, 81 Input unit, 82 Processor, 83 Memory, 84 Output unit, 100, 100A, 100B, 100C Operation management system, 241 Data sorting unit, 242, 242C Similarity determination model generation unit, 243, 243C similar data determination unit, 244, 251 data acquisition unit, 245 prediction model generation unit, 246 feature conversion unit, 247 second feature database, 252 inference unit.

Claims

1. a data acquisition unit that acquires, from the train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the situation of people flow of users using the train; a prediction unit that divides learning data including the operation data and the people flow data into past data and most recent data acquired after the past data, and predicts at least one of the operation status and the people flow status in the future from the prediction target data using a prediction model that has been trained using similar data from the past data that is similar to the most recent data and the most recent data; Equipped with The similar data is data of the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data. An operation prediction device characterized by:

2. the similarity determination model is generated by learning a first portion that is a part of the past data and the most recent data; a second portion of the past data excluding the first portion is input to the similarity determination model, thereby determining a similarity between the second portion of the past data and the most recent data; the similar data is data in the second portion of the past data that is determined to be similar to the most recent data based on the similarity.

2. The operation prediction device according to claim 1 .

3. the similarity determination model is generated by learning a first portion that is a part of the past data and a first portion that is a part of the most recent data; a second portion of the past data excluding the first portion and a second portion of the most recent data excluding the first portion are input to the similarity determination model, thereby determining a similarity between the second portion of the past data and the first portion of the most recent data and the most recent data; the similar data is data in the second portion of the past data that is determined to be similar to the most recent data by comparing the similarity calculated for the second portion of the past data with a threshold value; 2. The operation prediction device according to claim 1 .

4. the threshold is set based on the distribution of the similarity calculated for the second portion of the most recent data.

4. The operation prediction device according to claim 3.

5. classifying the most recent data based on the similarity calculated for data for each acquired date and time, and defining a frequency distribution for the classified data as a first distribution; classifying the past data determined to be similar data based on the similarity calculated for data for each acquired date and time, and defining a frequency distribution for the classified data as a second distribution; and adjusting the second distribution to approach the first distribution by deleting a portion of the past data determined to be similar data from the similar data.

5. The operation prediction device according to claim 2, wherein the operation prediction device is a device for predicting a traffic flow.

6. The learning data is linked to data on the date and time when the learning data was acquired, and the range of the most recent data is set by specifying a date and time range, and the past data and the most recent data are distinguished based on the date and time when the learning data was acquired.

5. The operation prediction device according to claim 1, wherein:

7. a data extraction unit that extracts the prediction target data from equipment data that is data output from the train; a feature extraction unit that converts the prediction target data into data for each unit, using a period from when the train departs from a first station to when the train arrives at a second station, which is the next stop, and departs from the second station, or a period from when the train arrives at the first station to when the train departs from the first station and arrives at the second station, which is the next stop, as a unit, to extract features that are the prediction target data for each station; Equipped with a target variable representing a result of separating the past data and the most recent data input to the similarity determination model is associated with the feature amount, for each of the past data and the most recent data input to the similarity determination model; 5. The operation prediction device according to claim 1, wherein:

8. a feature conversion unit that converts the first feature, which is the feature in units of stations, into a second feature in units of a set period; Equipped with the similarity determination model is generated by learning the past data, which is the second feature amount, and the most recent data, which is the second feature amount; the similar data is determined from the past data, which is the second feature amount; 8. The operation prediction device according to claim 7.

9. The operation prediction device according to claim 1 ; a countermeasure processing unit that formulates countermeasures according to the prediction results output by the operation prediction device; Equipped with the operation prediction device outputs the prediction result of at least one of the future operation status and the future pedestrian flow status to the countermeasure processing unit; An operation management device characterized by:

10. a communication unit that receives equipment data output from the train; a data collection device that collects the received equipment data and outputs the collected equipment data to the operation prediction device; Equipped with the operation prediction device includes: a data extraction unit that extracts the prediction target data from the equipment data; The operation management device according to claim 9, further comprising:

11. The operation management device according to claim 9 ; a train schedule creation device that creates a train schedule; Equipped with the traffic management device outputs information on the measures according to the prediction result of at least one of the future traffic status and the future pedestrian flow status to the diagram creation device; the diagram creation device creates the bus diagram based on the information on the measures; A traffic management system characterized by:

12. The operation management device according to claim 9 ; a data collection device that collects equipment data output from the train and outputs the collected equipment data to the traffic management device; An operation management system comprising:

13. The operation prediction device according to claim 1 ; a data collection device that collects equipment data output from the train and outputs the collected equipment data to the operation prediction device; an operation management device that formulates measures according to the prediction results output by the operation prediction device; Equipped with the operation prediction device outputs the prediction result of at least one of the future operation status and the future pedestrian flow status to the operation management device; A traffic management system characterized by:

14. a similarity determination model generation unit that divides learning data, which includes operation data indicating the operation status of trains and people flow data indicating the flow of people using the trains, into past data and most recent data acquired after the past data, and generates a similarity determination model that determines similar data in the past data that is similar to the most recent data by learning the past data and the most recent data; a similar data determination unit that determines the similar data from the past data based on the similarity determination model; a prediction model generation unit that generates a prediction model that predicts at least one of the operation status and the pedestrian flow status in the future by learning the most recent data and the similar data; A learning device comprising:

15. acquiring, from the train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the situation of people flow of users using the train; a step of dividing learning data including the operation data and the people flow data into past data and most recent data acquired after the past data, and predicting at least one of the operation status and the people flow status in the future from the prediction target data using a prediction model that has been trained on similar data from the past data that is similar to the most recent data and the most recent data; Including, The similar data is data of the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data. An operation prediction method characterized by:

16. A step of dividing learning data including operation data indicating the operation status of a train and people flow data indicating the situation of people flow of users using the train into past data and most recent data acquired after the past data; generating a similarity determination model that determines similar data of the past data that is similar to the most recent data by learning the past data and the most recent data; determining the similar data from the past data based on the similarity determination model; generating a prediction model that predicts at least one of the operation status and the pedestrian flow status in the future by learning from the most recent data and the similar data; A learning method comprising:

17. acquiring, from the train, prediction target data including operation data indicating the operation status of the train and people flow data indicating the situation of people flow of users using the train; a step of dividing learning data including the operation data and the people flow data into past data and most recent data acquired after the past data, and predicting at least one of the operation status and the people flow status in the future from the prediction target data using a prediction model that has been trained on similar data from the past data that is similar to the most recent data and the most recent data; An operation prediction program that causes a computer to execute The similar data is data of the past data that is determined to be similar to the most recent data using a similarity determination model that has been trained on the past data and the most recent data. An operation prediction program characterized by:

18. A step of dividing learning data including operation data indicating the operation status of a train and people flow data indicating the situation of people flow of users using the train into past data and most recent data acquired after the past data; generating a similarity determination model that determines similar data of the past data that is similar to the most recent data by learning the past data and the most recent data; determining the similar data from the past data based on the similarity determination model; generating a prediction model that predicts at least one of the operation status and the pedestrian flow status in the future by learning from the most recent data and the similar data; A learning program characterized by causing a computer to execute the above.

Images