JP3361017B2 - Comfort index PMV learning device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control technique for an air conditioner, and is useful for calculating an optimum room temperature set value for air conditioning control that maintains the comfort of residents in a building and saves energy. The invention relates to a device for learning the comfort index required in.

[0002]

2. Description of the Related Art The original purpose of air conditioning control is to ensure an appropriate indoor thermal environment. It is also desirable that this can be achieved with the least possible energy consumption.

In order to ensure an appropriate indoor thermal environment in consideration of human comfort, it is important to consider the human thermal sensation of heat and cold. The variables that affect this are:

(A) air temperature, (b) relative humidity, (c)
Average radiant temperature, (d) air velocity, (e) activity state (internal heat generation of human body), (f) clothing state The heat generation amount of a person is the radiation amount by convection, the heat radiation amount by radiation, the evaporation heat amount from a person, breathing When these heat balance equations are satisfied by the total of the heat radiation amount and the heat storage amount by the human body, the human body is thermally neutral and is in a comfortable state that is neither hot nor cold. On the contrary, when the heat balance equation is broken, the human body feels hot and cold.

In this regard, the Fanger of the Technical University of Denmark
In 1967, Professor announced the derivation of the comfort equation, and using this as a starting point, the heat load of the human body and the human thermal sensation were combined by statistical analysis from a questionnaire of a large number of European and American subjects, and PMV (Predicted MeanVote) : Predicted average answer)
Proposed. It has been taken up by the ISO standard in recent years and has become popular recently. PMV, which is an index of thermal sensation, is expressed as a numerical value according to the following 7-step evaluation scale. +3: Hot +2: Warm +1: Slightly warm 0: Neither, Comfortable -1: Slightly cool -2: Cool -3: Cold The comfortable range is -0.5 to +0.5.

Of the above six variables, the activity state, which represents the work intensity, is usually the metabolic rate met, and the clothing state is clo.
The unit of is used. MET (Met) A unit representing the amount of metabolism, which is based on resting metabolism in a thermally comfortable state.

1 met = 58.2 W / m 2 = 50 kcal / m 2 .h clo (black) A unit representing the heat insulation of clothes, and 1 clo is the temperature 21
℃, relative humidity 50%, airflow 5cm / s or less in the room, the amount of heat released from the body surface is in equilibrium with the metabolism of 1met, the value in the clothes state, converted to a normal thermal resistance value 1clo
= 0.155m2 ・ ° C / W = 0.18m2 ・ h ・ ° C / k
It becomes cal.

The detailed contents of the PMV are omitted here, but are described in the following references.

(A) Japanese Patent Application No. 3-288287 (Japanese Patent Laid-Open No. 5-126380) Name of the invention: Air-conditioning control device (b) Air Conditioning and Sanitary Engineering Society (ed): Air Conditioning and Sanitary Engineering Handbook 1, The changes to the 6 variables in PMV of Chapter 1 and Chapter 3 are very complicated because they are multidimensional, but the changes to one variable can be represented in a two-dimensional diagram. Examples of this characteristic are shown in FIGS.
Is. That is, FIG. 7 shows an example of the relationship between room temperature (air temperature) and PMV, FIG. 8 shows an example of the relationship between average radiation temperature and PMV, FIG. 9 shows an example of the relationship between relative humidity and PMV, and FIG. 10 shows wind velocity (air velocity). ) And PMV are plotted respectively.

The conditions of the variables not changed in each figure are as follows (for example, in FIG. 7, values other than room temperature are fixed to the following values). Solid line (winter) Dotted line (summer) Room temperature (air temperature) 22 ℃ Room temperature (air temperature) 25 ℃ Average radiant temperature 22 ℃ Average radiant temperature 25 ℃ Relative humidity 50% Relative humidity 50% Wind velocity (air velocity) 0.1m / s Wind velocity (air velocity) 0.1 m / s Clothes amount 1.0 clo Clothes amount 0.5 clo Activity amount 1.2 met Activity amount 1.2 met As can be seen from FIG. 10, when v (wind velocity) ≤ 0.1 m / s, PMV The value does not change with v.

The PMV of the comfort index is as described above.
It was obtained by Professor Fanger through a statistical analysis of a large number of Western subjects. However, the standard of human thermal sensation varies depending on race, social condition (customs and economic level), physiological climate adaptation effects due to geographical conditions, and facility environmental conditions. If it is used as it is for each Japanese air conditioning control system, it may deviate from the Japanese sense. Therefore, in the comfortable air conditioning control system, the comfort index corresponding to each individual air conditioning system and the occupant therein is learned by using the neuro technique, as indicated by reference numeral 11 in FIG. This Figure 11
It is quoted from the above-mentioned reference (Japanese Patent Application No. 3-288287).

The disclosure content of the above-mentioned reference is to input a plurality of process variables that affect the human thermal sensation, and to neural network from these input process variables and questionnaire data of the comfort level felt by the occupants. A neuro P that learns the comfort index PMV according to each air conditioning system and its occupants by the network (NN)
It is an air-conditioning control device that inputs a MV learning device and a learned PMV, and obtains a room temperature set value by a fuzzy arithmetic device so that the comfort index satisfies energy saving and falls within a comfortable range.

Multiple regression analysis is available as a learning method that can replace this neuro technique. This is a method of expressing a certain variate by a linear combination of a plurality of variates that are considered to be factors that determine the variate, and analyzing or learning the structure between them.

The resulting variable is y, and the possible variables are x1, x2, ..., Xr. Then, it is assumed that these data are given as follows.

[0015]

[Table 1] y is an objective variable or a dependent variable, x1, x2, ..., xr
Are called explanatory variables or independent variables. x1, x2,
A linear model for predicting and learning y from xr is yi = a0 + a1 · xi1 + a2 · xi2 +… + ar · xir
+ Ei ei represents an error component which cannot explain yi by xi1, xi2, ..., Xir. Where the sum of squares of error Σ (eix
ei) to minimize a0, a1, a2, ..., Ar

[0016]

[Equation 1] And A description of how to obtain this coefficient is omitted here.

When this regression coefficient is determined, the theoretical value of yi is

[0018]

[Equation 2] Calculated by

[0019]

When a neural network is used to learn PMV, which is often used as a reference for the degree of thermal comfort felt by humans, PMV, as already mentioned, has many variables that affect it. Since the variable is a non-linear function, the number of teacher signals required for learning becomes at least 50,000 and the number of learning calculations becomes enormous. However, it was still difficult to accurately learn 6 variables in all areas.

Further, even if an attempt is made to perform learning by using multiple regression analysis instead of the neural network, PMV becomes a non-linear model with respect to six variables affecting PMV, so that it is not possible to use the conventional multiple regression analysis as it is. Can not.

The present invention has been made in consideration of the above points, and is necessary for calculating the optimum room temperature set value in the air conditioning control which maintains the comfort of the occupant and realizes energy saving. To realize a method for significantly reducing the number of teacher signals and the number of times of learning and improving accuracy in a device that learns with a neural network to correct a comfort index based on PMV as a standard of residents' declarations. With the goal.

It is another object of the present invention to realize a method capable of learning and predicting a comfort index with high accuracy by using multiple regression analysis instead of a neural network in the same air conditioning control.

[0023]

In order to achieve the above object, the present invention provides an air conditioning process according to claim 1, in which the amount of clothes and the amount of activity affecting the PMV value are predetermined.
In a device for learning a comfort index PMV used for air conditioning control by a neural network, an air temperature, an average radiation temperature, a relative humidity, which are variables that affect the PMV value as an input to the neural network, and two of these three variables Comfort index P with a total of 6 variables
An MV learning device according to claim 2, wherein the comfort index PMV used for air conditioning control is learned by a neural network in an air conditioning process in which the amount of clothing and the amount of activity affecting the PMV value are predetermined. As inputs to, the air velocity of the variables affecting the PMV value, the air velocity squared, the six variables of the neural network input in claim 1, and the multiplication of the air velocity with each of the six variables, and the square of the air velocity. A comfort index PMV learning device for making a total of 20 by multiplying with each of the 6 variables, and a device for learning a comfort index PMV used for air conditioning control by a neural network according to claim 3, which affects the PMV value. For the amount of clothing and the amount of activity to be given, a combination of values used for each process to be controlled by air conditioning is determined in advance. , A comfort index for automatically selecting the corresponding neural network from the set clothing amount and activity amount when the neural network is provided with the number of combinations, PMV is learned for each neural network, and when this is used for air conditioning control PMV learning device,
The comfort index PMV learning device according to claim 4, wherein the comfort index PMV learning device according to claim 3 has the six variables according to claim 1 as an input to each neural network, and the comfort index PMV learning device according to claim 5, In the comfort PMV learning device of claim 3, a comfort index PMV learning device having 20 variables according to claim 2 as an input to each neural network, and a comfort PMV learning of claim 4 according to claim 6. When the comfort index PMV is used for air-conditioning control in the device, and when a value between the clothing amount and the activity amount is set to a predetermined combination case, two-dimensional output is performed using a plurality of neural network output values in the vicinity thereof. A comfort index PMV learning device for obtaining a PMV value by linear interpolation, and a comfort index PM for air conditioning control in the comfort PMV learning device according to claim 5 according to claim 7. When using a case where the amount of clothing and activity amount is set to a value between the predetermined combination case,
A comfort index P for obtaining a PMV value by two-dimensional linear interpolation using a plurality of neural network output values in the vicinity thereof
The MV learning device, the neural network in the comfort index PMV learning device according to claim 6 according to claim 8, when a teacher signal is given between combination cases in which the clothing amount and the activity amount are predetermined. In the comfort index PMV learning device according to claim 7, the comfort index PMV learning device for learning PMV using a teacher signal for a plurality of neural networks in the vicinity of the PMV learning process according to claim 7. , PMV learning of a neural network in the case where a teacher signal between a predetermined combination of the amount of clothing and the amount of activity is given, the teacher signal is used for a plurality of neural networks in the vicinity of the PMV learning.
A comfort index PMV learning device for learning MV and the comfort index PMV learning device according to any one of claims 1 to 9, wherein the neural network is replaced with a multiple regression analysis model. And a comfort index PMV learning device.

According to the present invention, by taking the above means, the neural network is divided for each case of the clothing amount and the activity amount to be used, and each neural network is allowed to share the learning only in that region, and The input of each neural network does not simply use the PMV variable itself, but considers the output characteristics of the PMV with respect to the variable and inputs the PMV variables in combination,
A comfort index PMV learning device that significantly reduces the number of teacher signals and the number of calculations and improves accuracy can be realized.

According to the present invention, the comfort index can be accurately and simply learned and predicted by using the multiple regression analysis.

[0026]

FIG. 1 shows a first embodiment of the present invention. If the target building for air conditioning control is, for example, a general office building, the activity amount that affects the PMV value is 1.2 me.
In most cases, the amount of clothing is predetermined, such as 0.5 clo during summer cooling and 1.0 clo during winter heating. In this case, these variables can be omitted as inputs to the neural network.

When the air velocity v is v ≦ 0.1 m / s, the value of PMV does not change with v (wind velocity and P
For the relationship of MV, see FIG. 10), and when the air conditioning control is the constant air volume method, v can be regarded as constant. Therefore, also in this case, v can be omitted as an input to the neural network.

As described in the prior art, room temperature and PM
The relationship between V, average radiation temperature and PMV, and relative humidity and PMV is almost linear if other conditions are fixed (see FIGS. 7 to 9). With this in mind,
As shown in FIG. 1, the inputs to the neural network are a total of 6 variables of temperature, average radiation temperature, relative humidity, and each multiplication.

The initial weight value at the start of learning the neural network is obtained in advance using the conventional Fanger's PMV calculation formula (the PMV calculation part in FIG. 1). The setting unit in FIG. 1 is for inputting, as a teacher signal, the data collected from the questionnaire about the comfort level of the occupant during the air conditioning control.

FIG. 2 shows a second embodiment of the present invention. As for the amount of clothes and the amount of activity, a predetermined air conditioning process is performed as in the first embodiment. However, when the air conditioning control system is the variable air volume system and the air velocity v ≧ 0.1 m / s, the PMV also changes with v. Air velocity and PMV
The relationship with and is close to a quadratic curve if other conditions are fixed (see FIG. 10). This, room temperature and PMV,
Taking into account the relationships between the average radiant temperature and PMV, and the relative humidity and PMV (see FIGS. 7 to 9), the airflow velocity and the square of airflow velocity, which are variables that affect the PMV value, are input to the neural network. , The six variables of the neural network input in the first embodiment, and the multiplication of the air velocity with each of the six variables, and the square of the air velocity with each of the six variables. That is, the following 20 pieces are used.

[0031]

[Equation 3] Ta: air temperature, Tr: average radiation temperature, Rh:
Relative humidity, v: air velocity FIG. 3 shows a third embodiment of the present invention. When considering a general office building as the target building, the normal clothing amount is summer:
0.5clo, winter: 1.0clo, with an activity level of 1.
It is 0 to 1.4 met (office work). Therefore, if 25 neural networks are prepared according to the following 5 × 5 combination case example and PMVs are learned for each neural network, it can be selected from any of them when applied to any general office building. can do.

[0032]

[Table 2] As the input to the neural network, Ta, Tr, and Rh (v is also input when PMV changes with v) in the example of FIG.

When these neural networks are used for air-conditioning control, appropriate clothing amounts and activity amounts are selected in advance from Table 2 and set in the neural network selection section.
The neural network selection unit selects one of the switch groups (SW1 to SW25) on the output side of the neural network and turns it on.

FIG. 4 shows a fourth embodiment of the present invention. In the fourth embodiment, the input to each neural network of the third embodiment described with reference to FIG. 3 is the same variables as those in the first embodiment, which are variables affecting the PMV value, that is, air temperature and average radiation. There are a total of 6 variables of temperature, relative humidity and each multiplication.

In the fifth embodiment of the present invention, the input to each neural network of FIG. 3 of the third embodiment shown in FIG. 3 is a variable that affects the PMV value similar to that of the second embodiment. The airflow velocity, the square of the airflow velocity, the six variables of the neural network input in the first embodiment, the multiplication of the airflow velocity with each of the six variables, and the multiplication of the square of the airflow velocity with each of the six variables, Total of 20
To be individual.

FIG. 5 is a diagram showing a sixth embodiment of the present invention. This sixth embodiment is characterized in that the comfort index PMV in the fourth embodiment shown in FIG. 4 is obtained, and the amount of clothing and the amount of activity are in a predetermined combination case (for example, Table 2). When a value between is set, the PMV value is obtained by two-dimensional linear interpolation using a plurality of neural network output values in the vicinity thereof.

A specific example of the two-dimensional linear interpolation method will be described with reference to FIG. It is assumed that the clothing amount of 0.6 clo and the activity amount of 1.15 met are preset in the neural network selection unit (point P). Among the neighboring points A, B, C, D, a neural network of three points (points A, B, D in the example) excluding the point farthest from the point P (point C in the example), The neural network selection unit selects. In the two-dimensional linear interpolation section, PM corrected by the following method
Obtain the V value.

Since the equation of the plane can be expressed by z = a.x + b.y + c, three points are enough to determine the equation of the plane.

From each input value to the neural network, the output value obtained by the neural network corresponding to the point A is zA, and from each input value to the neural network, the output value obtained from the neural network corresponding to the point B is zB. , ZD = 0.5 · a + 1.2 · b + c (10) zB = 0.5 · a + 1.0, where zD is the output value obtained by the neural network corresponding to point D from each input value to the neural network. * B + c (11) zD = 0.75 * a + 1.2 * b + c (12) From these three formulas, a, b, and c can be found. This value is a0, b
If 0 and c0, the corrected PMV is obtained as PMV = 0.6.a0 + 1.15.b0 + c0 (13).

The seventh embodiment of the present invention is characterized in that the comfort index PMV in the fifth embodiment is obtained, and the amount of clothing and the amount of activity are predetermined in a combination case (for example, Table 2). When the value of is set, the PMV value is obtained by two-dimensional linear interpolation using a plurality of neural network output values in the vicinity thereof. The specific method of the two-dimensional linear interpolation method is the same method as the sixth embodiment described with reference to FIG.

The eighth embodiment of the present invention is the same as the sixth embodiment, except that the PMV learning of the neural network in the case where the teacher signal is given between the combination cases in which the clothing amount and the activity amount are predetermined. At this time, a plurality of neural networks in the vicinity of the PMV are learned using the teacher signal.

A specific method example will be described below. Similar to the description in the sixth embodiment, the neural network selection unit previously has a clothing amount of 0.6 cl and an activity amount of 1.15 me.
It is assumed that the value of t has been set (point P in FIG. 6). At this time, it is assumed that the value of each input variable and the value of PMV (z1) are given as a teacher signal. The neural network selection unit uses the teacher signal to perform learning as a neural network, and in the same manner as in the sixth embodiment, points A,
Select B and D. The PMV value before learning at the point P for the input variable is 0.6 from the equations (10) to (13).
* A0 + 1.15 * b0 + c0. When this is set to z, the deviation Δe from the teacher signal is Δe = z1−z.

Therefore, the learning signal to the neural network corresponding to the point A is learned by assuming that zA in the equation (10) is added with Δe.

The teacher signal to the neural network corresponding to the point B is learned by assuming that zB in equation (11) is added with Δe.

The teacher signal to the neural network corresponding to the point D is learned by assuming that zD in the equation (12) is added with Δe.

The ninth embodiment of the present invention relates to the PMV learning of the neural network in the case where the teacher signal between the combination cases of the clothing amount and the activity amount in the seventh embodiment is given in advance. It has a characteristic, and allows a plurality of neural networks in the vicinity to learn PMV by using the teacher signal. A specific method example is the same as the method example described in the eighth embodiment.

The tenth embodiment of the present invention will be described below. In the tenth embodiment, a multiple regression analysis model is used instead of the neural network in the first embodiment or the second embodiment.

First, in order to replace the neural network in the first embodiment with a multiple regression analysis model, the multiple regression model is y = a0 + a1 .x1 + a2 .x2 + ... + a6 .x6 +
As ei, the explanatory variables x1, x2, ..., X6 are set as follows.

X1 = Ta, x2 = Tr, x3 = Rh, x
4 = TaTr, x5 = TaRh, x6 = TrRh As in the case of the neural network, a0, a1,
The initial value of a6 is prepared by using the conventional Fanger's PMV arithmetic expression and is obtained by the multiple regression analysis method. In addition, as with the neural network, the comfort index for each air conditioning system and the occupants there are multiple regressions using the data collected from the questionnaire about the occupants' comfort during air conditioning control. Learn by analytical methods.

To replace the neural network in the second embodiment with a multiple regression analysis model, the multiple regression model is y = a0 + a1.x1 + a2x2 + ... + a20.x20 + e.
As i, the above explanatory variables x1, x2, ..., X20 are the second
Input variables 1, 2, ..., 20 to the neural network shown in the embodiment of FIG.

In the following, also for the third to ninth embodiments, the tenth embodiment is obtained by replacing the neural network with the above multiple regression analysis model in the same manner.

[0052]

As described above, the present invention provides the comfort index,
In a device that learns with a neural network to correct the resident's declaration based on PMV, the neural network is divided for each case of clothing amount and activity amount used, and each neural network has only that region. Learning is shared, and the input of each neural network does not simply use the PMV variable itself.
Since measures such as inputting in combination with V variables are taken,
It is possible to provide a comfort index PMV learning device that significantly reduces the number of teacher signals and the number of calculations and improves accuracy.

Further, in a device for learning with a multiple regression analysis model for correcting based on PMV as a resident's declaration, a multiple regression analysis model is prepared for each case of clothing amount and activity amount used, Each multiple regression analysis model is made to share learning only in that area, and the input of each multiple regression analysis model does not simply use the PMV variable itself, but considers the output characteristics of the PMV for the variable and considers each PM
Since measures such as inputting in combination with V variables are taken,
A comfort index PMV learning device that significantly reduces the number of learning calculations and improves accuracy can be provided.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram showing a first embodiment of the present invention.

FIG. 2 is an overall configuration diagram showing a second embodiment of the present invention.

FIG. 3 is an overall configuration diagram showing a third embodiment of the present invention.

FIG. 4 is an overall configuration diagram showing a fourth embodiment of the present invention.

FIG. 5 is an overall configuration diagram showing a sixth embodiment of the present invention.

FIG. 6 is a diagram for explaining a specific example of the two-dimensional linear interpolation method according to claim 6;

FIG. 7 is a diagram showing a relationship between room temperature (air temperature) and a conventional PMV.

FIG. 8 is a diagram showing a relationship between an average radiation temperature and a conventional PMV.

FIG. 9 is a diagram showing the relationship between relative humidity and conventional PMV.

FIG. 10 is a diagram showing the relationship between wind speed (air velocity) and conventional PMV.

FIG. 11 is an overall configuration diagram of a conventional comfortable air conditioning control system.

[Explanation of symbols]

1A Neuro PMV learning device 2A Resident's comfort setting section 3A Neural network selection unit

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Nobutaka Nishimura Nobutaka Nishimura 1-1-1, Shibaura, Minato-ku, Tokyo Inside the Toshiba Headquarters office (56) Reference JP-A-6-147599 (JP, A) JP-A-3-125257 (JP, A) JP-A-5-280787 (JP, A) JP-A-4-328669 (JP, A) JP-A-4-320750 (JP, A) JP-A-6-82080 (JP, A) JP 5-126380 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) F24F 11/02

Claims (10)

(57) [Claims] 1. A comfort index PMV learning device for learning a comfort index PMV used for air conditioning control by a neural network in an air conditioning process in which the amount of clothing and the amount of activity that affect the PMV value are predetermined. The comfort index PMV is characterized in that, as the input of, the air temperature, the average radiation temperature, the relative humidity of the variables that affect the PMV value and the multiplication of two of these three variables in total are six variables. Learning device. 2. A comfort index PMV learning device for learning a comfort index PMV used for air conditioning control by a neural network in an air conditioning process in which the amount of clothing and the activity amount that affect the PMV value are predetermined. The airflow velocity, the square of the airflow velocity, which is a variable affecting the PMV value, as an input to
2 of the neural network input, and the multiplication of the air velocity with each of the 6 variables, and the square of the velocity with each of the 6 variables.
A comfort index PMV learning device characterized in that the number is 0. 3. A comfort index PMV learning device for learning a comfort index PMV used for air conditioning control with a neural network, wherein the amount of clothing and the amount of activity affecting the PMV value are used in advance for each air conditioning controlled process. Decide the combination of values,
A neural network is provided by the number of the combinations, PMV is learned for each neural network,
A comfort index PMV learning device characterized by automatically selecting a corresponding neural network from the set clothing amount and activity amount when using this for air conditioning control. 4. The comfort index PMV learning device according to claim 3, wherein the inputs to each neural network are the six variables of claim 1 and the comfort index PMV.
V learning device. 5. The comfort index PMV learning device according to claim 3, wherein the inputs to each neural network are the 20 variables in claim 2, and the comfort index P is characterized.
MV learning device. 6. The comfort index PMV learning device according to claim 4, when the comfort index PMV is used for air conditioning control, and when the clothing amount and the activity amount are set to values between predetermined combination cases, A comfort index PMV learning device, wherein PMV values are obtained by two-dimensional linear interpolation using a plurality of neural network output values in the vicinity thereof. 7. The comfort index PMV learning device according to claim 5, wherein when the comfort index PMV is used for air conditioning control, when the clothing amount and the activity amount are set to a value between predetermined combination cases, A comfort index PMV learning device, wherein PMV values are obtained by two-dimensional linear interpolation using a plurality of neural network output values in the vicinity thereof. 8. The comfort index PMV learning apparatus according to claim 6, wherein the learning signal is given during the PMV learning of the neural network when the teacher signal between the combination cases of the clothing amount and the activity amount is predetermined. A comfort index PMV learning device, characterized in that a plurality of nearby neural networks are trained to learn PMVs by using their teacher signals. 9. The comfort index PMV learning device according to claim 7, wherein the neural network PMV learning is performed when a teacher signal is given between a combination case in which the clothing amount and the activity amount are predetermined. A comfort index PMV learning device, characterized in that a plurality of nearby neural networks are trained to learn PMVs by using their teacher signals. 10. The comfort index PMV learning device according to claim 1, wherein the neural network is replaced with a multiple regression analysis model.