ON-DEMAND POWER CONTROL SYSTEM, ON-DEMAND POWER CONTROL SYSTEM PROGRAM, AND COMPUTER-READABLE RECORDING MEDIUM RECORDING THE SAME PROGRAM

Abstract

In order to estimate behavior of a person at home from a feature and a power consumption pattern of an electrical device, an on-demand power control system of the present invention includes initial human-induced probability value estimation means that estimates a state of the electrical device, and estimates an initial value of a human-induced probability of the electrical device based on the estimated state of the electrical device, human position estimation means for calling up the initial value of the human-induced probability of the electrical device and a likelihood map of this device from a memory, performing, for all samples, a process of referring to a sample human position selected from the likelihood map and calculating a weight of the device by multiplying a human position and the human-induced probability of the device, and estimating a probability of a human position at each time point until a final time; and human-induced probability re-estimation means for performing recalculation of the human-induced probability based on the human-induced probability and a human position probability, performing the recalculation of the human-induced probability until a value of the recalculation converges, and outputting the human-induced probability and the human position probability when the value converges.

Description

TECHNICAL FIELD

The present invention relates to an on-demand power control system, an on-demand power control system program, and a computer-readable recording medium recording the same program and, more particularly, to an on-demand power control system, an on-demand power control system program, and a computer-readable recording medium recording the same program which are capable of analyzing a unique feature of each electrical device collected by a smart tap and a power consumption pattern, and estimating behavior of a person at home.

BACKGROUND ART

An on-demand power control system is intended to implement energy management in households and offices. The system aims to make a 180-degree shift from a supplier-centric “push” power network to a user- or consumer-driven “pull” power network. The system is a system in which a home server infers “which one of the device requests is most important” from a user's usage pattern in response to requests for power from various devices at home (e.g., requests from an air conditioner and a light) and performs control so as to supply power to electrical devices beginning with an important one with high priority, i.e., performs on-demand power control. In the following, Energy on Demand control which is the on-demand power control will be referred to as “EoD control”, and a system thereof will be referred to as an “EoD control system”. The EoD control system is proposed by Professor Takashi Matsuyama, Kyoto University.

The greatest benefit of use of the system is that energy saving and CO₂ emissions reduction can be implemented from the demand side. For example, if a user sets instructions to make a 20% electric rate cut in the home server in advance, a user-centric effort to feed only power cut by 20% can be made by EoD control, and the system can implement energy saving and CO₂ emissions reduction.

As the EoD control system described above, there is known a home network including estimation means for estimating the function of an electrical device that is in operation and a positional relationship between a resident and the electrical device, and estimating the number of residents at home and the behavior of the resident(s) based on the operation status(es) of electrical device(s) (see Patent Literature 1). More specifically, the home network described above includes n electrical devices arranged at home, n modules for supplying power to the n electrical devices from outlets, and for detecting power use statuses of the electrical devices and the arranged positions thereof, detection means for detecting operation statuses of m electrical devices that are actually in operation, based on the power use statuses of the n electrical devices transmitted from the modules, and detecting mutual positional relationships of the m electrical devices that are in operation based on the n arranged positions transmitted from the n modules, and estimation means for determining feasibility of the operation statuses of the m electrical devices that are in operation and estimating the number of residents at home, based on the mutual positional relationships which have been detected and the operation statuses of the m electrical devices that are in operation.

Additionally, the module described above is called a “smart tap” (hereinafter, referred to as a “ST”) these days, and this ST is composed of voltage and current sensors which measure power, a semiconductor relay for power control, a ZigBee module for communication, and a microcomputer with a built-in DSP which performs overall control of the components and internal processing, and may calculate in detail current and voltage waveforms by high speed data sampling at 20 kHz (see Non Patent Literature 1).

The estimation means of Patent Literature 1 requires burdensome input operations to input to the modules in advance what each electrical device is, and measures the mutual positional relationships of the electrical devices and estimates the number of residents at home and the behavior of the resident(s) by attaching IC tags to outlets provided in the wall and providing IC tag readers to the modules. Thus, since the estimation means corresponds to human behavior estimation means, it will be referred to as the human behavior estimation means in the following. Regarding this human behavior estimation means, a problem is pointed out that, in spite of requiring burdensome labor such as an input operation of electrical devices, attachment of IC tags, installation of IC tag readers, and the like, only limited information regarding the number of residents at home and the behavior of the resident(s) based on the mutual positional relationships of the electrical devices may be obtained.

Accordingly, to solve the problem described above, a method of identifying what the electrical device being measured by the ST is being proposed (see Non Patent Literature 2). According to this method, a small number of features representing characteristics of the current waveform is extracted and transmitted to a server by the ST, and an electrical device is identified by using the obtained features by the server. The household commercial power is AC, and as can be seem from the voltage/current waveforms of a hair dryer and a vacuum cleaner in FIG. 7, the values and polarities of the current/voltage waveforms are constantly changing during use of the electrical devices, and thus, the method mentioned above allows identification of the electrical devices by comparing the phase differences of the current and voltage and the shapes of the waveforms according to the types of the electrical devices.

CITATION LIST

Patent Literature

Patent Literature 1

• International Publication No. WO 2008/152798

Non Patent Literature

Non Patent Literature 1

• “i-Energy and Smart Grid,” Professor Takashi Matsuyama, Graduate School, Kyoto University, p. 21, Jul. 29, 2009

Non Patent Literature 2

• “Electric Appliance Recognition from Power Sensing Data for Information-Power Integrated Network System” (Mobile P2P, ubiquitous networks, Ad hoc networks, Sensor network, General) by Takekazu Kato, Cho Hyun-Sang, Lee Dongwook, Tetsuo Toyomura, Tatsuya Yamazaki, The Institute of Electronics, Information and Communication Engineers, Technical report, USN, Ubiquitous Sensor Network, Vol. 108, NO. 399, (2009), pp. 133-138

SUMMARY OF INVENTION

Technical Problem

Regarding the human behavior estimation means of Patent Literature 1, a problem is pointed out that the information obtained is only limited information for estimating the number of residents at home and the behavior of the resident(s) based on the mutual positional relationships of the electrical devices. Now, due to the damage to the Fukushima No. 1 nuclear power plant caused by the Great East Japan Earthquake of March 2011, the Japanese government announced a policy to reduce electricity consumption by about 15% compared to the previous year, regarding a power-saving target within the jurisdictions of Tokyo Electric Power Company and Tohoku Electric Power Company during an on-peak period in summer which is a pillar of countermeasures against power shortages. Combined with this, users' desire to reduce the power consumption of electrical devices, even if only by a small amount, without impairing the Quality of Life (hereinafter, referred to as “QoL”) of the users is becoming strong. To reduce power of the electrical devices without impairing the QoL, to what extent the power consumption can be reduced while maintaining the QoL becomes important, and if the behavior of a person using an electrical device at home may be estimated from the unique feature of each electrical device and the power consumption pattern, a light which one forgot to turn off may be turned off, and an electrical device which one forgot to stop may be stopped, for example. Also, if an electrical device put to use may be identified by using the feature and the power consumption pattern, and the behavior of a person at home may be estimated based on the operation state of the electrical device which has been identified (on/off or high/medium/low of a switch, or the like), power may be saved while guaranteeing the QoL.

Accordingly, in view of the conventional problem as described above, the present invention aims to provide an EoD control system, an EoD control system program, and a computer-readable recording medium recording the same program which estimate the behavior of a person at home based on a feature of an electrical device and a power consumption pattern without estimating the number of persons based on the mutual positional relationships of electrical devices.

Solution to Problem

As a result of keen examination to achieve the above-described object, the present inventors have reached the present invention.

A human behavior estimation apparatus of the invention according to claim 1 of the present invention is an on-demand power control system including a power source including at least a commercial power source, a plurality of electrical devices, a smart tap connected to the electrical devices, a human behavior estimation apparatus, including a memory, for estimating behavior of a person in a living space, and a network to which the human behavior estimation apparatus is connected via the smart tap, wherein the human behavior estimation apparatus includes initial human-induced probability value estimation means for comparing a feature obtained from a power consumption pattern of an electrical device with a feature of learning data obtained in advance for each electrical device to estimate a state of the electrical device, and estimating an initial value of a human-induced probability of the electrical device based on the estimated state of the electrical device, human position estimation means for calling up the initial value of the human-induced probability of the electrical device and a likelihood map of the device from the memory, performing, for all samples, a process of referring to a sample human position selected from the likelihood map and calculating a weight of the device by multiplying a human position and the human-induced probability of the device, and estimating a probability of a human position at each time point until a final time, and human-induced probability re-estimation means for performing recalculation of the human-induced probability based on the human-induced probability and a human position probability, performing the recalculation of the human-induced probability until a value of the recalculation converges, and outputting the human-induced probability and the human position probability when the value is converged.

The human behavior estimation apparatus of the invention according to claim 2 of the present invention estimates the behavior of a person based on the human-induced probability that is a probability of a person performing a human-induced operation on a device and the human position probability that is a probability of a person existing at a position.

The human behavior estimation apparatus of the invention according to claim 3 of the present invention determines feature data of an operation state based solely on the power consumption pattern of each electrical device, and determines the human position and the human-induced probability by repeated calculation while taking a human-induced probability obtained from the data as an initial value, to estimate both the human position and the human-induced probability.

According to the human behavior estimation apparatus of the invention according to claim 4 of the present invention, the feature data of an operation state is composed of an id of an electrical device, an operation state of the device, and a feature of the device.

According to the human behavior estimation apparatus of the invention according to claim 5 of the present invention, the power consumption pattern of an electrical device is generated based on a change in an operation mode or a state of an electrical device due to operation of the device by a sequence of human states including movement and stopping.

According to the human behavior estimation apparatus of the invention according to claim 6 of the present invention, the feature of an electrical device is an average and dispersion of power consumption.

According to the human behavior estimation apparatus of the invention according to claim 7 of the present invention, the initial human-induced probability value estimation means is capable of estimating the initial value of the human-induced probability based on a probability of a state of an electrical device, by having a probability of a state shift of the electrical device being caused by a human-induced operation given in advance.

According to the human behavior estimation apparatus of the invention according to claim 8 of the present invention, the human position estimation means estimates a movement position of a person based on a human-induced probability according to which a two-dimensional position of the person at a time t may be determined by repeated calculation for a preceding time t−1.

According to the human behavior estimation apparatus of the invention according to claim 9 of the present invention, the human position estimation means uses a time-series filter to estimate the human position probability.

According to the human behavior estimation apparatus of the invention according to claim 10 of the present invention, the time-series filter is a particle filter, a Kalman filter or a moving average filter.

According to the human behavior estimation apparatus of the invention according to claim 11 of the present invention, the human position estimation means calculates the human-induced probability by using a likelihood map that is based on a relationship between a position of an electrical device and a human position.

According to the human behavior estimation apparatus of the invention according to claim 12 of the present invention, the likelihood map shows, using pixel values, an existence probability of a person who can press a power switch of an electrical device, with respect to an actual living space.

According to the human behavior estimation apparatus of the invention according to claim 13 of the present invention, in the likelihood map, calculation is performed using a pixel value of zero for a range that cannot be reached by a human hand if there is an obstacle.

According to the human behavior estimation apparatus of the invention according to claim 14 of the present invention, in the likelihood map, small pixel values are assigned to a wide range in a case where the electrical device may be operated by a remote control.

According to the human behavior estimation apparatus of the invention according to claim 15 of the present invention, in the likelihood map, in a case where the electrical device may be operated from a plurality of positions, large pixel values are assigned to a plurality of narrow ranges allowing operation.

According to the human behavior estimation apparatus of the invention according to claim 16 of the present invention, in the likelihood map, the pixel values are expressed based on a floor plan, the position/distance of an obstacle and a power switch, and a position of the electrical device.

According to the human behavior estimation apparatus of the invention according to claim 17 of the present invention, in a case where there is a plurality of persons, the human estimation apparatus performs calculation using a mixed normal distribution for the number of persons.

A program of the invention according to claim 18 of the present invention is a program for causing a computer to operate as a human behavior estimation apparatus on an on-demand power control system including a power source including at least a commercial power source, a plurality of electrical devices, a smart tap connected to the electrical devices, the human behavior estimation apparatus, including a memory, for estimating behavior of a person in a living space, and a network to which the human behavior estimation apparatus is connected via the smart tap, wherein the human behavior estimation apparatus includes initial estimated human-induced probability value setting means, human position estimation means, and human-induced probability re-estimation means, and wherein the program causes the computer to perform processes of: by the initial estimated human-induced probability value setting means, comparing a feature obtained from a power consumption pattern of an electrical device with a feature of learning data obtained in advance for each electrical device to estimate a state of the electrical device, and calculating an estimation of an initial value of a human-induced probability of the electrical device based on the estimated state of the electrical device, by the human position estimation means, calling up the initial value of the human-induced probability of the electrical device and a likelihood map of the device from the memory, performing, for all samples, a process of referring to a sample human position selected from the likelihood map and calculating a weight of the device by multiplying a human position and the human-induced probability of the device, and calculating an estimation of a probability of a human position at each time point until a final time, and by the human probability re-estimation means, performing recalculation of the human-induced probability based on the human-induced probability and a human position probability, performing the recalculation of the human-induced probability until a value of the recalculation converges, and outputting the human-induced probability and the human position probability when the value is converged.

A program of the invention according to claim 19 of the present invention is a program for causing a computer to operate as a human behavior estimation apparatus according to claim 1, the program causing the computer to perform a process of estimating, by the human behavior estimation apparatus, behavior of a person based on a human-induced probability that is a probability of a person performing a human-induced operation on a device and a human position probability that is a probability of a person existing at a position.

A program of the invention according to claim 20 of the present invention is a program for causing a computer to operate as a human behavior estimation apparatus according to claim 1, the program causing the computer to perform a process of, by the human behavior estimation apparatus, determining feature data of an operation state based solely on a power consumption pattern of each electrical device, and determining a human position and a human-induced operation by repeated calculation while taking a human-induced probability obtained from the data as an initial value, to estimate both the human position and a human-induced operation.

A recording medium of the invention according to claim 21 of the present invention is a computer-readable medium recording a program according to claim 18.

A recording medium of the invention according to claim 22 of the present invention is a computer-readable medium recording a program according to claim 19.

A recording medium of the invention according to claim 23 of the present invention is a computer-readable medium recording a program according to claim 20.

Advantageous Effects of Invention

A human behavior estimation apparatus of an on-demand power control system of the present invention is capable of estimating the behavior of a person at home through the daily life of a user in which the user uses electrical devices, and thus, power may be preferentially supplied to a necessary electrical device or unnecessary light may be turned off by laying out a plan regarding use of power or predicting an electrical device which is likely to be used next. This is an apparatus that is highly useful in saving power because power can be preferentially supplied to an electrical device needed by a user, or a light which the user forgot to turn off can be turned off and an electrical device which the user forgot to stop can be stopped, for example.

That the human behavior estimation apparatus is capable of estimating the behavior of a person at home was proven by the result of a real-life experiment conducted in a smart apartment room where subjects actually lived.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a schematic diagram showing the configuration of a communication network of an EoD control system.

FIG. 2 is a schematic diagram showing the configuration of a power network of the EoD control system according to the present invention.

FIG. 3 is an explanatory diagram showing a relation of connection among an outlet, a ST, and an electrical device.

FIG. 4 is a floor plan showing the arrangement of the ST for installing each electrical device in a model house.

FIG. 5 is a relational view showing the state of a person and operation of a device.

FIG. 6 is a relational view showing a relationship between the state of a device and a power consumption pattern.

FIG. 7 is a waveform diagram showing voltage/current waveforms of a hair dryer and a vacuum cleaner.

FIG. 8 is a functional block diagram showing functions of a human behavior estimation apparatus.

FIG. 9 is a diagram showing a likelihood map of an IH stove.

FIG. 10 is a diagram showing a likelihood map of a TV.

FIG. 11 is a diagram showing a likelihood map of a living room light.

FIG. 12 is a diagram showing a likelihood map of a washroom light.

FIG. 13 is a diagram showing a likelihood map of a corridor light.

FIG. 14 is a flow chart showing an overall process of the human behavior estimation apparatus.

FIG. 15 is a flow chart showing a process of initial value estimation of a human-induced probability.

FIG. 16 is a flow chart showing a process of human position estimation (one person).

FIG. 17 is a flow chart showing a succeeding process of the process of human position estimation (one person).

FIG. 18 is a flow chart showing a re-estimation process of the human-induced probability.

FIG. 19 is a flow chart showing a process of human position estimation (a plurality of persons).

FIG. 20 is a flow chart showing a succeeding process of the process of human position estimation (a plurality of persons).

FIG. 21 is a flow chart showing a succeeding process of the process of human position estimation (a plurality of persons).

FIG. 22 is a diagram showing a probability distribution of the human position obtained by human position estimation means.

FIG. 23 is a probability distribution diagram showing a probability distribution of sample points obtained by the human position estimation means.

FIG. 24 is an activity path diagram showing the path of activities of a person obtained as a result of estimation by the human position estimation means.

FIG. 25 is a dispersion diagram of positions of particles showing dispersion of a distribution and the time of occurrence of an event.

DESCRIPTION OF EMBODIMENTS

The configuration of a communication network of an EoD control system according to the present invention will be described with reference to FIG. 1.

FIG. 1 is a schematic diagram showing the configuration of a communication network of an EoD control system according to the present invention. An EoD control system 50 of the present invention is installed in an ordinary home, and is composed of a human behavior estimation apparatus, a ST 11, an electrical device (hereinafter, referred to simply as a “device”), and a power control apparatus 30. The human behavior estimation apparatus is connected to the STs 11 over a local area network (hereinafter referred to as a “LAN”) by a wired or wireless LAN. The LAN is merely an example, and the present invention is not limited to this. According to the present invention, the priority apparatus may be connected to the STs over a network such as WiFi, PLC, ZigBee, or specific low-power radio waves. The devices are connected to the STs through power cords. Accordingly, the STs can communicate with the human behavior estimation apparatus over the LAN.

Furthermore, with respect to the human behavior estimation apparatus, a knowledge database (a knowledge DB) 10 may be structured inside the human behavior estimation apparatus by using a directly connected or built-in memory. Power from a commercial power source is supplied via the power control apparatus 30 to the human behavior estimation apparatus and each device.

Additionally, a commercial power source is described as the power source of the EoD control system of the present invention, but the present invention is not limited to this, and photovoltaic power generation or a fuel cell may be used as the power source. Although an ordinary household will be described as the installation location of the EoD control system 50 according to the present invention, the present invention is not limited to this. Any location such as an office may be adopted as long as a ST can be installed. An external type ST which is connected to a power outlet will be described as a ST of the EoD control system according to the present invention. The present invention, however, is not limited to this, and an internal type one which is embedded in a power outlet may be employed.

FIG. 2 is a schematic diagram showing the configuration of a power system network of the EoD control system 50 shown in FIG. 1.

As has been described with reference to FIG. 1, the EoD control system 50 includes the power control apparatus 30. A commercial power source 32 is connected to the power control apparatus 30. The power control apparatus 30 is composed of, for example, a plurality of breakers (not shown) and includes one main breaker and a plurality of sub-breakers. Power (AC voltage) from the commercial power source 32 is given to the primary side of the main breaker, and an output from the secondary side of the main breaker is distributed among the plurality of sub-breakers. Note that the commercial power source 32 is connected to the primary side of the main breaker through a switch (not shown) for supplying/stopping supply of commercial current. The switch is turned on/off by a switching signal from the human behavior estimation apparatus.

The human behavior estimation apparatus and the plurality of devices 20 are connected to the output side of the power control apparatus 30, i.e., the secondary sides of the sub-breakers. Although not shown, the human behavior estimation apparatus is connected so as to be capable of receiving power from the power control apparatus 30 by inserting its attachment plug into, e.g., a wall socket. For the plurality of devices, the STs each include a plug which is an attachment plug and an outlet, and power from the commercial power source 32 is fed from the plug. The plurality of devices are connected so as to be capable of receiving power through plugs of the plurality of devices connected to the outlets.

As described above, in the EoD control system according to the present invention, not only the power network shown in FIG. 2 but also the communication network shown in FIG. 1 are constructed.

FIG. 3 is an explanatory diagram explaining a relation of connection among an outlet which is connected to a commercial power source and which is arranged in a wall, a ST, and a device. Referring to FIG. 3, a refrigerator 201 which is a device is composed of a plug unit 202 including an attachment plug and a cord 203, and the plug unit 202 of the refrigerator 201 is inserted into/removed from an outlet 114 of the ST. An outlet 41 is arranged in a wall 40, and commercial power is supplied to slots 411 in the outlet 41 through a power system at home. A plug 113 which is an attachment plug is inserted into/removed from the slots 411.

Regarding the structure of the ST, a known structure composed of voltage and current sensors which measure power, a semiconductor relay for power control, a ZigBee module (hereinafter, referred to as a “communication member”) for communication, and a microcomputer which performs overall control of the components and internal processing is used, as described above. The ST measures in real-time the consumption of power supplied to a device via the plug 113, and transmits the power consumption to the human behavior estimation apparatus via the communication member.

FIG. 4 is a floor plan showing the arrangement of the ST for installing each electrical device in a model house.

The size shown by the floor plan is 538 cm×605 cm, i.e. about 33 m². As shown in FIG. 4, devices used in an ordinary home, such as a TV, an air conditioner, an electric pot, a coffee maker, a rice cooker, a refrigerator, a microwave oven, a washing machine, a vacuum cleaner, a hair dryer and the like, are provided. The positions of the STs are shown in FIG. 4 as “ST”, and the ST is provided for each device. Among the devices, devices with a remote control are the TV, the air conditioner, a living room light, two kitchen lights, a bedroom light, and a VCR. The ST to be used at the time of using a portable device is five free outlets. Portable devices are the hair dryer, the vacuum cleaner, a mobile charger, a notebook PC, an air cleaner, an electronic toothbrush, and a night stand lamp. Numbers shown in FIG. 4 are the numbers in the id column of Table 1, and represent the names of the devices in the name column.

Table 1 is a table showing possible operation states (0 to 3) of each device.

TABLE 1
id	name	State 0	State 1	State 2	State 3
1	TV	Off	On
2	Air conditioner	Off	Ventilation	In operation
4	Pot	Off	On
5	Coffee maker	Off	On (Overheating)
6	Night stand lamp	Off	On
7	Rice cooker	Off	On
8	Refrigerator	On (Stopped)	On (Cooling)	Door is open
9	Microwave oven	Off	On
10	Washing machine	Off	On
11	Living room light and kitchen	Off	Living room	Living room +	Kitchen
	light 2			kitchen
12	Bedroom light	Off	On
13	Kitchen light 1	Off	On
15	Corridor light	Off	On
16	Washroom light	Off	On
17	Toilet light and ventilation fan	Off	Light on	Ventilation fan	Light + ventilation fan
18	Toilet seat with warm-water	Off	Warm seat	Nozzle cleaning	Cleaning
	shower feature
20	Air cleaner	Off	On
21	Vacuum cleaner	Off	High	Low
22	Hair dryer	Off	Drying mode	Set mode	Cool mode
24	Electric toothbrush	Off	On (Charging)
30	Bathroom light and ventilation	Off	Light on	Ventilation fan	Light + ventilation fan
	fan
40	Electric carpet	Off	One side	Both sides	Waiting
41	Heater	Off	Low	High
42	Router	Off	On
43	VCR	Off (Waiting)	Playing
44	IH	Off	High	Medium	Low
45	Mobile charger	Off	Charging
48	Notebook PC	Off	Charging	Charging + in use

Now, the “behavior of a person” is generally defined as a plurality of continuous actions and a sequence of human states such as stopping and movement of the person. The present invention takes notice that, in an ordinary living space, many actions involve operation of devices, and thus, the behavior is defined taking actions as a series of human-induced operations of a device and a human state as the position where a person performs an action.

That is, the following two objects are to be achieved by the present invention.

(1) Detection of a human-induced operation of a device

(2) Estimation of the position of a person

FIG. 5 is a relational view showing the state of a person and operation of a device.

FIG. 6 is a relational view showing a relationship between the state of a device and a power consumption pattern.

As shown in FIG. 5, a person performs operation of a device by a sequence of human states such as movement and stopping. When a device is operated, the state of the device and the operation mode are changed and are observed as a power consumption pattern of devices of a day indicating which device consumed how much power where and at what time as shown in FIG. 6. This power consumption pattern of a day is defined as the “power consumption pattern” of devices and is used in the following.

Additionally, the process described below to be performed by the human behavior estimation apparatus is performed off-line, and the human behavior estimation apparatus processes the power consumption pattern transmitted by the ST and performs off-line a process of indicating the activity path of a person where a probability distribution of a human position z

P_h⁽ⁱ⁾(z) and a set of probabilities of a device being operated

P_m⁽ⁱ⁾

are output as an estimation result.

FIG. 7 is a waveform diagram showing voltage/current waveforms of a hair dryer and a vacuum cleaner, and the diagram on the left is the waveforms of the hair dryer and the diagram on the right is the waveforms of the vacuum cleaner.

As shown in FIG. 7, the voltage waveform (the thick line) and the current waveform (the thin line) of each device are different depending on the device, and a device may be identified by analyzing the average and dispersion of the power based on these waveforms.

Table 2 shows each device and its operation states, 0 to 3 (on/off or high/medium/low of a switch, or the like), and the values of average and dispersion of power consumption (the average and dispersion will be hereinafter referred to as “feature(s)”), and the feature is used in the process of initial value estimation of a human-induced probability described later as learning data. State 0, state 1, state 2, and state 3 in Table 1 are the same as the operation states 0 to 3 of devices in Table 2, and indicate possible operation states of devices. As can be seen from Table 2, the feature changes according to a change in the operation state, and the value of the feature is different for each device. This Table 2 showing the operation states and features of devices (id) is defined and used as the “feature data of an operation state”.

	TABLE 2
	id	State	Average	Dispersion
	1	0	33.67584152	0.794761764
	1	1	519.9163551	191.5096297
	2	0	13.2	1.501183506
	2	1	92.51079325	1.575825427
	2	2	771.3062061	9834.533893
	4	0	0.037665621	0.00393113
	4	1	1178.754094	115228.527
	5	0	0	0
	5	1	883.2017	7.680225
	6	0	0	0
	6	1	40	0.124923425
	7	0	3.008685474	1.257193155
	7	1	511.1991327	16.39765424
	8	0	79.61447	0.510276
	8	1	0.689269	0.676579
	8	2	92.71951	4.610942
	9	0	2.871526283	0.036332956
	9	1	1342.814552	144.2288547
	10	0	1	0
	10	2	66.1586741	21794.22809
	11	0	0.313957949	0.000051
	11	1	73.88338248	0.662251365
	11	2	96.15896309	0.377936437
	11	3	24.54132468	0.202942811
	12	0	0.435026169	1.135486605
	12	1	83.54786	0.917483
	13	0	0	0
	13	1	24.12100601	6.205286
	15	0	0.339959	0.000303
	15	1	145.0075	0.094239
	16	0	0	0
	16	1	13.07276	0.038932
	17	0	0	0
	17	1	55.1002	0.01782
	17	2	12.08802	0.002011
	17	3	68.01902	0.002623
	18	0	0.455819	0.468028
	18	1	57.75753	0.008368
	18	2	149.9344	12142.89921
	18	3	122.4006	4.101377799
	20	0	0	0
	20	1	28.72909841	0.040321459
	21	0	0.828771283	0.013420972
	21	1	923.7198771	2500.292504
	21	2	270.2669604	102.1732589
	22	0	0	0
	22	1	1248.322651	144.3794678
	22	2	70.22042084	0.097136284
	22	3	59.67681294	0.09438221
	24	0	0	0
	24	1	0.8	0.19345
	30	0	0	0
	30	1	9.810464	0.035764
	30	2	15.93536	0.259133
	30	3	27.18636	0.443368
	40	0	0	0
	40	1	214.4565	15.95754
	40	2	0.366176	0.002291
	40	3	428.2257	122.0433
	41	0	0.412173	0.373117
	41	1	656.1504	152.1595
	41	2	1218.378	1186.601
	42	0	0	0
	42	1	30.245633	1.345523
	43	0	30.73303	0.376635
	43	1	28.54118	0.413536
	44	0	3.554298	0.487601
	44	1	1206.067	883.7798
	44	2	471.9262	7.701671
	44	3	270.34565	5.603545
	45	0	0	0
	45	1	2.854644	0.504454
	48	0	0	0
	48	1	15.13420972	0.871526283
	48	2	30.093702	5.339071

When position information of devices is given, if the timing of a human-induced operation performed by a person with respect to a device is known, it is possible to know that the person was near the device at that time, and by connecting these facts in a chronological order, the movement track of the person may be estimated. However, a device includes not only a state change due to a human-induced operation, but also an automatic state change, continuous load variations and the like, and it is difficult to identify a human-induced operation based solely on the power consumption pattern. If the position information of a person may be estimated at this time, since a human-induced operation is possible if the person is near the device and the operation is not possible if the person is far away from the device, it is possible to distinguish between a human-induced operation and a state change induced by other factors. To distinguish between the two, it is necessary to understand a “human-induced probability

p_m(s_a,t=1)”, which is the probability of a person performing a human-induced operation with respect to a device at a time t, and a “human position probabilityp_h(Z_t)”, which is the probability of a person being able to operate a device from a position Z_t. The “human-induced probabilityp_m(s_a,t=1)” and the “human position probabilityp_h(Z_t)” will be described below.

Whether a human-induced operation is performed on a certain device

aat a timetwill be expressed bys_a,t.

$\begin{array}{cc}{s}_{a,t}=\{\begin{array}{cc}1& \mathrm{Human}\text{-}\mathrm{induced}\mathrm{operation}\mathrm{is}\mathrm{performed}\\ 0& \mathrm{Human}\text{-}\mathrm{induced}\mathrm{operation}\mathrm{is}\mathrm{not}\mathrm{performed}\end{array}& \left(1\right)\end{array}$

Also, when the human position at a time

tis given as

Z_t,

and the power consumption pattern ofaup to the timetis given as

W_a,t,

the probabilityp_m(s_a,t=1)of a person performing a human-induced operation (hereinafter, referred to as a “human-induced probability”) onais expressed by the following equation.

p _m(s_a,t=1)=∫p_m(s_a,t=1|W_a,t,Z_t)p_h(Z_t)dZ_t=∫p_m(W_a,t|s_a,t=1)P_m(s_a,t=1|Z_t)p_h(Z_t)dZ_t  (2)

p_m(W_a,t|s_a,t=1)is the probability of a power consumption pattern

W_a,t

being obtained when a person operates the devicea,

P_h(Z_t)

is the probability distribution of the human position (hereinafter, referred to as a “human position probability”), andP_m(s_a,t=1|Z_t)is the probability of a person being able to operateafrom a certain position

Z_t.

According to this equation, it can be seen that, if the human position probability

P_h(Z_t)

is obtained, the human-induced probabilityP_m(s_a,t=1) may be estimated.

On the other hand, when the number of devices is given as

N,

whether a human-induced operation is performed on all the devices at a timet is expressed by a sets_t{s_a1,t, s_a2,t, . . . , s_aN,t},and a sequence of human-induced operations up to a timetis given asS_t={s₀, s₁, . . . , s_t},the human position may be determined by the following equation by Bayes' theorem.

$\begin{array}{cc}{p}_h\left(z_t\right)=p_h\left(z_t|S_t\right)=\frac{p_h\left(s_t|z_t\right)p_h\left(z_t|S_{t-1}\right)}{p\left(s_t|S_{t-1}\right)}& \left(3\right)\end{array}$

Moreover, sinceP_h(s_t|S_t-1)is not related to a probability variable

Z_t,

$\frac{1}{p\left(s_t|S_{t-1}\right)}$

is given as a normalization constantk_t,andk_tp_h(S_t|Z_t)p_h(Z_t|S_t-1)is given. Here,p_h(S_t|Z_t)is the human-induced probability for each device based on the human position

Z_t,

and may be calculated by the following equation.

$\begin{array}{cc}{p}_h\left(s_t|z_t\right)=\prod _a\left\{p_m\left(s_{a,t}=1|z_t\right)p_m\left(s_{a,t}=1\right)+\alpha p_m\left(s_{a,t}=0\right)\right\}& \left(4\right)\end{array}$

Here, the influence on a device on which the human-induced operation is not performed is irrelevant to the human position, and thus, a specific probability

αis given as a uniform distribution. Also,p_h(Z_t|S_t-1)in Equation (3) is a human position estimated based on the observation until a timet−1,and when assuming that the movement track of a person has a Markov property, modification to the following equation is possible.

p _h(Z_t|S_t-1)=∫P_h(Z_t|Z_t-1)p_h(Z_t-1|S_t-1)dZ_t-1  (5)

Here,

P_h(Z_t|Z_t-1)

is a movement model of a person, and

P_h(Z_t-1|S_t-1)

is a human position probability at a preceding time. Equations (3) and (5) are time series filter equations, and may be efficiently solved by a particle filter, a Kalman filter or a moving average filter, each being a time series filter, if the probability of a human-induced operation expressed by Equation (4) and the movement model are given. It can be said that, if a human-induced operationp_m(S_t)is obtained, for example, the human positionp_m(Z_t)may be determined by the particle filter.

Accordingly, it can be seen that, according to this problem, there is interdependence that, if the human-induced operation

p_m(S_t)is known, the human positionp_h(Z_t)may be determined, and if the human positionp_h(Z_t)is known, the human-induced operationp_m(S_t)may be determined.

The human behavior estimation apparatus of the present invention may estimate both of the human position and the human-induced operation by determining the feature data of the operation state based solely on the power consumption pattern of each device, and determining the human position and the human-induced operation by repeated calculation while taking the feature of the human-induced operation determined from the data as the initial value.

FIG. 8 is a functional block diagram showing functions of the human behavior estimation apparatus shown in FIG. 1.

The human behavior estimation apparatus is composed of initial human-induced probability value estimation means 120, human position estimation means 122, and human-induced probability re-estimation means 124.

This human behavior estimation apparatus 1 has (1) a function of setting, by the initial human-induced probability value estimation means 120, an initial value of a human-induced probability of a device by comparing the power consumption pattern of the device received from the ST with the feature and the operation probability of the device stored in a memory, (2) a function of estimating a human position by the human position estimation means 122 by using a time series filter and based on the initial value, and a floor plan, the arrangement of devices, the device operation probability at each position and a human movement model stored in the memory, and (3) a function of estimating, by the human-induced probability re-estimation means 124, a human-induced probability of the device being operated, based on the estimated human position, inputting the estimated human-induced probability to the human position estimation means 122, inputting the estimated human position to the human-induced probability re-estimation means 124, performing processes until the human position estimation process and the human-induced probability re-estimation process converge, and outputting the human position and the human-induced probability.

Probability calculation processes performed by (1) the initial human-induced probability value estimation means 120, (2) the human position estimation means 122, and (3) the human-induced probability re-estimation means 124 in FIG. 8 will be described below.

(1) Initial Human-Induced Probability Value Estimation Means

The probability calculation process of the initial human-induced probability value estimation means 120 will be described.

In the initial estimation, the probability of a human-induced operation (the human-induced probability) is estimated based on Equation (6) by using only the power consumption pattern

W_a,t

of each of the devicesa₁, a₂, . . . , a_N.

P _m ⁽⁰⁾(s_a,t=1)=p_m(s_a,t=1|W_a,t)  (6)

Each device has an operation mode unique to the device, and the operation mode is shifted by a human-induced operation or automatic control. That is, the human-induced operation and the automatic operation may be defined for each state transition like this. When a state

δis shifted to a stateδ′from a timet−1to a timet,the probability of this shift being caused by a human-induced operation is given asp_o(s_a,t=1|d_t-1=δ, d_t=δ′).Also, when the probability of the state of a household appliance beingδat the timetis given asp_o(d_t=δ|W_a,t)the human-induced probability for this household appliance is given as the following equation.

$\begin{array}{cc}{p}_m\left(s_{a,t}=1|W_{a,t}\right)=\sum _{d_{t-1},d_t}p_o\left(s_{a,t}=1|d_{t-1},d_t\right)p_o\left(d_{t-1}|W_{a,t-1}\right)p_o\left(d_t|W_{a,t}\right)& \left(7\right)\end{array}$

If the probabilityp_o(s_a,t=1|d_t-1, d_t).of a certain state shift being caused by a human-induced operation is given in advance, the initial value of the human-induced probability may be estimated from the probabilityp_o(d_t|W_a,t)regarding the state of the household appliance.

Which power consumption pattern observed by a smart tap corresponds to which household appliance is recognized in advance by the technique described in Non Patent Literature 2. At this time, estimation of the state of a device is performed by using the average and dispersion of the power consumption in a predetermined time section. A large number of average values and dispersion values of power consumption in each state are learned in advance as sample values for each state, with respect to all the devices. The average value of power consumption at the time when a household appliance

αis in a stateδis given asμ_α,δ,and the dispersion value is given asσ_α,δ,and these are normalized as below such that the average will be zero and dispersion will be one:{circumflex over (μ)}_α,δ, σ_α,δ.mpieces of data are selected and given as learning data for each state, andm′pieces of data which are different from the above and which are randomly extracted are given as unknown data, and the state of the unknown data is estimated by the nearest neighbor algorithm, and accuracy evaluation is performed. Accordingly, when the state of the deviceais assumed to beδ_i(i=1, 2, . . . , D),the probabilityp_o(δ_i|δ_k)(i, k=1, . . . , D)of being a true stateδ_i with respect to an estimated stateδ_k may be obtained. When the power consumption pattern

W_a,t

of the deviceais measured, the normalized average value{circumflex over (μ)}_a,t and dispersion value{circumflex over (σ)}_a,t of power consumption are compared with the learning data, and a stateδ_a,t is estimated. At this time, if the estimation result by the nearest neighbor algorithm is given asδ_a,t, the probability of being in the stateδ_a,t at the timetisp_o(d_t=δ_a,t|W_a,t)p_o(δ_a,t|δ_a,t).

To estimate the human-induced probability by Equation (6),

p_o(d_t=δ_a,t|W_a,t)has to be estimated for all the combinations of the states, but in reality, calculation is performed only for the state with the highest probability, for the sake of simplicity.

$\begin{array}{cc}{p}_m\left(s_{a,t}=1|W_{a,t}\right)=p_o\left(s_{a,t}=1|{\delta }_{a,t-1}^{*},{\delta }_{a,t}^{*}\right)p_o\left({\delta }_{a,t-1}^{*}|W_{a,t-1}\right)p_o\left({\delta }_{a,t}^{*}|W_{a,t}\right)\mathrm{Where},{\delta }_{a,t}^{*}=\underset{{\delta }_i}{\mathrm{argmin}}\left\{{\left({\hat{\mu }}_{a,{\delta }_i}-{\hat{\mu }}_{a,t}\right)}^2+{\left({\hat{\sigma }}_{a,{\delta }_i}-{\hat{\sigma }}_{a,t}\right)}^2\right\}& \left(8\right)\end{array}$

Also, when the human-induced probability is below a threshold

T_o,

filtering is performed such that the human-induced probability is assumed to be zero, to simplify the human position estimation.

$\begin{array}{cc}{p}_m^{\left(0\right)}\left(s_{a,t}=1\right)=\{\begin{array}{c}0\mathrm{if}p_m\left(s_{a,t}=1|W_{a,t}\right)<T_o\\ p_m\left(s_{a,t}=1|W_{a,t}\right)\mathrm{otherwize}\end{array}& \left(9\right)\end{array}$

Here, the combination of a household appliance and a time

(a,t)by whichp_m⁽⁰⁾(s_a,t=1)≧T_o is true is made an event sequenceE⁽⁰⁾={(a,t)|p_m⁽⁰⁾(s_a,t=1)≧T_o},and a combination of a time and a device with respect to which a human-induced operation has possibly been performed is indicated.

(2) Human Position Estimation Means

The probability calculation process of the human position estimation means 122 will be described.

A method of estimating a human position

p_h⁽ⁱ⁾(Z_t)based on a human-induced probabilityp_m^(i-1)(s_a,t=1)estimated according to Equations (3), (4) and (5) will be described. The internal state to be determined is a two-dimensional positionZ_t(x_t,y_t)of a person at the timet,and a movement position of the person is estimated based on a human-induced probabilityp_m^(i-1)(s_a,t=1)determined by repeated calculation for a preceding time (t−1).

According to Equation (3), when

k_t is given as a normalization constant, the human position probabilityp_h⁽ⁱ⁾(Z_t)at the timetmay be estimated by the following equation.

P _h ⁽ⁱ⁾(Z_t)=P_h⁽ⁱ⁾(Z_t|S_t)=k_tP_h⁽ⁱ⁾(S_t|Z_t)p_h⁽ⁱ⁾(Z_t|S_t-1)  (10)

Also, according to Equations (4) and (5), the following is true.

p _h ⁽ⁱ⁾(S_t|Z_t)=Π_a{p_m(s_a,t=1|Z_t)p_m^(i-1)(s_a,t=1)+αp_m^(i-1)(s_a,t=0)}  (11)

p _h ⁽ⁱ⁾(Z_t|S_t-1)=∫p_h(Z_t|Z_t-1)p_h⁽ⁱ⁾(Z_t-1|S_t-1)dZ_t-1  (12)

Here, the system model

P_h(Z_t|Z_t-1)

is a movement model of a person, and is a two-dimensional normal distribution

N(0, Σ₂),

according to the human behavior estimation apparatus of the present invention. Also, an observation model is estimated based on the human-induced probabilityp_m^(i-1)(s_a,t=1)estimated as expressed by Equation (11) and the probabilityp_m(s_a,t=1|Z)of the household applianceabeing able to be operated from a position

Z.

With respect to

$p_m\left(s_a=1|z\right),p_m\left(s_a=1|z\right)=\frac{1}{\sqrt{2\pi \sigma }}\exp \left(-\frac{{\left(z-x_a\right)}^2}{2{\sigma }^2}\right)$

is given by a normal distribution regarding distance of the household applianceafrom a position

X_a.

However, a likelihood map is given in advance such that zero is true for a position blocked by a wall or the like.

A particle filter is used to estimate the human position probability

p_h⁽ⁱ⁾(Z_t)according to Equations (10), (11) and (12). A particle filter is a method of performing estimation by approximating a probability distribution by using a large number of samples (particles) generated according to the probability distribution. A set of samples according to a prior distributionp_h⁽ⁱ⁾(Z_t|S_t-1) is given asQ_t|t={q_t|t-1^[t], . . . , q_t|t-1^[M]},and a set of samples according to a posterior distributionp_h⁽ⁱ⁾(Z_t|S_t)is given asQ_t|t={q_t|t^[t], . . . , q_t|t^[M]},[Initialization] A set of random samples

Q_0|0

is generated as an initial value.t:==1is given.[Prediction] A prediction sample q_t|t-1^[j], . . . , q_t-1|t-1^[j]+R(N(0, Σ₂))at a timetis generated according to the system model. Additionally,

R(N(0, Σ₂))

is a random number according to the two-dimensional normal distribution

N(0, Σ₂).

[Filter] A weightπ_t^[j]is calculated for each prediction sampleq_t|t-1^[j]according to the observation model by the following equation.

$\begin{array}{cc}{\pi }_t^{\left[j\right]}=\frac{p_h^{\left(i\right)}(s_t|z_t=q_{t|t-1}^{\left[j\right]}}{\sum _{j=1}^Mp_h^{\left(i\right)}\left(s_t|z_t=q_{t|t-1}^{\left[j\right]}\right)}& \left(13\right)\end{array}$

Here,

p_h^[i](s_t|Z_t=q_t|t-1^[j])is calculated according to Equation (11).

M

pieces ofq_t|t-1^[j]are sampled with replacement from

Q_t|t-1

at a rate proportional to respective weightsπ_t^[j]to obtainQ_t|t.Additionally, in the case wheres_t={0, 0, . . . , 0}is true, the prior distributionp_h^[i](Z_t|S_t-1)predicted from the distribution at a preceding time becomes the posterior distributionp_h^[i](Z_t|S_t)as it is, and thus,Q_t|t=Q_t|t-1 is given and this process may be omitted. This

Q_t|t

is taken as the set of samples for approximating the posterior distribution. Ift>t_end the process is repeated from the prediction based ont:=t+1.

Now, specific examples of likelihood maps will be described with reference to FIGS. 9 to 13.

The likelihood map shows, using pixel values, an existence probability of a person who can press a power switch of an electrical device, with respect to an actual living space.

FIG. 9 is a diagram showing the likelihood map of an IH stove, FIG. 10 is a diagram showing the likelihood map of a TV, FIG. 11 is a diagram showing the likelihood map of a living room light, FIG. 12 is a diagram showing the likelihood map of a washroom light, and FIG. 13 is a diagram showing the likelihood map of a corridor light. The pixel values of these likelihood maps are calculated based on the position of a switch, the floor plan, the arrangement of devices, and the length of a person's arm. It is represented that the whiter the screen (the greater the pixel value) is, the higher the probability of a person being at the position is, and the darker the screen is, the lower the probability of a person being at the position is. The likelihood maps of respective devices of FIGS. 9 to 13 will be described with reference to the floor plan of FIG. 4 showing the arranged positions of the devices. According to the likelihood map of an IH stove (see “44” in FIG. 4) of FIG. 9, it can be seen that a narrow range which is the position allowing manual operation of a person is white, and that the map becomes darker as the distance from the position increases, and that the map is black at a position of the wall. According to the likelihood map of a TV (see “1” in FIG. 4) of FIG. 10, since operation is performed using a remote control, it can be seen that a grey circle over a wide range becomes darker as the distance from the TV increases, unlike the case of a narrow white range for the IH stove. According to the likelihood map of a living room light (see “11” in FIG. 4) of FIG. 11, since operation using a remote control is possible, it can be seen that the range allowing a person to perform operation is grey, but positions of the walls are black. The likelihood map of a washroom light (see “16” in FIG. 4) of FIG. 12 is similar to that of the IH stove of FIG. 9, and it can be seen that a narrow range which is the position allowing a manual operation of a person is white, and that the map becomes darker as the distance from the position increases, and that the map is black at the position of the wall. According to the likelihood map of a corridor light of FIG. 13, since double pole switches are installed at two positions, i.e. near the front door (see “15” in FIG. 4) and near the living room door (see “15” in FIG. 4), it can be seen that two positions near the front door and near the living room door are white, and that positions where there are obstacles such as walls and shelves are deep black. As can be seen in the example of switch at the living room door, pixel values are calculated taking a range that can be reached if the distance is 65 cm or shorter, assuming a person's arm as 65 cm for example, as a range that a person's hand can reach in spite of an obstacle.

As can be understood from the description above, the pixel values of the likelihood map are calculated based on the use statuses of devices in an actual living space in such a way that the map is black in the case there is a fixed obstacle such as a wall or a shelf, the map is grey over a wide range in the case of a remote control, and the map is white at two positions in the case of double pole switches.

(3) Human-Induced Probability Re-Estimation Means

The probability calculation process of the human-induced probability re-estimation means 124 will be described.

A method of re-estimating a human-induced probability using a human position probability at each time point according to Equation (2) will be described. According to Equation (2), estimation by integration has to be performed for all the possible human positions, but here, to simplify the process, re-estimation is performed while excluding the calculation for a position with a low probability by determining an optimal human position with the highest probability,

{circumflex over (Z)}^(t)={{circumflex over (Z)}_t^(t)}(t=0, 1, . . . t_end).As a method of determining{circumflex over (Z)}^(t),a method of determining the same as an expected value at each time point by using a weighted average by likelihood is used in many cases with respect to real-time tracking and the like, but here, what is to be determined is the movement track of a person, and thus, a path of human positions whose product of probabilities along the path is the greatest is estimated as the optimal path, taking the temporal connection into account. This may be formulated as below by giving the number of samples in the particle filter as

M.

$\begin{array}{cc}\underset{j0,j1,\dots ,jt_{\mathrm{end}}}{\max }p\left(q_{1|0}^{\left[j0\right]},q_{2|1}^{\left[j1\right]},\dots ,q_{t_{\mathrm{end}}|t_{\mathrm{end}}-1}^{\left[jt_{\mathrm{end}}\right]}\right)& \left(14\right)\end{array}$

That is,

j₀, j₁, . . . , j_tend satisfying the above is to be determined. An estimated value{circumflex over (Z)}^(t) of a state to be determined by a sample sequence of these numbers is determined. Specifically, each sample keeps a history regarding from which of the samples at a preceding time point the sample is derived, and{circumflex over (Z)}^(t) is determined by following the history of the sample with the greatest weight att_end.The human-induced probability is updated by the following equation using this optimal path.

p _m ^(t)(s_a,t=1)=p_m(W_a,t|s_a,t=1)p_m(s_a,t=1|Z_t={circumflex over (Z)}^(t))  (15)

Update is performed by performing filtering in the same manner as Equation (9) with respect to a household appliance and a time whose human-induced probability is at or below a threshold

T_o,

and by removing a combination at or below the threshold

T_o

from the event sequence

E⁽ⁱ⁾.

E ⁽ⁱ⁾={(a,t)εE^(i-1)|p_m⁽ⁱ⁾(s_a,t=1)≧T_o}  (16)

The human-induced probability

p_m⁽ⁱ⁾(S_t)is estimated here, and human position estimation and human-induced probability re-estimation are repeated untilp_m⁽ⁱ⁾(S_t)is converged based oni:=i+1.When the number of repetitions by which convergence is achieved is given asi=i_f,the event sequence

E^(if)

which is obtained at this time may be said to be the sequence of human-induced operations obtained from

S_t.

Also, the optimal path{circumflex over (Z)}^(if) determined from the final probability distributionp_h^(if)(Z)is the movement track of a person.

FIG. 14 is a flow chart showing an overall process of the human behavior estimation apparatus of the present invention. The process of initial value estimation of a human-induced probability is performed in step S1, the process of estimation of a human position is performed in step S3 (one person) or step S7 (a plurality of persons), and the process of re-estimation of a human-induced probability is performed in step S11.

FIG. 15 is a flow chart showing a process of the initial value estimation of a human-induced probability in step S1 mentioned above.

As shown in FIG. 15, in step S11, the average value and the dispersion value of power consumption obtained by measuring the power consumption pattern of a device is compared with the learning data of the average value and the dispersion value of power consumption obtained in advance for each state of each device, and the state of the device is estimated. In step S13, the state of the device which is estimated is input to Equation (7) to estimate the initial value of the human-induced probability, and the initial value is stored in the memory as the human-induced probability.

FIG. 16 is a flow chart showing a process of estimation of a human position (one person) of step S3 mentioned above.

As shown in FIG. 16, in step S31, N samples are generated for random human positions and are made the initial value, and these are made the initial set of posterior samples (the time is an initial time (t:=1)). In step S33, each of the posterior samples at a time (t−1) is moved according to the movement model of a person, and a set of predicted samples at a time t is generated. In step S35, (1) one sample j is selected from the set of predicted samples. In step S37, (2a) one device is selected, in step S39, (2b) the likelihood map of the device and the human-induced probability of the device at a time t are called up from the memory, and in step S41, (2c) the weight of the device is calculated by referring to a sample human position selected from the likelihood map of the device and multiplying the human position and the human-induced probability of the device, and then, the process proceeds to (1).

As shown in FIG. 17, in step S43, (3) weights calculated by performing the processes of (2a), (2b) and (2c) described above on all the devices while changing the selected device are added up to thereby calculate the weight

π_t^[j]of the selected sample j. In step S45, the processes of (1) to (3) described above are performed on all the samples while changing the selected sample. In step S47, samples are duplicated by a number of pieces proportional to N×the weight of sampleπ_t^[j]for each sample in the set of predicted samples, and a set of posterior samples at the time t is generated. In step S49, whether the time t is the last time is determined, and if it is determined to be Yes in step S51, the process is ended. If it is determined to be No in step S51, t+1 is given as the time in step S53, and the process returns to step S33.

FIG. 18 is a flow chart showing a process of re-estimation of the human-induced probability of step S11 described above.

As shown in FIG. 18, in step S111, a position j on the path of a person satisfying Equation (14) is determined, and the human-induced probability is updated by Equation (15). In step S113, the event sequence is updated by removing a combination for which the human-induced probability is at or below a threshold T_o from an event sequence E⁽ⁱ⁾, based on Equation (16). In step S115, if convergence at the value the same as the preceding resulting value for the human-induced probability is achieved is determined, and if Yes, the process is ended, and if No, the process returns to step S31.

FIG. 19 is a flow chart showing a process of estimation of a human position (a plurality of persons) of step S7 described above.

As shown in FIG. 19, in step S71, N samples are generated for random human positions and are made the initial value, and these are made the initial set of posterior samples, and the time is made the initial time (t:=1). In step S73, each of the posterior samples at a time (t−1) is moved according to the movement model of a person, and a set of predicted samples at a time t is generated. In step S75, (1) one sample j is selected from the set of predicted samples, and in step S77, (2a) one device is selected, and in step S79, (2b) the likelihood map of the device and the human-induced probability of the device at a time t are called up from the memory. In step S81, (2c) the weight of the device is calculated by referring to a sample human position selected from the likelihood map of the device and multiplying the human position and the human-induced probability of the device, and then, the process proceeds to (3).

As shown in FIG. 20, in step S83, (3) weights calculated by performing the processes of (2a), (2b) and (2c) described above on all the devices while changing the selected device are added up to thereby calculate the weight

π_t^[j]of the selected sample j. In step S85, the processes of (1) to (3) described above are performed on all the samples while changing the selected sample. In step S87, samples are duplicated by a number of pieces proportional to N×the weight of sampleπ_t^[j]for each sample in the set of predicted samples, and a set of posterior samples at the time t is generated, and the process proceeds to (4).

As shown in FIG. 21, in step S95, an initial association probability and an association probability in a previous frame are determined for each sample, and in step S97, a mixed normal distribution for M persons is estimated, and in step S99, whether the estimated value is converged is determined, and in step S101, if Yes, the process is ended, and if No, the association probability of each sample is updated, and the process returns to step S97 and the same process is repeated until the estimated value is converged.

(Real Life Experiment)

The following real life experiment was conducted to prove that the human position estimation apparatus 1 is capable of estimating the behavior of a person at home based on the power consumption pattern in a smart apartment room.

The number of residents was one at a time and three in total, and 26 STs were installed. Devices were fixed to 21 of the STs, and the plugs of the devices were not inserted into or removed from the STs. The devices were arranged at positions shown in FIG. 4.

With respect to the life pattern, residents were out during the day, and were at home from night to morning. No other restrictions were imposed. The experiment was conducted on three subjects, three days for each subject. The three persons were asked to record their actions in the room and which devices they used. Three days were consecutive or non-consecutive. The three subjects are subjects A, B, and C. Basic information for each subject is as follows.

Subject A: Male in twenties, student, 2 days+1 day

Subject B: Female in twenties, student, 1 day×3

Subject C: Male in twenties, student, 3 days

The approximate power consumption of each hour was obtained for each subject based on the power consumption pattern of each subject. According to the power consumption patterns (not shown), the patterns of life that subject A returned home at about 10 o'clock at night, went to bed at about 3 o'clock in the morning, and went out at about 12 o'clock at noon, and that subject B usually slept about 7 hours a day, and that subject C used power more and over a longer period of time at night than in the morning can be grasped.

(Human Position Estimation Result)

FIG. 22 is a diagram showing a probability distribution of the human position obtained by human position estimation means. In FIG. 22, the top left is (a), the top right is (b), the bottom left is (c), and the bottom right is (d). The time of occurrence of an event of an electrical device is shown in (a) of FIG. 22. An electrical device was highly possibly operated, and the distribution is dense in the periphery. A state at a slightly later time is shown in (b) of FIG. 22. It can be seen that the distribution is slightly spread in all the directions based on the movement model of a person. The distribution after a sufficient time has elapsed thereafter with no event is shown in (c) of FIG. 22. Also, (d) of FIG. 22 shows a state where, although an event has occurred with respect to an electrical device, since the distribution density around the electrical device 8 is low, the distribution is spread not only around the electrical device but also over the entire area. This is a part of the analysis result of day 1 of subject B.

Next, the estimation result of subject C, from returning home at night on day 3 to leaving home in the morning, is shown in FIG. 23 as the result of human position estimation. An optimal path determined from a probability distribution

P_h⁽¹⁾

obtained by performing position estimation once is drawn in order. A small black circle is the estimation position of the person for each 20 frames. A big black circle with a number written inside drawn as well indicates the electrical device with respect to which an event included in the event sequence E⁽⁰⁾ used for the estimation has occurred, and its position. The position connected to the big black circle by a thick solid line is the estimation position at the time of the event.

To observe the distribution as a whole, a graph with time as the horizontal axis and dispersion of distribution as the vertical axis is shown in FIG. 24. FIG. 24 is a dispersion diagram of positions of particles showing dispersion of a distribution and the time of occurrence of an event. The greater of the dispersion in the x direction and the dispersion in the y direction is made the dispersion of the distribution for the sake of simplicity. This is a part of the analysis result of day 1 of subject B.

FIG. 25 is an activity path diagram showing the path of activities obtained by repeated calculation by human probability communication means 123. A path according to the initial set of events E⁽⁰⁾ and

P_h⁽¹⁾

is shown in (a) of FIG. 25, and an activity path of a person according to E⁽¹⁾ and

P_h⁽²⁾

at the same time is shown in (b) of FIG. 25.

This (b) of FIG. 25 shows the activity path of a person for which the probability distribution of a human position z

P_h⁽ⁱ⁾(z)

and a set of probabilities of devices being in operation

P_m⁽ⁱ⁾

have been output by the human behavior estimation apparatus as estimation results, and it is proven that the human behavior estimation apparatus is capable of estimating the behavior of a person at home.

Claims

1. An on-demand power control system comprising a power source including at least a commercial power source, a plurality of electrical devices, a smart tap connected to the electrical devices, a human behavior estimation apparatus, including a memory, for estimating behavior of a person in a living space, and a network to which the human behavior estimation apparatus is connected via the smart tap, wherein the human behavior estimation apparatus includes initial human-induced probability value estimation means for comparing a feature obtained from a power consumption pattern of an electrical device with a feature of learning data obtained in advance for each electrical device to estimate a state of the electrical device, and estimating an initial value of a human-induced probability of the electrical device based on the estimated state of the electrical device,human position estimation means for calling up the initial value of the human-induced probability of the electrical device and a likelihood map of the device from the memory, performing a process, for all samples, of referring to a sample human position selected from the likelihood map and calculating a weight of the device by multiplying the human position and the human-induced probability of the device, and estimating a probability of a human position at each time point until a final time, andhuman-induced probability re-estimation means for performing recalculation of the human-induced probability based on the human-induced probability and a human position probability, performing the recalculation of the human-induced probability until a value of the recalculation converges, and outputting the human-induced probability and the human position probability when the value is converged.

2. The on-demand power control system according to claim 1, wherein the human behavior estimation apparatus estimates the behavior of a person based on the human-induced probability that is a probability of a person performing a human-induced operation on a device and the human position probability that is a probability of a person existing at a position.

3. The on-demand power control system according to claim 2, wherein the human behavior estimation apparatus determines feature data of an operation state based solely on the power consumption pattern of each electrical device, and determines the human position and the human-induced probability by repeated calculation while taking a human-induced probability obtained from the data as an initial value, to estimate both the human position and the human-induced probability.

4. The on-demand power control system according to claim 3, wherein the feature data of an operation state is composed of an id of an electrical device, an operation state of the device, and a feature of the device.

5. The on-demand power control system according to claim 4, wherein the power consumption pattern of an electrical device is generated based on a change in an operation mode or a state of an electrical device due to operation of the device by a sequence of human states including movement and stopping.

6. The on-demand power control system according to claim 4, wherein the feature of an electrical device is an average and dispersion of power consumption.

7. The on-demand power control system according to claim 1, wherein the initial human-induced probability value estimation means is capable of estimating the initial value of the human-induced probability based on a probability of a state of an electrical device, by having a probability of a state shift of the electrical device being caused by a human-induced operation given in advance.

8. The on-demand power control system according to claim 1, wherein the human position estimation means estimates a movement position of a person based on a human-induced probability according to which a two-dimensional position of the person at a time t may be determined by repeated calculation for a preceding time t−1.

9. The on-demand power control system according to claim 8, wherein the human position estimation means uses a time-series filter to estimate the human position probability.

10. The on-demand power control system according to claim 9, wherein the time-series filter is a particle filter, a Kalman filter or a moving average filter.

11. The on-demand power control system according to any one of claims 8 to 10, wherein the human position estimation means calculates the human-induced probability by using a likelihood map that is based on a relationship between a position of an electrical device and a human position.

12. The on-demand power control system according to claim 11, wherein the likelihood map shows, by using a pixel value, an existence probability of a person who is capable of pressing a power switch of an electrical device, based on a use status of the electrical device in an actual living space.

13. The on-demand power control system according to claim 12, wherein, in the likelihood map, calculation is performed using a pixel value of zero for a range that cannot be reached by a human hand if there is an obstacle.

14. The on-demand power control system according to claim 12, wherein, in the likelihood map, small pixel values are assigned to a wide range in a case where the electrical device may be operated by a remote control.

15. The on-demand power control system according to claim 12, wherein, in the likelihood map, in a case where the electrical device may be operated from a plurality of positions, large pixel values are assigned to a plurality of narrow ranges allowing operation.

16. The on-demand power control system according to claim 13, wherein, in the likelihood map, the pixel values are expressed based on a floor plan, the obstacle, position/distance of a power switch, and a position of the electrical device.

17. The on-demand power control system according to claim 16, wherein, in a case where there are a plurality of persons, the human behavior estimation apparatus performs calculation using a mixed normal distribution for the number of persons.

18. A program for causing a computer to operate as a human behavior estimation apparatus on an on-demand power control system including a power source including at least a commercial power source, a plurality of electrical devices, a smart tap connected to the electrical devices, the human behavior estimation apparatus, including a memory, for estimating behavior of a person in a living space, and a network to which the human behavior estimation apparatus is connected via the smart tap, wherein the human behavior estimation apparatus includes initial human-induced probability value estimation means, human position estimation means, and human-induced probability re-estimation means, andwherein the program causes the computer to perform processes of:by the initial human-induced probability value estimation means, comparing a feature obtained from a power consumption pattern of an electrical device with a feature of learning data obtained in advance for each electrical device to estimate a state of the electrical device, and calculating an estimation of an initial value of a human-induced probability of the electrical device based on the estimated state of the electrical device,by the human position estimation means, calling up the initial value of the human-induced probability of the electrical device and a likelihood map of the device from the memory, performing, for all samples, a process of referring to a sample human position selected from the likelihood map and calculating a weight of the device by multiplying a human position and the human-induced probability of the device, and calculating an estimation of a probability of a human position at each time point until a final time, andby the human-induced probability re-estimation means, performing recalculation of the human-induced probability based on the human-induced probability and a human position probability, performing the recalculation of the human-induced probability until a value of the recalculation converges, and outputting the human-induced probability and the human position probability when the value is converged.

19. A program for causing a computer to operate as a human behavior estimation apparatus according to claim 1, the program causing the computer to perform a process of: estimating, behavior of a person based on a human-induced probability that is a probability of a person performing a human-induced operation on a device and a human position probability that is a probability of a person existing at a position.

20. A program for causing a computer to operate as a human behavior estimation apparatus according to claim 1, the program causing the computer to perform a process of: determining feature data of an operation state based solely on a power consumption pattern of each electrical device, and determining a human position and a human-induced probability by repeated calculation while taking a human-induced probability obtained from the data as an initial value, to estimate both the human position and a human-induced operation.

21. A computer-readable recording medium recording a program according to claim 18.

22. A computer-readable recording medium recording a program according to claim 19.

23. A computer-readable recording medium recording a program according to claim 20.