JP6761789B2 - People flow forecasters, systems and programs - Google Patents

Description

The present invention relates to a human flow predictor, a system and a program for predicting a future human flow or the like based on data of a past human flow or the like.

Conventionally, in a space with buildings, theme parks, shopping streets, etc. (hereinafter referred to as the target space), the flow of people (hereinafter referred to as the flow of people) and the traffic volume at any point (hereinafter referred to as the measurement point) can be grasped. Attempts have been made to utilize this in designing layouts in the target space and optimal operation of equipment such as air conditioning.

As means for measuring the traffic flow and the traffic volume (hereinafter, the traffic volume and the traffic volume are collectively referred to as a human flow or the like), a means using imaging by a camera (see, for example, Patent Document 1 and Patent Document 2), a point group using a laser. Means using (laser range finder, range scanner, laser scanner, hereinafter collectively referred to as LRF) (see, for example, Non-Patent Document 1), or means using an infrared sensor (see, for example, Non-Patent Document 2). Etc. have been proposed.

In addition, as a technique for predicting the flow of people after the measurement time from the measured flow of people, a technique of predicting the flow of people by combining an elevator and a crowd behavior model (see, for example, Patent Document 3), a flow of people in a target space, etc. A technique that replaces the prediction of the flow of people in the target space with the flow of people in another space that correlates with (see, for example, Patent Document 4), or a model used for some kind of prediction such as a crowd behavior model for each day and time zone. There have been proposed techniques (see, for example, Patent Document 5) that classify only by categorization and statically construct and use it.

Patent No. 4913895 Patent No. 5856456 Patent No. 4228862 Japanese Unexamined Patent Publication No. 2011-231946 Patent No. 5769860

Yusuke Wada, Yoshitaka Nakamura, Teruo Higashino, "Proposal of Human Flow Modeling Method Using Range Scanner in Space with Obstacles," Multimedia, Distributed, Emphasis and Mobile (DICOMO2011) Symposium, pp.1183-1192, 2011 .. Yoshihiko Hata, Hiroki Akada, Takashi Yoshikawa, Tomoaki Kado, Traffic Judgment Method in Corridor by Pyroelectric Infrared Sensor, Information Processing Society of Japan Research Report Ubiquitous Computing System (UBI), Vol. 2016-UBI-4, pp.1 -6, 2016.

However, the technique described in Patent Document 3 allocates the number of persons measured at the measurement points for each target area. For example, before the start of business of a commercial facility, the measurement of the flow of people and the like is still performed. At that time, it is not possible to predict the flow of people during business hours. That is, it is possible to predict the flow of people and the like only when the data measured immediately before exists. Similarly, it is not possible to predict future flow of people, etc.

In the first place, the method of using only the number of users of a specific device such as an elevator as described in Patent Document 3 for prediction cannot be applied to low-rise facilities where elevators are rarely used, for example. There was also the problem of being restricted.

Further, in the technology described in Patent Document 4, in a situation where the flow of people changes significantly, for example, before and after the start of business of a commercial facility, the number of people is increased as in the surrounding space before the start of business. Even if the data at a small time point is used to predict the flow of people during the busy hours after the start of business, the predicted value will be inaccurate.

In addition, in the technology described in Patent Document 5, in a situation where there is some fluctuation that affects the flow of people, etc., a statically constructed model is used to obtain the future flow of people, etc. over a long period of time. It is virtually impossible to make stable and accurate predictions.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a human flow prediction device, system and program capable of predicting future human flow data based on an appropriate dynamically updated human flow prediction model. To provide.

In order to solve the above problem, the first aspect of the present invention is a human flow prediction device that predicts the human flow data related to the target section in the target space in the target space through which a person passes, from the past human flow data. Based on a storage medium that stores a human flow prediction model for predicting future human flow data, a data acquisition unit that acquires human flow data related to the target section, and the human flow prediction model and the acquired human flow data. A human flow data prediction unit that predicts future human flow data related to the target section, and a plurality of past human flow data acquired by the data acquisition unit are used to calculate the feature amount of each data, and the calculated feature amount and the calculated feature amount. By executing a statistical analysis process using a value representing a partial autocorrelation with the acquired human flow data, a variable automatic selection unit that selects a feature amount satisfying a predetermined condition and the selected feature amount are selected. As an explanatory variable for predicting future human flow data, a human flow prediction model learning function unit for learning the human flow prediction model is provided.

In a second aspect of the present invention, the data acquisition unit includes external data including at least one of information representing a physical phenomenon related to the target space and information representing the implementation status of an event, and a facility in the target space. At least one of the energy management data related to the energy management is further acquired, and the variable automatic selection unit obtains the feature amount calculated from the human flow data acquired in the past predetermined period related to the target section, and the feature amount. Based on the feature amount calculated from at least one of the external data and the energy management data acquired in the past predetermined period, the feature amount satisfying the predetermined condition is selected by the statistical analysis process, respectively. The human flow prediction model learning function unit learns the human flow prediction model by using each of the selected feature quantities as an explanatory variable for predicting future human flow data, and the human flow data prediction unit is trained. Based on the above-mentioned human flow prediction model, the acquired human flow data, and at least one of the acquired external data and the energy management data, the future human flow data related to the target section is predicted. Is.

In the third aspect of the present invention, the human flow prediction model learning function unit dynamically updates the human flow prediction model by supervised learning using the acquired human flow data related to the target section. It is a thing.

In a fourth aspect of the present invention, the person flow prediction device predicts in advance the value indicated by the acquired person flow data related to the target section and the time when the person flow data is acquired in relation to the target section. A difference calculation unit for calculating the difference from the value indicated by the human flow data is further provided, and when the calculated difference value is larger than a predetermined threshold value, the human flow prediction model learning function unit up to the time. The human flow prediction model is retrained based on at least the human flow data acquired during the period.

A fifth aspect of the present invention is a human flow prediction device that predicts human flow data related to a target section in the target space in which a person passes, in order to predict future human flow data from past human flow data. A storage medium for storing the human flow prediction model, a data acquisition unit for acquiring the human flow data related to the target section, and future human flow data related to the target section based on the human flow prediction model and the acquired human flow data. A human flow data prediction unit that predicts the number of people, a human flow prediction model learning function unit that dynamically updates the human flow prediction model using at least the acquired human flow data, data representing the total traffic volume, and the target section. Bei example a completion unit which generates a supervised learning a relationship model representing the relationship between the human stream data according the pedestrian flow prediction model learning unit, the can be calculated from the past people flow data object space according to the target segments Based on the data representing the total traffic volume according to the above, the human flow prediction model for predicting the data representing the future total traffic volume related to the target space is learned, and the human flow data prediction unit is based on the human flow prediction model. wherein predicting a total traffic amount of the future according to the target space, based on said relationship model and the previous SL entire future according to predicted the target space was traffic amount predicted future human stream data according to the target segments Te It is something that I tried to do.

According to the first aspect of the present invention, the future people flow data related to the target area is based on the person flow prediction model in which the person flow data related to the target area is acquired and dynamically updated, and the acquired people flow data. Is expected. The flow prediction model is dynamically updated with at least the acquired flow data. In this way, by using an appropriate people flow prediction model that is dynamically updated, it is possible to stably and accurately predict future people flow and the like for a long period of time. In addition, by using an appropriate flow prediction model, for example, even during a period including a period in which there is a discontinuity in the flow of people, such as outside the business hours of the facility, for example, a period of one day to several weeks. , It becomes possible to predict the flow of people.
Further, according to the first aspect of the present invention, the feature amount of the data is calculated from the acquired data, and a value representing the correlation between the calculated feature amount and the acquired human flow data is used. By executing the statistical analysis process, a feature quantity satisfying a predetermined condition is selected. The above-mentioned human flow prediction model has been learned to use the selected features as explanatory variables for predicting future human flow data. Therefore, in order to predict the flow of people and the like, it is possible to select an explanatory variable that contributes to improving the accuracy of the prediction of the flow of people and the like from the acquired data.

According to the second aspect of the present invention, external data including at least one of information representing a physical phenomenon related to the target space and information representing the implementation status of an event, and energy management data relating to energy management in a facility in the target space. At least one of and is further acquired. The above-mentioned human flow prediction model is trained based on the human flow data acquired in the past predetermined period related to the target section and at least one of the external data and the energy management data acquired in the past predetermined period. Future human flow data related to the target section is predicted based on the learned human flow prediction model, the acquired human flow data, and at least one of the acquired external data and energy management data. In this way, it can have a great influence on the flow of people in a predetermined section, for example, in a commercial complex, a) an event held in the facility, b) replacement of stores in the facility, c) sales of seasonal products, etc. Or, it was learned in consideration of the influence of various external factors such as d) delay of transportation network and situation of extra flights, e) outdoor weather, etc. in facilities directly connected to public transportation. By using the people flow prediction model, it is possible to predict the future people flow, etc., taking into consideration the influence of these external factors.
Further, according to the second aspect of the present invention, the feature amount selected from the feature amount calculated from the human flow data and the feature amount calculated from at least one of the external data and the energy management data is set to the future human flow. The human flow prediction model is trained using it as an explanatory variable for predicting data. Therefore, it is possible to further improve the prediction accuracy of the flow of people and the like.

According to the third aspect of the present invention, the human flow prediction model is dynamically updated by supervised learning using the acquired human flow data related to the target section. For this reason, even if the accuracy of prediction by the human flow prediction model gradually declines due to the influence of various external factors, the human flow prediction model will reflect the influence of external factors by taking into account the appropriate number of updates. It can be updated each time to the one that has been made, and by using the person flow prediction model, it is possible to continue to accurately predict the future people flow and the like.

According to the fourth aspect of the present invention, the value indicated by the person flow data related to the acquired target section and the value indicated by the person flow data predicted in advance for the target section with respect to the time when the person flow data was acquired. The difference with is calculated. When the calculated difference value is larger than a predetermined threshold value, the human flow prediction model is retrained based on at least the human flow data acquired in the period up to the above time. Therefore, when the predicted value of the human flow or the like deviates greatly from the actual event, it is possible to relearn the human flow prediction model suitable for the current use.

According to the fifth aspect of the present invention, the traffic prediction model relates to the target space based on data representing the total traffic volume of the target space that can be calculated from the past traffic data of the target section. It predicts the data representing the future total traffic volume, and the future total traffic volume related to the target space is predicted based on the traffic prediction model. Based on the relational model that represents the relationship between the data representing the total traffic volume related to the target space and the human flow data related to the target space generated by supervised learning, and the future total traffic volume related to the predicted target space. Therefore, future data on the flow of people related to the target area is predicted. Therefore, for example, even if the sensor of the target section does not operate normally due to a failure or the like, it is possible to continuously predict the future flow of people related to the target section. Further, even when the measurement accuracy of the human flow sensor is low, it is possible to predict the future human flow data with improved accuracy as compared with the case of predicting the future human flow data without passing through the total traffic volume.

That is, according to the present invention, it is possible to provide a flow prediction device, a system and a program capable of predicting future flow data based on an appropriate dynamically updated flow prediction model.

The block diagram which shows the functional structure of the person flow prediction apparatus which concerns on 1st Embodiment of this invention. FIG. 5 is a flow chart showing an example of a human flow prediction process in an initial state executed by the control unit shown in FIG. FIG. 5 is a flow chart showing an example of a human flow prediction process in an initial state executed by the control unit shown in FIG. The figure which shows an example of the sensor data table stored in the storage unit shown in FIG. The flow diagram which shows an example of the human flow prediction model learning process executed by the control unit shown in FIG. The flow diagram which shows an example of the human flow prediction model learning process executed by the control unit shown in FIG. The figure which shows an example of the person flow prediction model table stored in the storage unit shown in FIG. The flow diagram which shows an example of the person flow prediction model update processing executed by the control unit shown in FIG. The flow diagram which shows an example of the person flow prediction model update processing executed by the control unit shown in FIG. The flow diagram which shows an example of the person flow prediction processing based on the learned person flow prediction model executed by the control unit shown in FIG. The flow diagram which shows an example of the relational model learning process executed by the control unit shown in FIG. The figure which shows an example of the relational model table stored in the storage unit shown in FIG. The flow diagram which shows an example of the relational model update process executed by the control unit shown in FIG. The flow diagram which shows an example of the relational model update process executed by the control unit shown in FIG. The flow diagram which shows an example of the person flow prediction processing based on the relational model executed by the control unit shown in FIG. The flow diagram which shows an example of the person flow prediction processing based on the relational model executed by the control unit shown in FIG. The flow diagram which shows an example of the variable automatic selection process executed by the control unit shown in FIG. The flow diagram which shows an example of the human flow prediction model relearning process at the time of error expansion executed by the control unit shown in FIG. The flow diagram which shows an example of the human flow prediction model relearning process at the time of error expansion executed by the control unit shown in FIG.

Hereinafter, embodiments according to the present invention will be described with reference to the drawings.
[First Embodiment]
(Constitution)
FIG. 1 is a block diagram showing a functional configuration of the human flow prediction device 1 according to the first embodiment of the present invention. In the target space such as a facility through which people pass, the person flow prediction device 1 illustrated in FIG. 1 determines the flow of people and the amount of traffic entering and exiting the target section for each one or more target sections existing in the target space. Future people flow data can be predicted based on people flow data representing people flow and the like.

The sensor 2 includes a human flow sensor that acquires human flow data related to the target section, and an external sensor that acquires external data related to the target space. The human flow data shall be acquired by, for example, imaging with a camera, a point cloud with a laser, an infrared sensor, or the like. The external data is, for example, information representing a physical phenomenon such as temperature, temperature, wind speed, atmospheric pressure, precipitation, and radio wave intensity data of a wireless access point in or around the target space. Note that the external data is not limited to those containing information representing physical phenomena related to the target space, for example, data such as special operation information of public transportation around the target space and implementation of events related to the target space. It may include information indicating the situation and the like.

The building energy management system (BEMS) 3 includes, for example, the set temperature of air conditioning in the target space, the power supply state of the fan coil unit (hereinafter referred to as FCU), the valve opening of the heat source of the FCU, the amount of water used, the amount of power consumption, and the like. The BEMS data, which is the BEMS data representing the above, is acquired. In the following, an example of executing the human flow prediction process using the BEMS data will be described, but the human flow prediction process may be executed using the energy management data related to energy management in other facilities in the target space.

The person flow prediction device 1 according to the present embodiment includes a control unit 11, a storage unit 12, an external interface unit 13, and an input / output interface unit 14.

The external interface unit 13 includes, for example, one or more wired or wireless communication interface units, and is related to the human flow data related to the target section from the human flow sensor included in the sensor 2 and the target space from the external sensor included in the sensor 2. BEMS data is acquired from BEMS 3 as external data, and the acquired human flow data, external data, and BEMS data are input to the control unit 11.

The input / output interface unit 14 inputs, for example, an operation signal or the like input by the input unit 4 including a keyboard, a mouse, or the like to the control unit 11, and causes the display unit 5 to display the display data output from the control unit 11. ..

The storage unit 12 uses a non-volatile memory such as an HDD (Hard Disc Drive) or SSD (Solid State Drive) that can be written and read at any time as a storage medium, and is used to realize the present embodiment. As a storage area, a sensor data table storage unit 121, a human flow prediction model table storage unit 122, an auxiliary relationship model table storage unit 123, and a human flow prediction data storage unit 124 are provided.

The sensor data table storage unit 121 is used to store the human flow data, the external data, the BEMS data, and the predicted future human flow data.

The human flow prediction model table storage unit 122 is used to store a human flow prediction model for predicting future human flow data from past human flow data.

The relational model table storage unit 123 stores, for example, a relational model that represents the relationship between the number of visitors to the target space, the number of participants, or both of the total traffic volume and the traffic data related to the target section. used.

The person flow prediction data storage unit 124 is used to store the predicted future person flow data. The predicted future human flow data stored in the human flow prediction data storage unit 124 is sequentially moved to, for example, the sensor data table storage unit 121. The predicted future human flow data may be stored directly in the sensor data table storage unit 121, and the storage unit 12 may not include the human flow prediction data storage unit 124.

The control unit 11 includes a CPU (Central Processing Unit), and in order to execute the processing function in the present embodiment, the control unit 11 includes a data acquisition unit 111, a human flow prediction unit 112, a variable automatic selection unit 113, and an output unit 114. I have.

The data acquisition unit 111 includes a human flow data acquisition unit 1111, an external data acquisition unit 1112, a BEMS data acquisition unit 1113, and a difference calculation unit 1114.

The human flow data acquisition unit 1111 acquires the human flow data related to the target section from the human flow sensor in the sensor 2 via the external interface unit 13, and stores the acquired human flow data in the sensor data table storage unit 121 of the storage unit 12. Execute the process.

The external data acquisition unit 1112 acquires external data related to the target space from the external sensor of the sensor 2 via the external interface unit 13, and stores the acquired external data in the sensor data table storage unit 121 of the storage unit 12. Execute the process to memorize.

The BEMS data acquisition unit 1113 acquires BEMS data from the BEMS 3 via the external interface unit 13, and executes a process of storing the acquired BEMS data in the sensor data table storage unit 121 of the storage unit 12.

The difference calculation unit 1114 receives the person flow data related to the acquired target section, and is stored in the sensor data table storage unit 121 of the storage unit 12 for the time when the person flow data is acquired, which is predicted in advance for the target section. Executes the process of reading the existing human flow data. After that, the difference calculation unit 1114 calculates the difference between the value indicated by the person flow data related to the acquired target section and the value indicated by the person flow data predicted in advance for the target section at the time when the person flow data was acquired. Execute the process to be performed.

The human flow prediction unit 112 includes a human flow prediction model learning function unit 1121, an auxiliary measuring instrument complementary function unit 1122, and a human flow data prediction unit 1123.

The human flow prediction model learning function unit 1121 acquires in the past predetermined period stored in the sensor data table storage unit 121 of the storage unit 12 in the human flow prediction model learning mode prior to the human flow prediction processing based on the learned human flow prediction model. The past people flow data, external data, and BEMS data related to the target section are read, and the past is based on the people flow data, external data, and BEMS data related to the target section acquired in the past predetermined period. A process of learning a human flow prediction model for predicting future human flow data from the human flow data and storing the learned human flow prediction model in the human flow prediction model table storage unit 122 of the storage unit 12 is executed. The predetermined period is an arbitrary period and may be set by the operator.

The human flow prediction model learning function unit 1121 receives the human flow data related to the acquired target section after the human flow prediction model is constructed in the human flow prediction model learning mode, and stores it in the human flow prediction model table storage unit 122 of the storage unit 12. The human flow prediction model is read out, and the read human flow prediction model is dynamically updated by supervised learning using the acquired human flow data, and the updated human flow prediction model is stored in the human flow prediction model table of the storage unit 12. The process of storing in the storage unit 122 is executed.

The person flow prediction model learning function unit 1121 predicts in advance the value indicated by the person flow data related to the acquired target section calculated by the difference calculation unit 1114 and the time when the person flow data is acquired in relation to the target section. When the information that the value of the difference from the value indicated by the human flow data is larger than the predetermined threshold value is received, the human flow data, the external data, and the BEMS data acquired during the period until the time when the person flow data is acquired are received. Based on the above, the human flow prediction model is relearned, and a process of storing the relearned human flow prediction model in the human flow prediction model table storage unit 122 of the storage unit 12 is executed.

In the human flow data prediction mode, the human flow data prediction unit 1123 reads out the human flow prediction model stored in the human flow prediction model table storage unit 122 of the storage unit 12, and based on the read out human flow prediction model, the sensor of the storage unit 12 A process of reading data used for predicting future human flow data related to the target section from the human flow data, external data, and BEMS data stored in the data table storage unit 121 is executed. After that, the human flow data prediction unit 1123 predicts future human flow data related to the target section based on the read human flow prediction model and the read data, and stores the predicted human flow data in the storage unit 12. The process of storing in the human flow prediction data storage unit 124 is executed. The prediction process may be executed, for example, in response to an operation signal or the like input by the input unit 4 via the input / output interface unit 14.

The variable automatic selection unit 113 calculates the feature amount of the data from the data acquired by the data acquisition unit 111, and statistically uses a value representing the correlation between the calculated feature amount and the acquired human flow data. By executing the analysis process, a process of selecting a feature amount satisfying a predetermined condition is executed. After that, the human flow prediction model learning function unit 1121 executes a process of learning a human flow prediction model that uses the selected feature amount as an explanatory variable for predicting future human flow data as the human flow prediction model. Can be done.

The output unit 114 reads the future human flow data related to the predicted target section stored in the human flow prediction data storage unit 124 of the storage unit 12, and reads the read human flow data via the input / output interface unit 14. A process of outputting to the display unit 5 and displaying the information of the data is executed.

In the measuring instrument complement function unit 1122, for example, in order for the above-mentioned human flow prediction model to predict future human flow data from past human flow data, the total traffic amount related to the target space that can be calculated from the past human flow data related to the target section. It is used when the data representing the future total traffic volume related to the target space is predicted based on the data representing the above.

When such a human flow prediction model is used, the measuring instrument complement function unit 1122 is stored in the sensor data table storage unit 121 of the storage unit 12 in the relational model learning mode prior to the human flow prediction processing based on the learned human flow prediction model. The human flow data, external data, and BEMS data related to the target section acquired in the past predetermined period are read out, and the human flow data, external data, and BEMS related to the target section acquired in the past predetermined period are read out. Based on the data, a relational model showing the relationship between the data representing the total traffic volume and the human flow data related to the target section is generated by supervised learning, and the generated relational model is stored in the relational model table storage unit of the storage unit 12. The process of storing in 123 is executed. The predetermined period is an arbitrary period and may be set by the operator. Further, the measuring instrument complement function unit 1122 can execute a process of updating and relearning the relational model stored in the relational model table storage unit 123 of the storage unit 12, similarly to the human flow prediction model learning function unit 1121. ..

In such a case, the human flow data prediction unit 1123 reads out the human flow prediction model stored in the human flow prediction model table storage unit 122 of the storage unit 12 in the human flow data prediction mode, and based on the read out human flow prediction model. Predict the future total traffic volume for the target space. After that, the human flow data prediction unit 1123 reads out the relational model stored in the relational model table storage unit 123 of the storage unit 12, the read relational model, and the future total traffic volume related to the predicted target space. Based on, the process of predicting the future flow data related to the target section is executed.

(motion)
Next, the operation of the human flow prediction device 1 configured as described above will be described.

(1) Human flow data prediction processing in the initial state The human flow prediction device 1 of the present invention executes a process of predicting future human flow data based on the learned human flow prediction model. For example, a human flow prediction model is learned. Previously, a process of predicting future people flow data may be executed based on a predetermined people flow prediction model as described below.

2A and 2B are flow charts showing an example of prediction processing of future human flow data based on a predetermined human flow prediction model executed by the control unit 11 shown in FIG. The sensor data storage process from step S101 to step S103 in FIG. 2A and the human flow data prediction process from step S104 to step S107 in FIG. 2B are illustrated and described as a continuous flow. The data prediction process can be executed at an arbitrary timing regardless of the sensor data storage process. For example, the human flow data prediction process may be executed at an arbitrary timing requested by the operator.

First, in step S101, pedestrian flow sensor of the sensor 2 measures the human flow or the like according to target segments of the sensor in the target space passing a person, pedestrian flow together with the identifier I _i of the sensor was measured human stream data Input to the prediction device 1. As the identifier _Ii , one identifier may be set for one sensor, and when one sensor measures at a plurality of measurement points, each measurement point may be set. A plurality of different identifiers may be set. Further, the identifier _Ii is set by, for example, the installer of the sensor. These identifiers I _f to be described by using as an identifier, for example the sensor throughout this specification also _applies to I _e like. Further, each of the identifiers I _i , _{If f} , I _e, etc. may specify an arbitrary sensor, or may represent a different identifier corresponding to the sensor each time.

In step S102, the control unit 11 of the pedestrian flow prediction apparatus 1 under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111 to obtain a corresponding identifier I _i with the human stream data at time t.

In step S103, the control unit 11 stores the acquired human flow data together with the identifier _Ii and the time t in the sensor data table storage unit 121 of the storage unit 12 under the control of the human flow data acquisition unit 1111 of the data acquisition unit 111. Let me. Incidentally, in the sensor data table storage unit 121, if the predicted human stream data identified by the identifier I _i at time t is stored stores the pedestrian flow data obtained by overwriting the data The same shall apply hereinafter. However, this does not apply if the storage unit 12 stores the predicted human flow data in a location different from that of the sensor data table storage unit 121. By repeatedly executing the processes of steps S101 to S103, for example, human flow data from one or more arbitrary number of sensors are stored in the sensor data table storage unit 121 together with the corresponding identifier and time, respectively. ..

FIG. 3 is a diagram showing an example of a sensor data table stored in the sensor data table storage unit 121 of the storage unit 12 shown in FIG. The sensor data table is a list of sensor data including human flow data. In the example shown in FIG. 3, as sensor data, the number of people existing in the target section at the time of measurement is stored. By dividing the sensor data table logically or physically, the sensor data table may be stored in different sensor data tables depending on the sensor, the measurement point, the measurement time, or the like. A reference to a place where the target data is stored may be provided by the storage device itself, or the information may be included in a request statement (hereinafter referred to as a query) issued by each component of the present invention when requesting data. It may be solved by including.

Here, the human flow data represents, for example, the following, but is not limited to this.
-The number of people who passed the measurement point per arbitrary unit time such as 1 minute (hereinafter referred to as the number of people who passed)
・ Number of people within a radius of a meter at the measurement point (hereinafter referred to as the number of people staying)
-Movement direction of people who have passed the measurement point-For data that is the measurement point where the target space is connected to the outside as the measurement point, for example, the total passage of the target space obtained by adding all the number of people passing by for each measurement time Quantity Then, in step S104, the control unit 11 determines the identifier _If of the human flow data to be predicted and the time t _f to be predicted under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112, and the identifier _{If f} determines the identifier _If. The storage unit 12 is requested to receive the identified past n times of human flow data. The determination of the identifier _If and the time t _f may be, for example, input by the operator via the input unit 4, and the same applies to the following. The past n times of human flow data are determined by, for example, a predetermined human flow prediction model stored in the human flow prediction model table storage unit 122 of the storage unit 12.

In step S105, the storage unit 12 returns the requested past n times of human flow data to the control unit 11 from the sensor data table storage unit 121.

In step S106, the control unit 11 acquires the returned past n times of human flow data under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112, and acquires the acquired past as the maximum likelihood estimation value of the human flow and the like. The average value of the human flow data for n times is calculated and used as the predicted value of the human flow data.

In step S107, the control unit 11, under the control of the human stream data prediction unit 1123 of the pedestrian flow prediction unit 112, a predicted value of the human stream data, as a predictive value for human stream data identified by the identifier I _f at time t _f, It is stored in the human flow prediction data storage unit 124 of the storage unit 12.

(2) Human flow prediction model learning process FIGS. 4A and 4B are flows showing an example of a human flow prediction model learning process executed prior to the human flow prediction process based on the learned human flow prediction model by the control unit shown in FIG. It is a figure. The sensor data storage process from step S201 to step S207 in FIG. 4A and the human flow prediction model learning process from step S208 to step S213 in FIG. 4B are illustrated and described as a continuous flow. The human flow prediction model learning process can be executed at an arbitrary timing regardless of the sensor data storage process. For example, the human flow prediction model learning process may be executed at an arbitrary timing requested by the operator.

First, in step S201, pedestrian flow sensor of the sensor 2 measures the human flow or the like according to target segments of the sensor in the target space passing a person, pedestrian flow together with the identifier I _i of the sensor was measured human stream data Input to the prediction device 1.

In step S202, the control unit 11 of the pedestrian flow prediction apparatus 1 acquires the corresponding identifier _{I i} under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111, at time _{t 1} with the human stream data.

On the other hand, in step S203, the external sensor of the sensor 2 measures a physical phenomenon related to the target space, the external data measured inputs together with the identifier I _i of the sensor to the pedestrian flow prediction apparatus 1. Note that the external data is not limited to those containing information representing physical phenomena related to the target space, and may include, for example, information representing the implementation status of events related to the target space, and may also include the following. The same is true.

In step S204, the control unit 11 of the pedestrian flow prediction apparatus 1 acquires the corresponding identifier _{I i} under the control of an external data acquisition unit 1112 of the data acquisition unit 111, at time _{t 2} with the external data.

In step S205, BEMS3 inputs BEMS data in the target space to the pedestrian flow prediction apparatus 1 together with the identifier _{I i} of the data. In addition, instead of the BEMS data, energy management data related to energy management in other facilities in the target space may be input, and the same applies to the following.

In step S206, the control unit 11 of the pedestrian flow prediction apparatus 1 acquires the identifier _{I i} corresponding with the BEMS data in BEMS under the control of the data acquisition unit 1113, the time _{t 3} of the data acquisition unit 111.

The order of each step from step S201 to step S206 is only for convenience, and each step of step S201, step S203, and step S205 may be executed in any order, and steps S202, step S204, and step S204, and each step may be executed in any order. Each step of step S206 may be executed after the corresponding step of step S201, step S203, and step S205, respectively. Further, steps S203 to S206 are supplementary, and the operation of each step may not be executed.

Next, in step S207, the control unit 11 acquires the acquired human flow data, external data, and under the control of the human flow data acquisition unit 1111, the external data acquisition unit 1112, and the BEMS data acquisition unit 1113 of the data acquisition unit 111. The BEMS data is stored in the sensor data table storage unit 121 of the storage unit 12 together with the identifier and time of each data. By repeatedly executing the processes of steps S201 to S207, for example, human flow data, external data, and BEMS data from one or more arbitrary number of sensors are stored in the sensor data table together with the corresponding identifiers and times, respectively. It is stored in the part 121.

Subsequently, in step S208, the control unit 11 determines an identifier _If that specifies the area people flow or the total traffic volume to be predicted under the control of the person flow prediction model learning function unit 1121 of the person flow prediction unit 112, and determines the identifier I _{f. The} storage unit 12 is requested to have data for the past M days (hereinafter referred to as an objective variable) identified by _f .

In step S209, the storage unit 12 returns the requested human flow data for the past M days to the control unit 11 from the sensor data table storage unit 121.

In step S210, the control unit 11, under the control of the human flow prediction model learning unit 1121 of the pedestrian flow prediction unit 112, a value indicated by the S data identified by the identifier I _f at time t and S was S _t , for example, the time granularity, such as 1 hour Delta] t in the case where the value of the objective variable at time t-n × Delta] t before by n unit time when a (hereinafter, the unit time and called) is expressed as S _t-n, Partial autocorrelation of j unit time difference This is the unit time difference in which the absolute value of partial autocorrelation | PACF (j) | is the maximum in the past m days with respect to PACF (j).
To decide.

In step S211 the control unit 11 was determined under the control of the human flow prediction model learning function unit 1121 of the human flow prediction unit 112.
Based on the,
Is one of the explanatory variables, and is the average value of the past m days from the time t of S.
Is one of the explanatory variables. After that, for example, when a multiple regression model in which the base function φ (x) is a linear function of an explanatory variable as a multiple regression analysis is used as a human flow prediction model, the partial regression coefficients a ₁ and a _{2 of the} human flow prediction model are used. In order to determine, ..., the data for the past M days (where M> m) stored in the sensor data table storage unit 121 of the storage unit 12 is read out, and the above explanatory variable corresponding to the objective variable at time t is used. A set of all combined data
Was a one data set, followed by a time t-Delta] t target variable S _{t-Delta] t} to the set of data that combines all of the above described variable corresponding to
Is used as one data set, and D data sets are created by repeating this until the data read by the above explanatory variables cannot be calculated.

In step S212, the control unit 11 sets the partial regression coefficients a ₁ , a ₂ , ... Can be constructed from the above D data sets as unknown variables under the control of the human flow prediction model learning function unit 1121 of the human flow prediction unit 112. The partial regression coefficient is calculated by solving D simultaneous equations by a method such as the least squares method.

In step S213, the control unit 11, a human stream under the control of the human flow prediction model learning unit 1121 of the prediction unit 112, the identifier _{I al} explanatory variables _{x i} corresponding to partial regression coefficients calculated _{a l} and _{a l,} time A human flow prediction model for predicting future human flow data, including the relative time j from t or the calculation formula of the explanatory variable, and the objective variable identifier _If , is stored in the human flow prediction model table storage unit 122 of the storage unit 12. Remember.

Although the case where only the human flow data is used for the calculation of the explanatory variables has been described above, the external data and the BEMS data may also be used for the calculation of the explanatory variables.

When the calculation formula is used, the storage unit 12 stores, for example, the calculation formula itself including the necessary sensor data identifier together with the identifier of the calculation formula.

FIG. 5 is a diagram showing an example of a human flow prediction model table stored in the human flow prediction model table storage unit 122 of the storage unit 12 shown in FIG. 1.

In the first line, when the human flow data identified by the identifier "0000abcd" is used as the objective variable, for example, the data of "ID: 0000abcd, j = 24 × 3600" indicating the personal flow data at the same time on the previous day is the explanatory variable. Indicates that it is used as. Further, in the third line, when the human flow data identified by the identifier "0000abcd" is used as the objective variable, for example, "average (ID: 0000abcd, 5x) indicating the average value of the values indicated by the personal flow data for the past 5 days" is shown. 24 × 3600) ”data also indicates that it is used as an explanatory variable.

Here, the explanatory variables include, for example, the following in addition to those described above, but the explanatory variables are not limited thereto.
・ The unit time difference is the highest h in order from the one with the largest | PACF (j) |
When
・ When the days of the week in the seven-day system are 1, 2, ... 7 in order from Monday to Sunday, if the day of the week to which time t belongs is k, the kth element is 1 and the other elements are 0. [rho compared with the case of calculating the correlation coefficient S _t'certain such 7-dimensional day vector is day w ₁ of S _t different day of the w ₂ [rho, for example, a threshold value [rho _theta such 0.8 > and [rho _theta becomes the first category the day, any day with the day of the week which does not satisfy the above conditions with a c-number of categories such as its day only one category, as well defined c dimension day with the day the vector Hours in the 24-hour system to which the category vector t belongs
・ Variables that quantify the seasons such as the month or four seasons to which t belongs ・ The above month or four seasons are the number of each element, 12 or 4, and only the element to which t belongs is 1, and the other elements are 0. At a position close to the target space on the day or time to which the meteorological data predicted by the weather forecast or the predicted value of meteorological data such as precipitation probability, expected temperature, etc. on the day or time to which the moon vector or seasonal vector t belongs Authenticity value of whether or not there is a special operation of a certain public transportation ・ Authenticity value of whether or not there is an event in the target space or its neighboring facilities on the day or time to which t belongs ・ The scale and target of the above event Event feature vector that vectorizes features such as age group and participation fee In addition to the multiple regression model described above, the human flow prediction model includes, but is not limited to, the following.
・ Of the explanatory variables x _i , when the mean of D used for learning is μ _i and the standard deviation is σ _i , the mean is 0 and the variance is 1.
Multiple regression model using as an explanatory variable instead of x _i・ Polynomial multiple regression model with φ (x) as a polynomial of the explanatory variable ・ Vector weight with φ (x) as the base function that allows multidimensional explanatory variables Regression model ・ Multiple regression model including interaction terms with the product of different explanatory variables as φ (x) as the base function ・ Multiple regression model including any multiple effects among the above effects In addition, the derivation of the partial regression coefficient is , In addition to those using the minimum square method described above, there are the following, but the present invention is not limited to this.
-A method that uses an approximate solution such as a stochastic gradient descent method instead of the least-squares method-A method that applies the least-squares method after adding the terms of L1 regularization or L2 regularization-Multiple methods among the above methods Method of combining In addition, instead of storing the calculation formula of the explanatory variable in the storage unit 12, the data calculated by the calculation formula (hereinafter referred to as secondary data) is stored in the storage unit 12 as data, and a new identifier is stored. It may be given as an identifier of the secondary data and replaced by storing the identifier of the secondary data.

(3) People flow prediction model update process FIGS. 6A and 6B are flow charts showing an example of a person flow prediction model update process executed by the control unit 11 shown in FIG. In the human flow prediction model update process, the human flow prediction model is dynamically updated by supervised learning using the acquired human flow data.

First, in step S301, pedestrian flow sensor of the sensor 2 measures the human flow or the like according to target segments of the sensor in the target space passing a person, pedestrian flow together with the identifier I _i of the sensor was measured human stream data Input to the prediction device 1.

In step S302, the control unit 11 of the pedestrian flow prediction apparatus 1 under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111 to obtain a corresponding identifier I _i with the human stream data at time t.

In step S303, the control unit 11 stores the acquired human flow data together with the identifier _Ii and the time t in the sensor data table storage unit 121 of the storage unit 12 under the control of the human flow data acquisition unit 1111 of the data acquisition unit 111. Let me. Thereafter, the control unit 11, notification that under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111, a person stream data identified by the identifier I _i stored in the sensor data table storage section has been updated (hereinafter , Update notification) is transmitted to the human flow prediction model learning function unit 1121 of the human flow prediction unit 112.

Subsequently, in step S304, the control unit 11, under the control of the human flow prediction model learning unit 1121 of the pedestrian flow prediction unit 112, in response to receiving the update notification, predicted human stream data S which is identified by the identifier I _i The storage unit 12 is requested to have a human flow prediction model for this purpose.

In step S305, the storage unit 12 has a human flow prediction model for predicting the requested human flow data S for the control unit 11, and a human flow prediction model for predicting the human flow data S in the above-mentioned human flow prediction model learning process. The identifiers or calculation formulas of the explanatory variables required for the data set used for the training of the above are returned from the human flow prediction model table storage unit 122.

In step S306, the control unit 11 controls sensor data such as human flow data, external data, and BEMS data required for at least one of the above data sets under the control of the human flow prediction model learning function unit 1121 of the human flow prediction unit 112. Is requested from the storage unit 12.

In step S307, the storage unit 12 returns the requested sensor data to the control unit 11 from the sensor data table storage unit 121.

In step S308, the control unit 11 creates a data set from the returned sensor data under the control of the human flow prediction model learning function unit 1121 of the human flow prediction unit 112.

In step S309, the control unit 11 uses one or more created data sets under the control of the human flow prediction model learning function unit 1121 of the human flow prediction unit 112 to obtain a partial regression coefficient by a method such as a stochastic gradient descent method. Calculate and update.

In step S310, the control unit 11, under the control of the human flow prediction model learning unit 1121 of the pedestrian flow prediction unit 112, the identifier _{I al} explanatory variables _{x i} corresponding to the partial regression coefficients updated _{a l} and _{a l,} time The human flow prediction model including the relative time j from t, the calculation formula of the explanatory variable, and the identifier _{If of the} objective variable is stored in the human flow prediction model table storage unit 122 of the storage unit 12.

By repeating the processes from step S304 to step S310 every time new data such as human flow is stored in the sensor data table storage unit 121, the value of the partial regression coefficient is sequentially updated in the human flow prediction model. Is updated (hereinafter, this process is referred to as online learning).

(4) Human flow data prediction processing based on the learned human flow prediction model FIG. 7 shows future human flow based on the learned and updated human flow prediction model as described above executed by the control unit 11 shown in FIG. It is a flow chart which shows an example of the process of predicting data.

First, in step S401, the control unit 11 of the human flow prediction device 1 determines the identifier _If of the human flow data to be predicted and the time t'to be predicted under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112. A human flow prediction model learned in the above-mentioned human flow prediction model learning process or an updated human flow prediction model in the above-mentioned human flow prediction model update process, in which the value indicated by the human flow data identified by the identifier _If is used as the objective variable. , Request the storage unit 12.

In step S402, the storage unit 12 returns the requested human flow prediction model to the control unit 11 from the human flow prediction model table storage unit 122.

In step S403, the control unit 11 is based on the identifier of the explanatory variable included in the returned human flow prediction model and the relative time or the calculation formula of the explanatory variable under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112. The storage unit 12 is requested to have sensor data such as human flow data, external data, and BEMS data, which are necessary for the storage unit 12.

In step S404, the storage unit 12 returns the requested sensor data to the control unit 11 from the sensor data table storage unit 121.

In step S405, the control unit 11 calculates explanatory variables from the returned sensor data as necessary under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112.

In step S406, the control unit 11, under the control of the human stream data prediction unit 1123 of the pedestrian flow prediction unit 112, for example, the explanatory variables corresponding to the partial regression coefficients a _i and x _i contained in human flow prediction model is a regression model When you do, add all these products together
Was calculated and calculated
Is the predicted value of the human flow data identified by the identifier _If of time t'. In the above description and drawings, the case where the human flow prediction model is a multiple regression model is described as an example, but the human flow prediction model may be another model.

In step S407, the control unit 11 stores the predicted value of the human flow data in the human flow prediction data storage unit 124 of the storage unit 12 under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112.

The predicted value stored in the human flow prediction data storage unit 124 is stored in, for example, the same as the sensor data table shown in FIG.

(5) Relational Model Learning Process FIG. 8 is a flow diagram showing an example of the relational model learning process executed by the control unit 11 shown in FIG.

First, in step S501, the control unit 11 of the human flow prediction device 1 determines the identifier _If of the human flow data measured by the measuring instrument to be complemented under the control of the measuring instrument complement function unit 1122 of the human flow prediction unit 112. Then, the storage unit 12 is requested to have the total traffic volume and the human flow data identified by the identifier _If .

In step S502, the storage unit 12 returns the requested total traffic volume and the human flow data identified by the identifier _If to the control unit 11 from the sensor data table storage unit 121.

In step S503, the control unit 11 sets the person flow data _St at time t as the objective variable and the total traffic volume X _t at time _t as the explanatory variable under the control of the measuring instrument complement function unit 1122 of the person flow prediction unit 112. As one, the partial regression coefficient is calculated in the same manner as that executed in the above-mentioned human flow prediction model learning process.

In step S504, the control unit 11 includes the calculated partial regression coefficient, the corresponding explanatory variable identifier, and the objective variable identifier under the control of the measuring instrument complement function unit 1122 of the human flow prediction unit 112, and the total traffic volume. The relational model representing the relationship between the data representing the above and the human flow data is stored in the relational model table storage unit 123 of the storage unit 12.

FIG. 9 is a diagram showing an example of a relational model table stored in the relational model table storage unit 123 of the storage unit 12 shown in FIG.

Here, the explanatory variables include those shown below in addition to those described above, but are not limited thereto.
· Respect partial cross-correlation PCCF (j) between the total traffic amount X _t-j target variable S _t and a unit time difference j when using the same transcription as in the above-mentioned pedestrian flow prediction model learning processing of the last m days The unit time difference in which | PCCF (j) | is the maximum in the range
about
The total traffic volume and unit time difference are the top h in order from the one with the largest absolute value of partial cross-correlation | PCCF (j) |
When
・ When the days of the week in the seven-day system are 1, 2, ... 7 in order from Monday to Sunday, if the day of the week to which time t belongs is k, the kth element is 1 and the other elements are 0. [rho compared with the case of calculating the correlation coefficient S _t'certain such 7-dimensional day vector is day w ₁ of S _t different day of the w ₂ [rho, for example, a threshold value [rho _theta of 0.8, etc. > and [rho _theta becomes the first category the day, any day with the day of the week which does not satisfy the above conditions with a c-number of categories such as its day only one category, as well defined c dimension day with the day the vector Hours in the 24-hour system to which the category vector t belongs
・ Variables that quantify the seasons such as the month or four seasons to which t belongs ・ The above month or four seasons are the number of each element, 12 or 4, and only the element to which t belongs is 1, and the other elements are 0. At a position close to the target space on the day or time to which the meteorological data predicted by the weather forecast or the predicted value of the meteorological data such as the precipitation probability, the expected temperature, etc. Authenticity value of whether or not there is a special operation of a certain public transportation ・ Authenticity value of whether or not there is an event in the target space or its neighboring facilities on the day or time to which t belongs ・ The scale and target of the above event Event feature vector that vectorizes features such as age group and participation fee In addition to the multiple regression model described above, the relational models are as shown below, but are not limited to these.
・ Of the explanatory variables X _i , when the mean of D used for learning is μ _i and the standard deviation is σ _i , the mean is 0 and the variance is 1.
Multiple regression model using as an explanatory variable instead of X _i・ Polynomial multiple regression model with φ as a polynomial of the explanatory variable ・ Vector multiple regression model with a base function that allows multidimensional explanatory variables as φ ・ Different as φ Multiple regression model including interaction terms with the product of explanatory variables as the base function ・ Multiple regression model including any multiple effects among the above effects In addition, the minimum square method described above is used to derive the partial regression coefficient. In addition to the ones shown below, there are, but are not limited to.
-A method that uses an approximate solution such as stochastic gradient descent instead of the least-squares method-A method that applies the least-squares method after adding regularization terms such as L1 regularization or L2 regularization-Multiple of the above methods How to combine the methods of

(6) Relational model update processing FIGS. 10A and 10B are flow charts showing an example of the relational model update processing executed by the control unit 11 shown in FIG. In the relational model update process, the relational model is dynamically updated by supervised learning using the acquired human flow data.

First, in step S601, pedestrian flow sensor of the sensor 2 measures the human flow or the like according to target segments of the sensor in the target space passing a person, pedestrian flow together with the identifier I _i of the sensor was measured human stream data Input to the prediction device 1.

In step S602, the control unit 11 of the pedestrian flow prediction apparatus 1 under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111 to obtain a corresponding identifier I _i with the human stream data at time t.

In step S603, the control unit 11 stores the acquired human flow data together with the identifier _Ii and the time t in the sensor data table storage unit 121 of the storage unit 12 under the control of the human flow data acquisition unit 1111 of the data acquisition unit 111. Let me. Thereafter, the control unit 11, notification that under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111, a person stream data identified by the identifier I _i stored in the sensor data table storage section has been updated (hereinafter , Update notification) is transmitted to the measuring instrument complement function unit 1122 of the human flow prediction unit 112.

In step S604, the control unit 11, under control of the instrument completion unit 1122 of the pedestrian flow prediction unit 112, in response to receiving the update notification, the relationship for predicting human stream data S which is identified by the identifier I _i Request the model from the storage unit 12.

In step S605, the storage unit 12 learns the relational model for predicting the requested human flow data S for the control unit 11 and the relational model for predicting the human flow data S in the above-mentioned relational model learning process. The identifier or calculation formula of the explanatory variable required for the data set used is returned from the relational model table storage unit 123.

In step S606, the control unit 11 receives sensor data such as human flow data, external data, and BEMS data required for at least one of the above data sets under the control of the measuring instrument complementary function unit 1122 of the human flow prediction unit 112. , Request to the storage unit 12.

In step S607, the storage unit 12 returns the requested sensor data to the control unit 11 from the sensor data table storage unit 121.

In step S608, the control unit 11 creates a data set from the returned sensor data under the control of the measuring instrument complement function unit 1122 of the human flow prediction unit 112.

In step S609, the control unit 11 calculates the partial regression coefficient by a method such as a stochastic gradient descent method using one or more created data sets under the control of the measuring instrument complement function unit 1122 of the human flow prediction unit 112. And update.

In step S610, the control unit 11 includes the updated partial regression coefficient, the corresponding explanatory variable identifier, and the objective variable identifier under the control of the measuring instrument complement function unit 1122 of the human flow prediction unit 112. The relational model representing the relationship between the data representing the above and the human flow data is stored in the relational model table storage unit 123 of the storage unit 12.

By repeating the processes from step S604 to step S610 every time new data such as human flow is stored in the sensor data table storage unit 121, the relational model is sequentially updated in the value of the partial regression coefficient. Updated (online learning).

(7) People flow data prediction processing based on the relational model FIGS. 11A and 11B are flow charts showing an example of the people flow data prediction processing based on the relational model executed by the control unit 11 shown in FIG.

First, in step S701, the control unit 11 of the human flow prediction device 1 determines the identifier _If of the human flow data to be predicted and the time t'to be predicted under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112. The storage unit stores the relational model learned in the above-mentioned relational model learning process or the relational model updated in the above-mentioned relational model update process, which represents the relationship between the total traffic volume and the human flow data identified by the identifier _If. Request to 12.

In step S702, the storage unit 12 returns the requested relational model to the control unit 11 from the relational model table storage unit 123.

In step S703, the control unit 11 is a human flow prediction model learned in the above-mentioned human flow prediction model learning process for predicting the total traffic volume under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112, or The storage unit 12 is requested to use the human flow prediction model updated in the above-mentioned human flow prediction model update process.

In step S704, the storage unit 12 returns the requested human flow prediction model to the control unit 11 from the human flow prediction model table storage unit 122.

The order of each step from step S701 to step S704 is only for convenience, and each step of step S701 and step S703 may be executed in any order, and each step of step S702 and step S704, respectively. It may be executed after the corresponding step of step S701 and step S703.

In step S705, the control unit 11 is based on the identifier of the explanatory variable included in the returned human flow prediction model and the relative time or the calculation formula of the explanatory variable under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112. The storage unit 12 is requested to have sensor data such as human flow data, external data, and BEMS data, which are necessary for the storage unit 12.

In step S706, the storage unit 12 returns the requested sensor data to the control unit 11 from the sensor data table storage unit 121.

In step S707, the control unit 11 calculates explanatory variables from the returned sensor data as necessary under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112.

In step S708, the control unit 11 predicts the total traffic volume under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112, similarly to the human flow data prediction process based on the learned human flow prediction model described above. For example, the predicted value of the total traffic volume at time t'is calculated from the human flow prediction model which is a multiple regression model.

In step S709, the control unit 11 controls the total traffic volume at the predicted time t under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112.
And, for example, sensor data such as human flow data, external data, and BEMS data required based on the identifiers of other explanatory variables included in the relational model, which is a multiple regression model, and the relative time or the calculation formula of the explanatory variables are stored. acquired by requesting the unit 12, when the explanatory variables corresponding to the partial regression coefficients a _i in the relation model was x _i, and adds all these products
And this
Is the predicted value of the human flow data identified by the identifier _If of time t'. In the above description and drawings, the case where the human flow prediction model and the relational model are multiple regression models is described as an example, but the human flow prediction model and the relational model may be other models.

In step S710, the control unit 11 stores the predicted value of the human flow data in the human flow prediction data storage unit 124 of the storage unit 12 under the control of the human flow data prediction unit 1123 of the human flow prediction unit 112.

The predicted value stored in the human flow prediction data storage unit 124 is stored in, for example, the same as the sensor data table shown in FIG.

(8) Variable automatic selection process FIG. 12 is a flow chart showing an example of the variable automatic selection process executed by the control unit 11 shown in FIG. In the variable automatic selection process, the variable used as the explanatory variable in the human flow prediction model learned as described above is automatically selected.

First, in step S801, the control unit 11 of the pedestrian flow prediction apparatus 1 is extracted under the control of the human flow prediction model learning unit 1121 of the pedestrian flow prediction unit 112, an identifier I _f target variable S _t, the relevant feature amounts The identifier I _e of the sensor data such as the human flow data, the external data, and the BEMS data is determined, and the determined identifier I _f of the objective variable _{St and} the identifier I of the sensor data for which the related feature quantity is to be extracted. _e and are transmitted to the variable automatic selection unit 113. The feature amount is, for example, an amount that expresses the feature of the data by adding an arbitrary linear or non-linear transformation to the sensor data. The determination of the identifier _Ie of the sensor data may be, for example, automatically determined for each sensor identifier in order or randomly, or may be determined by the operator.

In step S802, the control unit 11 requests the storage unit 12 for the sensor data s for the past M days identified from the received identifier _Ie under the control of the variable automatic selection unit 113.

In step S803, the storage unit 12 returns the requested sensor data s for the past M days to the control unit 11 from the sensor data table storage unit 121.

In step S804, under the control of the variable automatic selection unit 113, the control unit 11 calculates data obtained by processing the sensor data s identified by the identifier _Ie , for example, a differential value s ′ obtained by differentiating the sensor data s. ..

In step S805, the control unit 11, under the control of variable automatic selection unit 113 calculates cross-correlation function γ a (jΔt) and as explanatory variables the differential value s'like _{x t.}

In step S806, the control unit 11 compares the maximum value γ _MAX of the calculated absolute value of the cross-correlation function γ (jΔt) with the threshold value γ _θ such as 0.8 under the control of the variable automatic selection unit 113. To do. The comparison result, if the gamma _MAX> gamma _theta satisfying of the above explanatory variables will be determined to be appropriate explanatory variables relating to the target variable S _t. On the other hand, if the above condition is not satisfied, the process returns to step S804, and the same operation is repeated for s ″ in which the sensor data s is subjected to other conversion. In addition, instead of the statistical analysis using the cross-correlation function, a statistical analysis using a correlation coefficient or the like may be used.

The operation from step S801 to step S806 until the above condition is satisfied is repeated for the number of necessary explanatory variables.

Thereafter, in step S807, the control unit 11 of the pedestrian flow prediction apparatus 1 under the control of the human flow prediction model learning unit 1121 of the pedestrian flow prediction unit 112, an identifier I _f target variable S _t, the above conditions are met A human flow prediction model including the above-mentioned processing calculation formula of the above-mentioned identifier _Ie related to the explanatory variable and the above-mentioned processing of the sensor data identified by the identifier _Ie , and the partial regression coefficient calculated by the same method as described above. The learned and learned human flow prediction model is stored in the human flow prediction model table storage unit 122 of the storage unit 12.

Here, the calculation formula of the explanatory variable includes, for example, the following in addition to the above differential value, but is not limited to this.
・ Average value of a collection of m consecutive sensor data s
Or standard deviation σ _m
- sensor data s collected by m pieces continuous in complex sequence F _i obtained by frequency conversion such as discrete Fourier transform or discrete wavelet transform unit inverse time Delta] t measurement of sensor data and f _{s =} 1 / Δt sometimes (hereinafter, referred to as the sampling frequency _{f s)} i th power spectrum of complex sequence _{F i} for the frequency resolution _{Δf = f s / m | F} i | 2 or declination arg _(F i)
・ The power of the above explanatory variables

(9) Human flow prediction model re-learning process FIGS. 13A and 13B are flow charts showing an example of a human flow prediction model re-learning process when an error is expanded, which is executed by the control unit 11 shown in FIG.

First, in step S901, the pedestrian flow sensor of the sensor 2 measures the human flow or the like according to target segments of the sensor in the target space passing a person, the measured person stream data S _t of the sensor identifier I _i At the same time, it is input to the person flow prediction device 1.

In step S902, the control unit 11 of the pedestrian flow prediction apparatus 1 acquires the corresponding identifier _{I i} under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111, at time t together with the human stream data _{S t.}

In step S903, the control unit 11 refers to the sensor data table stored in the sensor data table storage unit 121 of the storage unit 12 under the control of the human flow data acquisition unit 1111 of the data acquisition unit 111, and identifies the identifier I at time t. _When the predicted value of the human flow data identified by _i is stored, the human flow prediction data representing the predicted value is stored before overwriting the acquired human flow data _St in the sensor data table storage unit 121. Request to unit 12.

In step S904, the storage unit 12 returns the requested human flow prediction data to the control unit 11 from the sensor data table storage unit 121.

In step S905, the control unit 11 controls the flow prediction data under the control of the difference calculation unit 1114 of the data acquisition unit 111.
Calculating the difference [Delta] S _t with the value indicated by _the value and the human stream data S _t indicated.

In step S906, the control unit 11, under the control of the human stream data acquisition unit 1111 of the data acquisition unit 111, a who were stream data S _t acquired from the pedestrian flow sensor together with the identifier I _i and time t, the sensor data storage unit 12 It is stored in the table storage unit 121.

In step S907, the control unit 11, under the control of the difference calculation section 1114 of the data acquisition unit 111, it is compared with a threshold value θ a predetermined the difference [Delta] S _t. If the result of the comparison satisfies the condition of | ΔS _t |> θ, the process proceeds to the subsequent step S908. On the other hand, if the above conditions are not satisfied, the processing of the subsequent step is not executed. An example of calculating the difference between the value indicated by the human flow data and the value indicated by the human flow prediction data and comparing the calculated difference with the threshold value has been described. For example, the value indicated by the human flow data and the value indicated by the human flow prediction data are shown. The ratio to the value may be calculated and the calculated ratio compared to the threshold.

In step S908, the control unit 11 considers that the prediction is not normal under the control of the difference calculation unit 1114 of the data acquisition unit 111, and causes the human flow prediction model learning function unit 1121 to learn the identifier I of the human flow data that requires learning. _i and a signal indicating that learning is required (hereinafter, the transmitted signals are collectively referred to as a re-learning request notification) are transmitted together.

In step S909, the control unit 11, under the control of the human flow prediction model learning unit 1121 of the pedestrian flow prediction unit 112, in response to receipt of the retraining request notification, pedestrian flow past M days identified by the identifier I _i Data is requested from the storage unit 12.

After that, from step S910 to step S914, the same processing as in steps S209 to S213 of FIG. 4B described in relation to the above-described human flow prediction model learning process is executed, and the human flow prediction model table storage unit 122 of the storage unit 12 is subjected to the same processing. so that the human flow prediction model for predicting human stream data identified by the identifier I _i is stored.

In the above, an example in which the same processing as the above-mentioned human flow prediction model learning processing is executed in response to the reception of the re-learning request notification has been described, but this is only an example, for example, the re-learning request notification. Depending on the receipt, the process using the above-mentioned human flow prediction model update process, relational model learning process, relational model update process, human flow data prediction process based on the relational model, variable automatic selection process, etc. is executed. You may.

(10) People flow prediction data output processing First, the control unit 11 of the people flow prediction device 1 determines the future time t and the identifier _If of the people flow data to be output under the control of the output unit 114. However, these may be described in the human flow prediction device 1 as a setting file, or the operation signal or the like input by the input unit 4 may be notified to the output unit 114.

Subsequently, the control unit 11, under the control of the output unit 114, refers to the human flow prediction data storage unit 124 or sensor data table storage unit 121 of the storage unit 12, at time t, pedestrian flow prediction is identified by an identifier I _f Check for data.

Under the control of the output unit 114, the control unit 11 outputs the person flow prediction data to the display unit 5 when the person flow prediction data exists, and outputs a NULL value when the person flow prediction data does not exist. Output. The destination from which the person flow prediction data is output may be some device other than the display unit 5.

(effect)
As described in detail above, in the first embodiment of the present invention, the following effects can be achieved.

(1) Under the control of the data acquisition unit 111, the human flow sensor in the sensor 2, the external sensor in the sensor 2, and the BEMS 3 receive the human flow data, the external data, and the BEMS data related to the target section in the target space, respectively. To be acquired. Under the control of the human flow prediction model learning function unit 1121, the past human flow data is based on the human flow data acquired in the past predetermined period related to the target section, and the external data and BEMS data acquired in the past predetermined period. A people flow prediction model for predicting future people flow data is learned from. After that, under the control of the human flow data prediction unit 1123, the future human flow data related to the target section is predicted based on the learned human flow prediction model, the acquired human flow data, the acquired external data, and the BEMS data. Will be done.

By using the flow prediction model learned in this way, for example, a period including a period in which there is a discontinuity in the flow of people such as outside the business hours of the facility, for example, a period of one day to several weeks. However, it is possible to predict the flow of people.

In addition, in this way, it is possible to have a great influence on the flow of people in a predetermined section, for example, in a commercial complex, a) an event held in the facility, b) replacement of stores in the facility, c) seasonal products. Learning in consideration of the influence of various external factors such as d) transportation network delays and temporary flight conditions, e) outdoor weather, etc. in product sales, etc., or facilities directly connected to public transportation By using the created human flow prediction model, it is possible to predict future human flow, etc., taking into consideration the effects of these external factors.

(2) Under the control of the human flow prediction model learning function unit 1121, the human flow prediction model is dynamically updated by supervised learning using the acquired human flow data related to the target section.

For this reason, even if the accuracy of prediction by the human flow prediction model gradually declines due to the influence of various external factors, the human flow prediction model will reflect the influence of external factors by taking into account the appropriate number of updates. It can be updated each time to the one that has been made, and by using the person flow prediction model, it is possible to continue to accurately predict the future people flow and the like.

(3) Under the control of the human flow prediction model learning function unit 1121, the future total traffic related to the target space is based on the data representing the total traffic volume related to the target space that can be calculated from the past human flow data related to the target section. The above-mentioned human flow prediction model that predicts the data representing the quantity is learned. Further, under the control of the measuring instrument complement function unit 1122, a relational model representing the relationship between the data representing the total traffic volume and the human flow data related to the target section is generated by supervised learning. After that, under the control of the person flow data prediction unit 1123, the future total traffic amount related to the target space is predicted based on the person flow prediction model, and the relational model and the future total traffic related to the predicted target space are predicted. Future traffic data for the target plot is predicted based on the quantity.

Therefore, for example, even if the sensor of the target section does not operate normally due to a failure or the like, it is possible to continuously predict the future flow of people related to the target section. Further, even when the measurement accuracy of the human flow sensor is low, it is possible to predict the future human flow data with improved accuracy as compared with the case of predicting the future human flow data without passing through the total traffic volume.

(4) Under the control of the variable automatic selection unit 113, the feature amount of the data is calculated from the data acquired by the data acquisition unit 111, and the correlation between the calculated feature amount and the acquired human flow data is shown. By executing the statistical analysis process using the values, the feature quantities satisfying the predetermined conditions are selected. Then, under the control of the human flow prediction model learning function unit 1121, the human flow prediction model using the selected feature amount as an explanatory variable for predicting future human flow data is learned.

Therefore, in order to predict the flow of people and the like, it is possible to select an explanatory variable that contributes to improving the accuracy of the prediction of the flow of people and the like from the acquired data.

(5) Under the control of the difference calculation unit 1114, the value indicated by the person flow data related to the acquired target section and the value indicated by the person flow data predicted in advance for the target section at the time when the person flow data was acquired. The difference is calculated. When the calculated difference value is larger than a predetermined threshold value, at least the human flow data acquired during the period up to the time when the human flow data is acquired under the control of the human flow prediction model learning function unit 1121. Based on this, the human flow prediction model is retrained.

Therefore, when the predicted value of the human flow or the like deviates greatly from the actual event, it is possible to relearn the human flow prediction model suitable for the current use.

[Other Embodiments]
The present invention is not limited to the first embodiment. For example, in the first embodiment described above, the person flow prediction device is illustrated and described as one device, but each functional unit included in the person flow prediction device is spatially distributed and can communicate via a network. It may be provided in any device among a plurality of devices connected to the state. Further, each functional unit included in the human flow prediction device may be configured as software on a computer on a network such as a cloud or a virtual machine using virtual technology.

In addition, the types and configurations of devices such as the human flow prediction device, and the configurations of information stored in the sensor data table, the human flow prediction model table, and the relational model table are also variously modified without departing from the gist of the present invention. It is possible to carry out.

In short, the present invention is not limited to the first embodiment as it is, and at the implementation stage, the components can be modified and embodied within a range that does not deviate from the gist thereof. In addition, various inventions can be formed by an appropriate combination of the plurality of components disclosed in the first embodiment. For example, some components may be deleted from all the components shown in the first embodiment. In addition, components from different embodiments may be combined as appropriate.

1 ... People flow prediction device, 11 ... Control unit, 111 ... Data acquisition unit 1111 ... People flow data acquisition unit 1112 ... External data acquisition unit 1113 ... BEMS data acquisition unit 1114 ... Difference calculation unit, 112 ... People flow prediction unit, 1121 ... Human flow prediction model learning function unit 1122 ... Measuring instrument complement function unit 1123 ... Human flow data prediction unit, 113 ... Variable automatic selection unit, 114 ... Output unit, 12 ... Storage unit, 121 ... Sensor data table storage unit, 122 ... Human flow prediction model table storage unit, 123 ... Relationship model table storage unit, 124 ... Human flow prediction data storage unit, 13 ... External interface unit, 14 ... Input / output interface unit, 2 ... Sensor, 3 ... BEMS, 4 ... Input unit, 5 ... Display

Claims (7)

It is a human flow prediction device that predicts the human flow data related to the target section in the target space in the target space through which people pass.
A storage medium that stores a human flow prediction model for predicting future human flow data from past human flow data,
A data acquisition unit that acquires human flow data related to the target section, and
A human flow data prediction unit that predicts future human flow data related to the target section based on the human flow prediction model and the acquired human flow data.
The feature amount of each data was calculated from a plurality of past human flow data acquired by the data acquisition unit, and a value representing the partial autocorrelation between the calculated feature amount and the acquired human flow data was used. A variable automatic selection unit that selects features that satisfy predetermined conditions by executing statistical analysis processing,
A human flow prediction device including a human flow prediction model learning function unit that learns the human flow prediction model as an explanatory variable for predicting future human flow data using the selected feature amount . The data acquisition unit includes external data including at least one of information representing a physical phenomenon related to the target space and information representing the implementation status of an event, and energy management data related to energy management in a facility in the target space. Get at least one more of
The variable automatic selection unit is at least the feature amount calculated from the human flow data acquired in the past predetermined period related to the target section, the external data acquired in the past predetermined period, and the energy management data. Based on the feature amount calculated from one of them, the feature amount satisfying the predetermined condition is selected by the statistical analysis process.
The human flow prediction model learning function unit learns the human flow prediction model by using each of the selected feature quantities as an explanatory variable for predicting future human flow data .
The human flow data prediction unit is based on the learned human flow prediction model, the acquired human flow data, and at least one of the acquired external data and the energy management data, and the future related to the target section. The person flow prediction device according to claim 1, which predicts the person flow data of the above. The human flow prediction device according to claim 1 or 2, wherein the human flow prediction model learning function unit updates the human flow prediction model by supervised learning using the acquired human flow data related to the target section. A difference calculation unit that calculates the difference between the acquired value indicated by the person flow data related to the target section and the value indicated by the person flow data predicted in advance for the target section at the time when the person flow data was acquired. Further prepare
The human flow prediction model learning function unit learns the human flow prediction model based on at least the human flow data acquired during the period up to the time when the calculated difference value is larger than a predetermined threshold value. The person flow prediction device according to any one of claims 1 to 3, which is to be corrected. It is a human flow prediction device that predicts the human flow data related to the target section in the target space in the target space through which people pass.
A storage medium that stores a human flow prediction model for predicting future human flow data from past human flow data,
A data acquisition unit that acquires human flow data related to the target section, and
A human flow data prediction unit that predicts future human flow data related to the target section based on the human flow prediction model and the acquired human flow data.
A human flow prediction model learning function unit that dynamically updates the human flow prediction model using at least the acquired human flow data,
A complementary function unit that generates a relational model representing the relationship between the data representing the total traffic volume and the human flow data related to the target section by supervised learning.
Bei to give a,
The human flow prediction model learning function unit represents the future total traffic volume related to the target space based on the data representing the total traffic volume related to the target space that can be calculated from the past human flow data related to the target section. Learn the above-mentioned human flow prediction model that predicts data,
The human flow data prediction unit predicts the future total traffic volume of the target space based on the human flow prediction model, and is based on the relational model and the future total traffic volume of the predicted target space. A human flow prediction device that predicts future human flow data related to the target section. By arranging each component of the person flow prediction device according to any one of claims 1 to 5 in a distributed manner and connecting the component members in a communicable state via a network, the person flow prediction that predicts the future person flow data. system. A program that causes a computer to function as each part of the human flow prediction device according to any one of claims 1 to 5 .