JP6151202B2 - Program, apparatus, and method for estimating user attribute from power consumption in feature time zone - Google Patents

Description

  The present invention relates to a user information estimation technique that analyzes a user-related amount such as power consumed by a user and estimates predetermined information about the user.

  2. Description of the Related Art In recent years, attention has been focused on systems that provide support and advice related to life and work, such as household management support and power saving advice, according to user attributes of the user, such as the number of household members and household member information.

  As an example of such a system, Patent Literature 1 discloses a household management support system capable of providing more accurate and accurate advice on household expenditure in consideration of income disparity. In this system, after registering user information such as household category, number of households, prefecture name, and annual income, and daily household balance, the result of comparing the household balance with the calculated balance indicator is output.

  Patent Document 2 discloses a power saving support system that determines power saving advice based on the power usage measured by the power measurement sensor and information on the environment output from the environment sensor and presents it to the user. Yes. Here, an embodiment is also disclosed in which an inquiry about power saving advice is presented to the user, personal attribute information is input from the user as part of an answer to the inquiry, and the advice information provided is improved based on the personal attribute information. ing.

JP 2010-79760 A JP 2011-66956 A

  However, the prior arts disclosed in Patent Documents 1 and 2 rely on self-reporting by the user to obtain user attributes used when generating and improving advice. That is, the user himself / herself needs to explicitly input user attributes to the system.

  In the marketing field, there are many systems that analyze customer information. However, user attributes that are the basic information for the analysis, such as the number of households and the characteristics of household members, are subject to questionnaire surveys on households. Must go and get directly from customers.

  As described above, acquisition of user attributes necessary for estimating information related to service provision to a user has conventionally been dependent on input by each user or a questionnaire survey, which has been a laborious and time-consuming operation. .

  Therefore, the present invention provides a user attribute estimation program, apparatus, and method capable of estimating the user attribute of the user with higher reliability, particularly based on the amount of power consumed by the user as the life-related quantity. Objective.

According to the present invention, there is provided a user attribute estimation program for causing a computer mounted on a device that estimates a user attribute of an estimation target user based on a power consumption amount of the estimation target user to function.
Weather data management means for acquiring weather-related values indicating information related to the weather in the area related to the user in a predetermined period;
Based on the power consumption data for the user measured during a given period, the device-derived power amount is calculated based on the minimum value of the power consumption amount in each time zone, and the power consumption amount and device-derived power in that time zone are calculated. Active power extraction means for calculating the amount of active power corresponding to the amount of power related to the user's activity based on the amount;
An active power vector having an active power amount in each time zone as an element is generated from a user's active power amount in a predetermined period for each unit period constituting the predetermined period, and the active power amount is characterized according to the weather. A clustering process is performed to select an active power vector related to a unit period in which a weather-related value has a predetermined value in a characteristic time zone that can occur, and to perform a clustering process to classify the selected active power vector into a predetermined number of clusters. Representative vector generation means for generating a characteristic time representative vector representing the amount of active power in the characteristic time zone of the user based on the active power vector belonging to
Feature vector generation means for generating a feature vector including, as elements, an element of an active power vector in one unit period of the user, an element of a feature time representative vector generated for the user, and a weather-related value in one unit period ,
An estimation model construction means for constructing an estimation model by performing machine learning using a teacher data set including a feature vector and a user attribute value indicating a user attribute of a user related to the feature vector as teacher data;
A feature vector including an element of an active power vector in one unit period of the estimation target user, an element of a feature time representative vector generated for the estimation target user, and a weather-related value in the one unit period is constructed. There is provided a user attribute estimation program that causes a computer to function as user attribute estimation means that inputs to the estimation model and acquires an estimated value of the user attribute of the estimation target user.

  As one embodiment of the user attribute estimation program according to the present invention, the representative vector generation means sets an evening zone, which is a predetermined time during which the amount of sunlight before sunset or sunset falls as a characteristic time zone, and in the evening zone It is also preferable to select an active power vector related to a unit period indicating a weather-related value corresponding to sunny or sunshine.

  Further, in the user attribute estimation program according to the present invention, the representative vector generation means sets a characteristic time zone composed of a plurality of time zones, and a plurality of representative power amounts in each of the plurality of time zones. It is also preferable to generate a feature time representative vector having elements.

  Furthermore, as another embodiment of the user attribute estimation program according to the present invention, one day is set as the unit period, and the feature vector generation means includes a daily value indicating whether or not this one day is a weekday as a further element. It is also preferable to generate a feature vector.

Furthermore, as still another embodiment of the user attribute estimation program according to the present invention, the weather data management means further obtains an air temperature value indicating the temperature in the area related to the user in a predetermined period, and the user attribute estimation program includes: further,
An average active power calculating means for calculating an average active power amount that is an average of active power amounts in each time zone for a plurality of users in a neighboring area considered to be under the same weather condition;
Based on the temperature value in the time zone related to the calculated average active power amount, the computer functions as an active power estimation model construction means for constructing an active power estimation model that models the relationship between the temperature and the active power amount,
The feature vector generation means, for the generated feature vector, from the temperature value in each time zone in the unit period related to the feature vector, using the activity power estimation model, the general activity power amount that is the activity power amount in the feature time zone And calculating the similarity between the calculated general activity power amount in the feature time zone and the element of the active power vector in the feature time zone of the feature vector, and the reliability is calculated according to the similarity. It is also preferable to exclude feature vectors that are considered low from the record for user attribute estimation.

Furthermore, as still another embodiment of the user attribute estimation program according to the present invention, the weather data management means further accumulates an air temperature value indicating an air temperature in the area related to the user in a predetermined period,
The user attribute estimation program further includes:
Teacher data generation means for generating a teacher data set from user attribute declaration or survey data;
Average active power calculating means for calculating an average active power amount that is an average of active power amounts in each time period for a plurality of neighboring users considered to be under the same weather condition;
Based on the temperature value in the time zone related to the calculated average active power amount, the computer functions as an active power estimation model construction means for constructing an active power estimation model that models the relationship between the temperature and the active power amount,
The teacher data generation means is a feature vector estimated by calculating the amount of active power in each time zone using the active power estimation model from the temperature value in each time zone in the unit period related to the acquired user attribute. It is also preferable to generate an estimated feature vector and generate teacher data from the estimated feature vector and the acquired user attribute.

According to the present invention, the apparatus further estimates the user attribute of the estimation target user based on the power consumption of the estimation target user,
Weather data management means for acquiring weather-related values indicating information related to the weather in the area related to the user in a predetermined period;
Based on the power consumption data for the user measured during a given period, the device-derived power amount is calculated based on the minimum value of the power consumption amount in each time zone, and the power consumption amount and device-derived power in that time zone are calculated. Active power extraction means for calculating the amount of active power corresponding to the amount of power related to the user's activity based on the amount;
An active power vector having an active power amount in each time zone as an element is generated from a user's active power amount in a predetermined period for each unit period constituting the predetermined period, and the active power amount is characterized according to the weather. The number of vectors to which clustering is performed to select an active power vector related to a unit period in which a weather-related value indicates a predetermined value in a feature time zone that can occur, and to classify the selected active power vector into a predetermined number of clusters. Representative vector generation means for generating a feature time representative vector representing the amount of active power in the feature time zone of the user based on the active power vector belonging to the cluster with the largest number of
Feature vector generation means for generating a feature vector including, as elements, an element of an active power vector in one unit period of the user, an element of a feature time representative vector generated for the user, and a weather related value in the one unit period When,
An estimation model construction means for constructing an estimation model by performing machine learning using a teacher data set including a feature vector and a user attribute value indicating a user attribute of a user related to the feature vector as teacher data;
A feature vector including an element of an active power vector in one unit period of the estimation target user, an element of a feature time representative vector generated for the estimation target user, and a weather-related value in the one unit period is constructed. There is provided a user attribute estimation device having user attribute estimation means for inputting to the estimation model and acquiring an estimated value of the user attribute of the estimation target user.

According to the present invention, there is further provided a user attribute estimation method by a computer mounted on a device that estimates a user attribute of the estimation target user based on the power consumption of the estimation target user,
A first step of acquiring a weather-related value indicating information relating to weather in a region related to the user in a predetermined period;
Based on the power consumption data for the user measured during a given period, the device-derived power amount is calculated based on the minimum value of the power consumption amount in each time zone, and the power consumption amount and device-derived power in that time zone are calculated. A second step of calculating an amount of active power corresponding to the amount of power related to the user's activity based on the amount;
An active power vector having an active power amount in each time zone as an element is generated from a user's active power amount in a predetermined period for each unit period constituting the predetermined period, and the active power amount is characterized according to the weather. A clustering process is performed to select an active power vector related to a unit period in which a weather-related value has a predetermined value in a characteristic time zone that can occur, and to perform a clustering process to classify the selected active power vector into a predetermined number of clusters. A third step of generating a feature time representative vector representing the amount of active power in the feature time zone of the user based on the activity power vector belonging to
A fourth step of generating a feature vector including, as elements, an element of an active power vector in one unit period of the user, an element of a feature time representative vector generated for the user, and a weather-related value in the one unit period When,
A fifth step of constructing an estimation model by performing machine learning using a teacher data set including, as teacher data, a set of a feature vector and a user attribute value indicating a user attribute of a user related to the feature vector;
A feature vector including an element of an active power vector in one unit period of the estimation target user, an element of a feature time representative vector generated for the estimation target user, and a weather-related value in the one unit period is constructed. And a sixth step of obtaining an estimated value of the user attribute of the estimation target user.

  According to the user attribute estimation program, apparatus, and method of the present invention, the user attribute of the user can be estimated with higher reliability based on the power consumption by the user.

It is a schematic diagram in one Embodiment of the user attribute estimation system by this invention. It is a functional block diagram which shows the function structure in one Embodiment of the user attribute estimation apparatus by this invention. It is explanatory drawing which shows one Embodiment of a feature vector, and the table which shows one Embodiment of the relationship between the feature vector corresponding to a user attribute. It is a graph which shows one Embodiment of an active power estimation model. It is a table which shows the Example which provides a reliability flag to a feature vector. 5 is a flowchart illustrating an embodiment of a user attribute estimation method according to the present invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[User attribute estimation system]
FIG. 1 is a schematic diagram of an embodiment of a user attribute estimation system according to the present invention.

  In FIG. 1, the user attribute estimation system according to the present embodiment is installed in the carrier communication network 6 and determines the user attribute of the user based on (a) the user's power consumption and (b) the current day's weather information. -The user attribute estimation apparatus 1 to be estimated is included. Here, in this embodiment, a user is one household (user household) or a member of the household. Furthermore, the user attribute is a static attribute in the present embodiment, and refers to a living behavior characteristic of the household member in the “characteristic time zone”.

  In addition, the “characteristic time zone” is a time zone in which a feature can be generated in the amount of active power according to the weather. In this embodiment, the evening time zone (evening), that is, 16:00 to 19:00 (16:00, Time zone and 18:00 hour). Actually, the evening zone is a time zone in which the amount of sunlight before sunset or before sunset is reduced, and by reducing the amount of sunlight in this way, an action to turn on lighting in each household can occur in the evening zone. It is.

  Further, as user attributes related to this evening zone (characteristic time zone), for example, “evening at home” in which the household member is at home in the evening zone, “evening” when the household member is out in the evening zone “Out-of-house type”, “Return-to-dinner type” where the household member returns home before dinner, “House-at-home and lighting in the evening” "Patience type".

  Similarly, according to FIG. 1, the power consumption of the user (one household) is transmitted from the home gateway (HGW) 5 installed in the home, which is the stay target area of the household members, through the carrier communication network 6 in this embodiment. Is transmitted to the user attribute estimation device 1. Here, the power consumption may be measured by the smart meter 4, and the measured value may be output to the HGW 5 via, for example, a HEMS (Home Energy Management System). Moreover, as a change mode, the smart meter 4 can transmit power consumption to MDMS (Meter data Management System) (not shown), and the user attribute estimation apparatus 1 can also acquire power consumption from this MDMS. It is.

  The carrier communication network 6 can be a fixed access network such as an optical fiber network or an ADSL (Asymmetric Digital Subscriber Line). In addition, wireless LAN (Local Area Network) such as Wi-Fi (registered trademark), WiMAX (Worldwide Interoperability for Microwave Access), LTE (Long Term Evolution) or 3G (3rd Generation) or other wireless access network may be used. Is possible.

  The smart meter 4 measures the amount of power consumed by the user at home, and outputs a measured value of the amount of power consumed per unit time to the HGW 5 via, for example, HEMS. Here, instead of the smart meter, a CT sensor installed on a distribution board, a clamp meter (overhead ammeter), or an outlet tap (tap wattmeter) installed at an outlet may be used. In this case, the individual power consumption of in-house power distribution is measured by a CT sensor, an outlet tap, etc., and the measured value of the power consumption is output to the HGW 5.

  The HGW 5 is a device that relays and controls communication between a home network provided in a home, a set top box (not shown) (STB), and the like, and the carrier communication network 6. For example, a personal computer (PC) is connected to the home network via a wireless access point (AP). The HGW 5 further transmits the measured value of the power consumption per unit time acquired from the smart meter 4 to the user attribute estimation device 1 via the carrier communication network 6.

  The user information management device 2 is installed, for example, in the carrier communication network 6 and receives user attribute information declarations from each user (household) in the form of answering a questionnaire survey, for example. Store and manage in association with the ID. The user inputs, for example, static attributes such as his / her home / outing situation and lighting usage situation in the evening zone at a predetermined date and time to the user information management apparatus 2 from the HGW 5 via the carrier communication network 6. May be. The user attribute estimation device 1 acquires user attribute information associated with the user ID from the user information management device 2.

  The weather information management server 3 provides the past, present, or future (forecast) weather information such as weather, temperature, illuminance, humidity, and rainfall for each date and time of a predetermined period via the Internet 7 and the operator communication network 6. It transmits to the attribute estimation apparatus 1 addressed. In particular, it is also preferable that weather information including a weather-related value indicating information related to the weather is transmitted, for example. Here, the weather-related value may include a weather value that distinguishes “sunny”, “cloudy”, or “rain / snow”, a sunshine value that indicates the amount of sunlight, a temperature value, and the like.

Specifically, the user attribute estimation device 1
(A) “active power vector” element corresponding to the amount of power related to the user's activity in a certain day as a unit period, and the user's active power amount in the evening zone as the “characteristic time zone” Generate a “feature vector” that includes the elements of the “feature time representative vector” and the weather-related values for this one day as elements,
(B) Machine learning is performed using a teacher data set including, as teacher data, a set of a “feature vector” and a user attribute value of the user related to the feature vector (life behavior characteristics of the household member in the evening) Build an estimation model,
(C) Features including an element of “active power vector” for one day of the estimation target user, an element of “feature time representative vector” generated for the estimation target user, and a weather-related value for the one day The vector is input to the constructed estimation model, and the estimated value of the user attribute of the estimation target user is acquired.

Thus, in the user attribute estimation device 1, the “feature vector” constituting the teacher data is
(A) An element of a “characteristic time representative vector” representing the user's active energy during the evening,
(B) Includes weather-related values. Here, the evening zone is a time zone in which the characteristics of each household can appear prominently in the amount of active power depending on the weather. Specifically, lighting is turned on in each household due to a decrease in the amount of sunlight. It is a time zone in which behavior can occur characteristically. Therefore, when the teacher data including the combination of (a) and (b) is used for machine learning, the characteristics of the user (household) are more clearly reflected in the output from the estimated model to be constructed. An accurate estimate is obtained.

  In addition, once such an estimation model is generated, user attributes can be estimated more accurately and with high reliability based on power consumption by the user without relying on user input or questionnaire surveys. it can. In particular, it is possible to accurately and accurately estimate living behavior characteristics in the user's characteristic time zone (evening time zone).

[User attribute estimation device 1]
FIG. 2 is a functional block diagram showing a functional configuration in an embodiment of the user attribute estimation apparatus according to the present invention.

  According to FIG. 2, the user attribute estimation device 1 includes a communication interface unit 101 and a processor memory. Here, the processor memory realizes its function by executing a program for causing the computer installed in the user attribute estimation apparatus 1 to function.

  Here, the processor memory includes a power consumption data management unit 102, a weather data management unit 103, a user attribute management unit 104, an activity power extraction unit 111, a representative vector generation unit 112, and a feature vector generation unit 113. The teacher data generation unit 114, the user attribute estimation model construction unit 115, the user attribute estimation unit 116, the average activity power calculation unit 121, and the activity power estimation model construction unit 122. In addition, according to FIG. 2, the flow of the process which connected each function structure part with the arrow is understood also as one Embodiment of the user attribute estimation method by this invention.

The communication interface unit 101
(A) The measured power consumption data is received from the smart meter 4 or the HGW 5 installed in the user's home (stay area), and collected from the user information management device 2 by the user ID and declaration, etc. Receiving information on a set of user attributes, and receiving weather information including weather-related values at each date and time in a predetermined period from the weather information management server 3,
(B) The estimated user attribute value of the estimation target user household output from the user attribute estimation unit 116 is transmitted to an external notification destination. For example, the estimated value can be presented via a website or application.

  The power consumption data management unit 102 accumulates and manages power consumption data generated from the received measurement value of power consumption. Note that the power consumption data may be accumulated outside the apparatus 1, for example, in a storage in the user's home.

  Here, a 24-dimensional power consumption vector is generated from the power consumption in the hourly time zone constituting one day (basic period), and 90 power consumptions corresponding to each day in a predetermined period, for example, 90 days. Vectors may be acquired and stored / managed. At this time, the power consumption amount in each time zone may be an integrated value in the one hour, or when the raw data is, for example, a 15-minute value, the center of the four values in the one hour. It may be a value. Furthermore, one time slot | zone is not limited to 1 hour, For example, it is good also as 2 hours, 3 hours, etc. In addition, time zones having different time lengths may be set.

  The weather data management unit 103 acquires a weather-related value indicating the weather in the area related to the user household in each time zone in each day for a predetermined period, for example, 90 days, for example, “sunny” or “rain / snow”, Accumulate and manage. Furthermore, it is also preferable to further acquire, store and manage the temperature value indicating the temperature in the area related to the user household during the predetermined period.

  The user attribute management unit 104 acquires, stores, and manages information on a set of a user ID and a user attribute. Here, the user ID is an identification number of a user household or household member. In addition, in the present embodiment, the user attributes are the life behavior characteristics of household members in the evening zone as the characteristic time zone, for example, “evening home type”, “evening going out type”, “homecoming before dinner”, “home evening” And “lighting endurance type”.

The active power extraction unit 111
(A) The minimum value of the power consumption in each time zone is determined from the data of the power consumption of the user household measured for a predetermined period, for example, 30 days, and this minimum value is determined for the device in each time zone. The amount of power
(B) The active power amount is calculated based on the power consumption amount and the device-derived power amount in the time period.

  Here, the device-derived power amount is a power consumption amount mainly generated from the device other than the user-derived power amount generated when the member of the user household intentionally performs the ON operation or use operation on the device. is there. Standby power, background power consumption, and the like correspond to this. For example, the minimum value among the 30 data of the power consumption amount at 16:00 in 30 days is set as the device-derived power amount at 16:00 in the user household. It should be noted that the device-induced power amount may vary due to changes in seasons and living conditions, and it is also preferable that the device-induced power amount is updated in accordance with the user attribute estimation process or periodically.

Furthermore, the amount of active power is the amount of power related to the activity of the user household. It is calculated using the following formula in each time zone.
(1) (Activity energy) = (Total energy consumption)-(Equipment-induced energy)
Here, when the calculated amount of active power is equal to or greater than a predetermined threshold, it is possible to interpret that a person (household member) is active in the user household in the time period. This threshold value may be different for each season, for example, summer, winter, and summer / winter.

First, the representative vector generation unit 112
(A) For each day constituting the 90 days, an active power vector having the active power amount in each time zone as an element is generated from the active power amount in a predetermined period of the user household, for example, 90 days,
(B) An active power vector for one day is selected such that the weather-related value indicates a predetermined value (in this embodiment, “sunny” or “sunshine”) in the evening zone (characteristic time zone).

  Here, when the time zone is set to 1 hour, the active power vector is a 24-dimensional vector having the active power amount in each time zone as an element on each day of 90 days. As a result, a total of 90 24-dimensional active power vectors (records) are generated. Further, if the weather-related value in the evening zone in the 90 days is “sunny” or “daylight present” is 70 days, a 24-dimensional active power vector corresponding to each of these 70 days out of these 90 records, That is, 70 records are selected.

The representative vector generation unit 112 then
(C) performing a hierarchical clustering process for classifying the selected active power vector into a predetermined number of clusters;
(D) Based on the active power vector belonging to the largest cluster, that is, the largest cluster, an evening band (feature time) representative vector (lighting ON time representative vector) representing the amount of active power in the evening band of the user household. Generate.

  This clustering process is performed for the selected 70 records in the above example. Here, the number of clusters is set to a fixed value, for example, 5 clusters. Alternatively, the number of clusters may be determined using a statistic criterion such as AIC (Akaike's Information Criteria) or BIC (Bayesian Information Criteria). Note that if clustering processing is performed using all active power vectors regardless of weather-related values, records of a plurality of weather-related values are mixed in each cluster. On the other hand, by selecting the active power vector in advance based on the weather-related value as described above, it is possible to reliably generate an evening zone representative vector having a desired weather value (“sunny”) characteristic. Become.

The representative vector generation unit 112 sets the cluster having the maximum number of element records as the representative cluster of the user household after the clustering process is completed. The evening zone representative vector is a representative (average) vector of this representative cluster. For example, the evening zone is composed of three time zones of 16:00, 17:00 and 18:00, and the elements (power consumption) of representative vectors in these time zones are l1, l2 and l3, respectively. Then
(2) (Evening zone representative vector) = (l1, l2, l3)
It becomes.

  The feature vector generation unit 113 includes (a) an element of an active power vector for one day (unit period) of a user household, and (a) an evening zone (feature time) representative vector (lighting ON time representative) generated for the user household. A feature vector including elements of (vector) and (c) weather-related values in the one unit period as elements is generated. This feature vector preferably further includes a daily value indicating whether or not the one day is a weekday as a further element. Actually, for example, whether or not you are at home in the evening zone, and when to turn on the lighting, often differs between weekdays and holidays. For this reason, the daily value is an effective parameter for improving the accuracy of user attribute estimation described later.

  FIG. 3 is an explanatory diagram showing an embodiment of a feature vector, and a table showing an embodiment of a relationship between a user attribute and a corresponding feature vector.

According to the embodiment shown in FIG. 3A, the feature vector is
(A) Active power vector elements: a1, a2, a3, a4, a5, a6, a7,
(A) Evening zone representative vector elements: l1, l2, l3,
(C) Weather-related value vector elements: w1, w2, w3, w4, and (d) Daily value: h
Is a 15-dimensional vector.

  Here, a1, a2, a3, a4, a5, a6, and a7 of (a) are 0-5 o'clock, 6-10 o'clock, 11-15 o'clock, It is power consumption in the hour zone, the 17:00 zone, the 18:00 zone, and the 19-23 hour zone. Thus, a3, a4 and a5 indicate the power consumption in the evening zone. L1, l2, and l3 in (a) are elements of the vector shown in the above-described equation (2). W1, w2, w3, and w4 in (c) are the weather-related values (“Sunny” or “Rain” at 0-7 o'clock, 8-15 o'clock, 16-18 o'clock, and 19: 00-23: 00, respectively.・ Snow ”). Thus, w3 indicates the weather-related value in the evening zone. Further, h in (D) is a daily value indicating whether the day is a weekday or a holiday.

  Next, FIG. 3B shows an example of feature vectors corresponding to user attributes. According to the figure, the user attribute of the user 0001 is “evening at home”. Here, when both the feature vector whose weather related value w3 in the evening zone is “sunny” and the feature vector which is “rain” in the user 0001 are seen, the elements (l1, l2, l3) of the evening zone representative vector are Over time, it has increased from 0.47 (kWh) → 0.60 → 0.71. As a result, it is understood that the “evening at home type” tends to turn on the indoor lighting when it becomes dim in the evening.

  In addition, in the feature vector of “sunny” in the evening zone for the same user 0001 (“evening at home”), the amount of active power (a4, a5, a6) in the evening zone is 0.40 (kWh) → 0 over time. It has changed from .57 to 0.68. Accordingly, it is understood that, in the “evening at home type”, the power consumption increases in the evening zone (to turn on the indoor lighting) on a sunny day. On the other hand, in the feature vector of “rain” in the evening zone, the amount of active power (a4, a5, a6) in the evening zone changes from 0.60 (kWh) → 0.67 → 0.78 over time. . As a result, it is understood that in the “evening at home type”, the power consumption increases early on a rainy day, so that the increase in power consumption in the evening zone is relatively small.

  Furthermore, the user attribute of the user 0002 is “at home in the evening and lighting endurance type”. Here, when both the feature vector whose weather-related value w3 in the evening zone is “sunny” and the feature vector which is “rain” in the user 0002 are seen, the elements (l1, l2, l3) of the evening zone representative vector are Over time, it has increased from 0.47 (kWh) → 0.49 → 0.71. As a result, it is understood that the “evening at home and lighting endurance type” tends not to turn on the lighting in the room even when it becomes dim in the evening.

  In addition, in the feature vector of “sunny” in the evening zone for the same user 0002 (“evening at home and lighting endurance type”), the amount of active power (a4, a5, a6) in the evening zone is 0.40 (kWh) over time. ) → 0.47 → 0.68. As a result, it is understood that the increase in power consumption is relatively small in the evening zone on a sunny day in the “evening at home and lighting endurance type”. On the other hand, in the feature vector “rain” in the evening zone, the amount of active power (a4, a5, a6) in the evening zone changes from 0.45 (kWh) → 0.57 → 0.78 over time. . Thus, it is understood that in the “evening at home and lighting endurance type”, the increase in power consumption is faster on a rainy day than on a sunny day.

  Furthermore, the user attribute of the user 0003 is “evening out type”. Here, when both the feature vector whose weather-related value w3 in the evening zone is “sunny” and the feature vector which is “rain” in the user 0003 are seen, the elements (l1, l2, l3) of the evening zone representative vector are As time passes, it changes from 0.47 (kWh) to 0.35 to 0.42. As a result, it is understood that the “home in the evening and lighting endurance type” tends to not turn on the room lighting because it is absent even if it becomes dim in the evening.

  In addition, in the feature vector of “sunny” in the evening zone for the same user 0003 (“evening out type”), the amount of active power (a4, a5, a6) in the evening zone is 0.40 (kWh) → 0 over time. It has changed from .47 to 0.38. Further, even in the “rain” feature vector, the amount of active power (a4, a5, a6) in the evening zone changes from 0.40 (kWh) → 0.47 → 0.38 over time. As a result, it is understood that in the “evening out type”, the power consumption fluctuation in the evening zone is small regardless of the weather.

  As described above, the feature vector according to the present invention has features that sufficiently reflect user attributes with respect to power consumption and weather-related values. It is understood that the input value to is very good.

  Of course, the number of elements of the active power vector, the number of elements of the evening zone representative vector, and the number of elements of the weather-related value vector that constitute the feature vector are not limited to the above-mentioned numbers, and are set in the set time zone. The number of elements can be set accordingly. Furthermore, the weather-related values (elements of the weather-related value vector) constituting the feature vector are not limited to the weather values (for example, “sunny” or “rain / snow”). For example, a sunshine value indicating the amount of sunshine (for example, “large amount of sunshine”, “medium amount of sunshine”, or “small amount of sunshine”) may be employed as the weather-related value. Further, a form in which the daily value h is not included in the feature vector is also possible.

Returning to FIG. 2, the teacher data generation unit 114 generates a teacher data set from user attribute declarations or survey data. The teacher dataset
(A) a feature vector for a certain day (basic period) of a certain user household;
(B) A set of teacher data including a pair with a user attribute value indicating a user attribute of the user household. The teacher data set can be composed of 90000 teacher data corresponding to 90 days of 1000 households, for example. Here, when generating the teacher data, for example, the device ID of the HGW 5 that is the transmission source of the power consumption information and the user ID are associated with each other, and the power consumption used to generate the feature vector When the device ID and the user ID associated with the user attribute value indicate the same user (household), it is also preferable that the feature vector and the user attribute value are combined into teacher data.

  The user attribute estimation model construction unit 115 performs machine learning using the generated teacher data set to construct an estimation model. Various well-known methods such as a random forest method and a logistic regression analysis method can be applied to the machine learning algorithm used here.

The user attribute estimation unit 116 inputs a feature vector for one day (unit period) of the estimation target user household into the constructed estimation model, and acquires an estimated value of the user attribute of the estimation target user household. Here, the input feature vector is
(A) an element of an active power vector for one day (unit period) of an estimation target user household;
(B) an element of a characteristic time representative vector generated for the estimation target user household;
(C) the weather-related values for the day,
Is preferably included as an element. Furthermore, (d) the daily value (“weekdays” or “holiday”) for the day may be included as an element. For example, when each of the two feature vectors for the user 0001 shown in FIG. 3B is input to the constructed estimation model, the estimated value that the user attribute is “evening at home” as an output from the estimation model Will be acquired.

  In the embodiment described above, the evening zone is adopted as the characteristic time zone, and the user attribute is estimated on the assumption of the user's action to turn on the illumination in the evening zone. Is not limited to this. For example, it is possible to assume a predetermined time in the morning as the characteristic time zone and assume a user's action of starting the washing machine or the like on a sunny day.

  In addition, the user attributes estimated by the method of the present invention are not limited to the living behavior characteristics related to the evening zone. For example, since the feature vector includes elements of the active power vector (a1, a2, a3, a4, a5, a6, a7 in this embodiment), for example, “daytime home type” or “daytime home and It is also possible to estimate attributes such as “power saving type”.

  Below, the function of the average active power calculation part 121 and the active power estimation model construction part 122 is demonstrated. These functional components consider the average power consumption of households in the vicinity of the estimation target user household, select feature vectors, assist teacher data generation, and improve the reliability of the estimated values of user attributes to be acquired. Increases the effect. As an embodiment of the present invention, a functional configuration that does not employ the average active power calculation unit 121 and the active power estimation model construction unit 122 described below is naturally possible.

  The average active power calculating unit 121 calculates an average active power amount that is an average of active power amounts in each time period for a plurality of user households in a neighboring area considered to be under the same weather condition.

  The active power estimation model construction unit 122 constructs an active power estimation model that models the relationship between the temperature and the active power amount based on the calculated temperature value in the time zone related to the average active power amount. Here, the active power estimation model classifies a set of average active energy data by time zone, by weather-related values, and by daily values (weekdays / holidays), and then shows the relationship between temperature and active energy. It is also preferable to build together. For specific modeling, various well-known methods, such as a method using a spline curve, a method of defining a change point and performing regression analysis, and the like can be applied. The contents of the active power estimation model will be described below.

  FIG. 4 is a graph illustrating an embodiment of an active power estimation model.

  As shown in FIG. 4 (A), in the active power estimation model of the present embodiment, the temperature and the active power for each hour of each hour in one day (basic period) that is “sunny” and “weekdays”. The relationship is calculated. FIG. 4B shows a graph showing the relationship between the air temperature in the midnight range and the active power. The curve in the graph indicates “general activity power” that is the activity energy for each temperature value. In this way, by using the constructed active power estimation model, if the temperature value in the target time zone is known, it is the average active power amount of households in the neighboring area in that time zone. The “active energy” is estimated.

Returning to FIG. 2, the feature vector generation unit 113
(A) For the generated feature vector, from the temperature value in each time zone in one day (unit period) related to the feature vector, the activity power estimation model constructed by the activity power estimation model construction unit 122 is used to set the evening band. Calculate “general activity energy” in (characteristic time zone)
(B) calculating the degree of similarity between the calculated “general activity power amount” in the evening band and the element of the activity power vector related to the evening band among the feature vectors;
(C) A reliability flag = 0 is assigned to a feature vector that is considered to have low reliability according to the calculated similarity, and the feature vector is excluded from the record for user attribute estimation.

  FIG. 5 is a table showing an embodiment in which a reliability flag is assigned to a feature vector.

Hereinafter, an embodiment for obtaining the reliability of the feature vector generated by the feature vector generation unit 113 will be described. In each example of FIGS. 5A and 5B,
(A) elements (a4, a5, a6) of the active power vector in the evening zone (16, 17 and 18:00) in the feature vectors related to “sunny” and “weekdays”;
(B) Calculate cosine similarity with general activity energy (a4 ', a5', a6 ') in the same daytime ("Sunny" and "Weekday") evening zone (16, 17 and 18:00) Has been. When the calculated cosine similarity is less than a predetermined threshold (0.8 in this embodiment), a reliability flag = 0 is assigned to the feature vector of the reliability check target.

  In the embodiment of FIG. 5A, the cosine similarity is 0.99 (> 0.8), and the reliability flag = 1 is assigned to this feature vector. As a result, the feature vector is used as a record for user attribute estimation. On the other hand, in the embodiment of FIG. 5B, the cosine similarity is 0.78 (<0.8), and a reliability flag = 0 is assigned to this feature vector. As a result, the feature vector is excluded from the record for user attribute estimation. In this way, by comparing with the households in the general neighborhood area using the amount of general activity power, the abnormal or extraordinary day of power usage in the user household is excluded from the estimation material, and the estimated value Reliability can be increased.

  FIG. 6 is a flowchart illustrating an embodiment of a user attribute estimation method according to the present invention.

First, steps S601 to S606 are a learning process.
(S601) A weather-related value (for example, a value indicating “sunny” or “rain / snow”) in a predetermined period (for example, 90 days) is acquired.
(S602) The power consumption amount in each time zone in a predetermined period of the user household is acquired, and the device-derived power amount of the user household is calculated from the power consumption amount.
(S603) An active power vector of the user household is calculated from the calculated device-derived power amount.

(S604) The active power vector of the user household for a predetermined period is clustered to generate a representative vector for the characteristic time (evening band) of the user household.
(S605) A feature vector in one basic period (one day) of the user household is generated using the generated feature time representative vector.
(S606) An estimation model is constructed using a set of teacher data including a set of feature vectors and user attributes.

Next, steps S607 to S609 are an estimation process.
(S607) A feature vector related to the estimation target user household is generated.
(S608) The generated feature vector is input to the estimation model constructed in step S606.
(S609) The estimated value of the user attribute in the estimation target user household is acquired as an output from the estimation model.

  As described above in detail, according to the user attribute estimation program, the user attribute estimation device, and the user attribute estimation method of the present invention, (a) the user (household) active power amount in the characteristic time zone is represented by “ Teacher data is configured using “feature vectors” including elements of “feature time representative vectors” and (a) weather-related values. Here, the characteristic time zone is a time zone in which the characteristics of each household can appear in the active power amount according to the weather. Therefore, by using such teacher data for machine learning, the characteristics of the user (household) are more clearly reflected in the output from the estimated model to be constructed, and a more accurate estimated value can be acquired. . Furthermore, once such an estimation model is generated, user attributes can be estimated with higher accuracy and higher reliability based on the amount of power consumed by the user without relying on user input or questionnaire surveys.

  In addition, based on the life behavior characteristic in the characteristic time zone of the user (household) thus estimated, appropriate support information about the user's power consumption, for example, power saving advice and household management advice, and further, the life behavior characteristic It is also possible to provide product / service information suitable for the customer. This also makes it possible to realize energy saving and marketing that is more suitable for the needs of users (households).

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 User attribute estimation apparatus 101 Communication interface part 102 Power consumption data management part 103 Weather data management part 104 User attribute management part 111 Activity power extraction part 112 Representative vector generation part 113 Feature vector generation part 114 Teacher data generation part 115 User attribute estimation model Construction unit 116 User attribute estimation unit 121 Average activity power calculation unit 122 Activity power estimation model construction unit 2 User information management device 3 Weather information management server 4 Smart meter 5 HGW
6 Business communication network 7 Internet

Claims (7)

A user attribute estimation program that causes a computer mounted on a device that estimates a user attribute of the estimation target user based on the power consumption of the estimation target user to function,
Weather data management means for acquiring weather-related values indicating information related to the weather in the area related to the user in a predetermined period;
Based on the power consumption data of the user measured during the predetermined period, the device-derived power amount is calculated based on the minimum value of the power consumption amount in each time zone, and the power consumption amount in the time zone and the Active power extraction means for calculating an active power amount corresponding to the power amount related to the user's activity based on the device-derived power amount;
For each unit period constituting the predetermined period, an active power vector having the active power amount in each time zone as an element is generated from the active power amount of the user for the predetermined period, and the active power amount is determined according to the weather. A clustering process is performed to select an active power vector related to a unit period in which a weather-related value has a predetermined value in a characteristic time zone in which a feature can occur, and to classify the selected active power vector into a predetermined number of clusters. Representative vector generation means for generating a feature time representative vector representing the amount of active power in the feature time zone of the user based on the active power vector belonging to the cluster of
A feature vector for generating a feature vector including, as elements, an element of an active power vector in one unit period of the user, an element of a feature time representative vector generated for the user, and a weather-related value in the one unit period Generating means;
An estimation model construction means for constructing an estimation model by performing machine learning using a teacher data set including, as teacher data, a set of the feature vector and a user attribute value indicating a user attribute of the user related to the feature vector;
A feature vector including, as elements, an element of an active power vector in one unit period of the estimation target user, an element of a feature time representative vector generated for the estimation target user, and a weather-related value in the one unit period A user attribute estimation program for causing a computer to function as user attribute estimation means for inputting to a constructed estimation model and acquiring an estimated value of a user attribute of the estimation target user.   The representative vector generation means sets an evening zone, which is a predetermined time when the amount of sunlight before sunset or sunset falls as the characteristic time zone, and indicates a weather-related value corresponding to sunny or sunshine in the evening zone. The user attribute estimation program according to claim 1, wherein an active power vector related to a unit period is selected.   The representative vector generating means sets a characteristic time zone composed of a plurality of time zones, and generates a characteristic time representative vector having a plurality of elements each representing the amount of active power in each of the plurality of time zones. The user attribute estimation program according to claim 1 or 2. One day is set as the unit period,
4. The user according to claim 1, wherein the feature vector generation unit generates a feature vector including a daily value indicating whether or not the day is a weekday as an additional element. 5. Attribute estimation program. The weather data management means further acquires an air temperature value indicating an air temperature in the area related to the user in the predetermined period,
The user attribute estimation program further includes:
An average active power calculating means for calculating an average active power amount that is an average of active power amounts in each time zone for a plurality of users in a neighboring area considered to be under the same weather condition;
Based on the temperature value in the time zone related to the calculated average active power amount, the computer functions as an active power estimation model construction means for constructing an active power estimation model that models the relationship between the temperature and the active power amount,
The feature vector generation means is an active power amount in the feature time zone using the active power estimation model from the temperature value in each time zone in the unit period related to the feature vector for the generated feature vector Calculating the amount of active power, calculating the degree of similarity between the calculated general activity power amount in the feature time zone and the element of the active power vector related to the feature time zone of the feature vector, 5. The user attribute estimation program according to claim 1, wherein a feature vector that is regarded as having low reliability is excluded from a record for user attribute estimation. An apparatus for estimating a user attribute of an estimation target user based on a power consumption amount of the estimation target user,
Weather data management means for acquiring weather-related values indicating information related to the weather in the area related to the user in a predetermined period;
Based on the power consumption data of the user measured during the predetermined period, the device-derived power amount is calculated based on the minimum value of the power consumption amount in each time zone, and the power consumption amount in the time zone and the Active power extraction means for calculating an active power amount corresponding to the power amount related to the user's activity based on the device-derived power amount;
For each unit period constituting the predetermined period, an active power vector having the active power amount in each time zone as an element is generated from the active power amount of the user for the predetermined period, and the active power amount is determined according to the weather. A clustering process is performed to select an active power vector related to a unit period in which a weather-related value has a predetermined value in a characteristic time zone in which a feature can occur, and to classify the selected active power vector into a predetermined number of clusters. Representative vector generation means for generating a feature time representative vector representing the amount of active power in the feature time zone of the user based on the active power vector belonging to the cluster of
A feature vector for generating a feature vector including, as elements, an element of an active power vector in one unit period of the user, an element of a feature time representative vector generated for the user, and a weather-related value in the one unit period Generating means;
An estimation model construction means for constructing an estimation model by performing machine learning using a teacher data set including, as teacher data, a set of the feature vector and a user attribute value indicating a user attribute of the user related to the feature vector;
A feature vector including, as elements, an element of an active power vector in one unit period of the estimation target user, an element of a feature time representative vector generated for the estimation target user, and a weather-related value in the one unit period A user attribute estimation device comprising: a user attribute estimation unit configured to input to the constructed estimation model and obtain an estimated value of a user attribute of the estimation target user. A user attribute estimation method using a computer mounted on a device that estimates the user attribute of the estimation target user based on the power consumption of the estimation target user,
A first step of acquiring a weather-related value indicating information relating to weather in a region related to the user in a predetermined period;
Based on the power consumption data of the user measured during the predetermined period, the device-derived power amount is calculated based on the minimum value of the power consumption amount in each time zone, and the power consumption amount in the time zone and the A second step of calculating an amount of active power corresponding to the amount of power related to the activity of the user based on the amount of power derived from the device;
For each unit period constituting the predetermined period, an active power vector having the active power amount in each time zone as an element is generated from the active power amount of the user for the predetermined period, and the active power amount is determined according to the weather. A clustering process is performed to select an active power vector related to a unit period in which a weather-related value has a predetermined value in a characteristic time zone in which a feature can occur, and to classify the selected active power vector into a predetermined number of clusters. A third step of generating a feature time representative vector representative of the amount of active power in the feature time zone of the user based on the active power vector belonging to the cluster of
A fourth feature vector is generated that includes, as elements, an element of an active power vector in one unit period of the user, an element of a feature time representative vector generated for the user, and a weather-related value in the one unit period. And the steps
A fifth step of constructing an estimation model by performing machine learning using a teacher data set including, as teacher data, a set of the feature vector and a user attribute value indicating a user attribute of the user related to the feature vector;
A feature vector including, as elements, an element of an active power vector in one unit period of the estimation target user, an element of a feature time representative vector generated for the estimation target user, and a weather-related value in the one unit period And a sixth step of obtaining an estimated value of the user attribute of the estimation target user by inputting to the constructed estimation model.

Images