JP2802469B2 - State prediction 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 state predicting apparatus for predicting an air-conditioning load or the like using an input / output model.

[0002]

2. Description of the Related Art Conventionally, when a load state of an air conditioning system or the like is predicted, a new situation is predicted using an autoregressive model described later or a model based on a combination of basic patterns. That is, the autoregressive model is a pattern matching model of an air-conditioning load created by determining a parameter of a mathematical expression using a case that has recently occurred, that is, a recent case. The parameters of the autoregressive curve are determined by calculating the following equation.

[0003]

(Equation 1)

Here, X (t) is heat load data as observation time series data, e (t) is white noise time series data, a
(I), b (i), p, and q are parameters, and these parameters are determined based on data for 10 days. When a model is created using a combination of basic parameters, basic patterns for weekdays, Saturdays (half-holiday) and holidays are determined from past statistical data, and a model is created using a combination of these patterns. That is, this basic pattern indicates a daily load every hour, and the basic pattern follows the actual load fluctuation by smoothing the load at the same time from the past to the present with respect to this basic pattern. Is what you do.

[0005]

However, in such an autoregressive model, data that can be actually handled due to limitations in memory capacity and calculation time includes several types of variable data such as loads for several hours and error values. It is. Therefore, what can be represented by these variables is actually the load pattern, and the external factors such as the outside temperature, which are the actual factors of the load, and the amount of change in the load caused by the external factors, etc. Factors are not simultaneously considered as variables,
There has been a problem that when a model is corrected, it cannot be corrected accurately, so that load prediction cannot be correctly performed.

Further, when correcting a model, it is difficult to determine how to treat a factor with respect to a correction target. In the case of performing a correction based on a number of factors, a correction formula or the like is used. Since it is difficult to determine, a phenomenon cannot be accurately described as a result, and accurate load prediction cannot be performed. In addition, because of the smoothing based on the data for the past 10 days, there is also a drawback that, for example, a hot day continues for 10 days and the 11th day suddenly cools down, so that it is inferior in the ability to follow a sudden change.

On the other hand, in the case of the prediction model based on the combination of the basic patterns, the load is made to follow the actual load fluctuation by smoothing. There is a disadvantage that. In addition, since a fixed pattern called the basic pattern is used, it is necessary to prepare a number of patterns in order to perform accurate description, and even if many patterns are prepared, This causes a problem of selecting a pattern. As described above, since the conventional method is based on the pattern, it is difficult to consider external factors, and in particular, it cannot cope with a sudden load change that cannot be described with a fixed pattern determined sequentially. was there. Further, since the predicted value is calculated using a mathematical model subjected to pattern matching, it is difficult to determine whether a new situation can be described by the generated model, and thus it is difficult to determine the credibility of the predicted value. There was a problem.

Accordingly, an object of the present invention is to make it possible to predict a sudden load change and determine the authenticity of the predicted value.

[0009]

SUMMARY OF THE INVENTION In order to solve such a problem, the present invention provides a method for predicting a state that will occur several hours later based on a relationship with an input factor. Means for categorizing according to a predetermined criterion, means for inputting case data of each type categorized according to a predetermined criterion, extracting factors for recognition and generating a recognition model based on the extracted factors, and Means to generate a causal relationship model between the input factors and the state several hours after the input of case data of each class, and to use the recognition model when a new situation occurs, Means for calculating the attribution of a class, means for calculating a predicted value in a new situation using a causal relationship model, and calculation of a final inference value by weighted averaging of the attribution and the predicted value of each class. Means, Means for calculating the credibility of the measured values, means for learning the new situation and the state that has actually occurred several hours later based on the recognition model, and means for learning the new situation and the state that has actually occurred several hours later. For learning based on the causal relation model.

In addition, when predicting the heat load of the air conditioning system, data of the outside air temperature, the change amount of the outside air temperature, the discomfort index, the heat load, and the change amount of the heat load are used as input factors, and the output factors of the input factors are used. Input factor determining means for determining how many hours before the input factor used for predicting the thermal load several hours later should be used, using each data of the heat load occurring several hours later, and this input factor determining means Input factor symbolizing means for converting the input factor determined by the above into symbolized data, modeling means for generating a model by associating the input factor with the output factor based on the symbolized data, and a new input factor. And an inference means for inferring using the above model when the inference is performed.

[0011]

When performing state prediction, a factor is extracted from each type of case data classified according to a predetermined criterion, and a recognition model is generated based on this, and a causal relationship model is generated from each type of case data. Is done. Then, when a new situation arises, the attribution to each class is calculated using the recognition model, the predicted value in the new situation is calculated using the causal relationship model, and the credibility of this predicted value is calculated. Is calculated. Then, a final inference value is obtained by performing a weighted average of the attribute values and the predicted values of these classes.

Further, when predicting the heat load of the air conditioning system, data of the outside air temperature, the change amount of the outside air temperature, the discomfort index, the heat load and the change amount of the heat load are used as input factors. The data of the heat load that occurs several hours later is used as the output factor, and the input factor used to predict the heat load several hours later is used to determine how many hours ago the data should be used. Is converted into symbolized data, and a model that associates input factors with output factors is generated based on the symbolized data. When a new input factor is input, the heat load is predicted several hours later using the model. Is performed.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing one embodiment of a system to which a state prediction device according to the present invention is applied. This system is an air conditioning load prediction system for district cooling and heating. In the figure, reference numeral 1 denotes a storage device for storing case data such as an outside air temperature observed at certain time intervals in the past and a discomfort index at this time, and 2 denotes a data classification unit for classifying the case data for each month. Numeral 10 creates a recognition model from the case data for each month, and when new data such as outside temperature is actually input, recognizes the attribution to each month based on this recognition model and performs load prediction. The state recognition automatic generation device 20 creates a causal relationship model for each month from the case data for each month, and when new data is input, makes inferences based on the causal relationship model and calculates its credibility. An inference apparatus 30 is a load predicted value display unit, 40 is a credibility integration unit, and 50 is a credibility display unit.

Here, the state recognition automatic generation device 10 includes a recognition model generation unit 11, a recognition model learning unit 12, a recognition model storage unit 13, a recognition model situation input unit 14, an attribute recognition unit 15, a final predicted value determination unit. 16. Then, the recognition model generating means 11 extracts factors for recognition from past case data to generate a recognition model based on each factor, and the recognition model learning means 12 generates a recognition model when inputting new case data such as outside air temperature. And updates the above recognition model, and the like, and these recognition models are stored in the recognition model storage unit 13. On the other hand, when the new case data is input from the recognition model situation input unit 14, the attribution recognizing unit 15 recognizes the attribution to each factor based on the recognition model in the storage unit 13, and finalizes the recognition result. In addition to sending the result to the value determining means 16, the final predicted value determining means 16 outputs a weighted average of the recognition result and the inference result of the inference apparatus 20, and outputs the result as a final load predicted value.

The inference apparatus 20 includes a causal relation model generating means 21, a causal relation model learning means 22, a causal relation model storage unit 23, a causal relation model situation input unit 24, an inference means 25, and a credibility calculation means 26. Have been. Then, the causal relationship model generating means 21 generates a model of a causal relation between the input factor and the result several hours later, for example, room temperature, with respect to the past case data, and the causal relation model learning means 22 generates a new case model. When data is input, it learns the data and updates the causal relation model, and these causal relation models are stored in the causal relation model storage unit 2.
3 is stored. On the other hand, when the new case data is input from the causal relationship model situation input unit 24, the inference means 25 makes an inference based on the causal relationship model in the storage unit 13 for this new case data, and outputs the inference result to the final prediction value determination. The load is sent to the means 16 and integrated with the recognition result by the automatic state recognition generation device 10 to display the predicted value on the load predicted value display section 30. Further, the inference means 25 sends the inference result to the credibility calculation means 26 to cause the credibility calculation unit 26 to perform credibility calculation and display the result on the credibility display unit 50. When inferring new case data, the inference means 25 sets the topology.
Inferring based on. The phase refers to a structure given to a set so that the concept of continuity can be defined, and indicates, for example, a distance between similar new case data and past case data, similarity between similar cases, and the like.

FIG. 2 is a functional block diagram showing a main part of the state predicting apparatus. The state predicting model generating unit A includes the data classifying unit 2, the recognition model generating unit 11 and the causal relation model generating unit 21. Recognition model learning means 12
A model learning unit B including a causal relationship model learning unit 22, a situation prediction unit C including an attribute recognition unit 15, an inference unit 25, and a final predicted value determination unit 16, a credibility calculation unit 26, and a load predicted value display. Unit 30 and credibility display unit 50
And a result display section D.

Here, referring to FIGS. 3 to 8, the data classification unit 2 and the recognition model generation unit 1 in the state prediction model generation unit A will be described.
The details of each function of the attribute recognition unit 15 in the state prediction unit 1 and the state prediction unit C will be described. Now, X _j (i = 1,..., N) is an input variable such as an outside air temperature and a discomfort index, and y is an output variable such as a heat load several hours later. Suppose there is a state model throughout the year like this. When creating a state model, creating a model for each variable, for example, an annual range, would result in a very complicated input / output space. , the state model M _i
(I = 1,..., 12) (that is, this month is a classification criterion for the data categorization of the data categorizing unit 2 and categorizes the yearly state). That is, for each set of input variables for each month, a model in the variable range for each month is created according to equation (2). For the sake of simplicity, the discussion proceeds assuming that an identification variable described later is included in the input variables.

[0018]

(Equation 2)

Further, the annual range for each input variable is determined according to equation (3). The situation of this annual range is as shown in FIG.

[0020]

(Equation 3)

Next, in order to create a hierarchy, a membership function for each input variable is created. Here, first, as shown in FIG. 4, each input variable is divided into m equal parts with respect to the annual range (digitization). I do.

Then, with respect to the range divided into m equal parts, a calculation is made as to which digit and how many digits the month data of each input variable belongs to. As a result of this calculation, for example, the data distribution such as the outside air temperature and the discomfort index is as shown in FIG.

Next, in order to calculate the possibility distribution and use it as a membership function, the maximum value (maxNO.ofi) of the data in the data distribution of each month as shown in FIG.
Perform normalization with. As a result, the membership function A _xj becomes as shown in FIG.

When the membership function of each state model is determined in this manner, factors for characterizing each state model are extracted by selecting variables for hierarchization.
That is, hierarchization is performed using variables that can identify the features of the model of each month among the variables X _j (j = 1,..., N). That is, here, the center of gravity C (A _xj ^Mi ) of the membership function of each month as shown in FIG. 8 is calculated, and the calculated value and the center of gravity C (A _xj ^M−2 ) of two months before and after each month are calculated. ), C (A _xj ^{M + 2} )
Calculate the distance to each. Then, according to equation (4), if the calculated distance is equal to or more than a certain threshold value q with respect to the annual range, the distance is automatically selected as the above-mentioned identification variable. Further, by displaying the distribution of such a membership function and notifying the operator, the operator can easily select an identification variable.

[0025]

(Equation 4)

When this discriminant variable is selected, in the present embodiment, the distance between the center of gravity value of the membership function of a certain month and the center of gravity values of two months before and after that is calculated, and one month before and after is calculated. Absent. This is to make the boundaries of each month ambiguous. Generally, the discrimination variable may be selected by setting the distance between the respective barycentric values C (A _xj ^Mi ) to a certain threshold value q 'or more. Further, it is easy for the operator to intuitively determine the distribution of the displayed membership functions.

Here, the number of membership functions defined for a certain range is counted, and the number is defined as N.
The relationship between the membership function N and the threshold value q is expressed by equation (5).
When the condition is satisfied, it is considered that the membership function is arranged at a position where it can be identified.

[0028]

(Equation 5)

Thus, the recognition model generating means 11
As a result, factors for characterizing each state model are extracted, and a recognition model is generated based on the factors.
Next, the membership recognizing unit 15 recognizes the membership (possibility) indicating which month the user is most likely to belong to the new input data. That is, the above-mentioned identification variable X _i (i
= 1 ′,..., N ′), the input data Xi
Given (t) (i = 1,..., N), the attribution Pos (Mi) to each month is determined based on the equation (6). That is,

[0030]

(Equation 6)

Therefore, the identification variables X _i (i = 1 ′,.
.., n ') means that the hierarchization has been performed. Next, the inference value of the hierarchical model thus determined is determined. That is, the input data Xi (t) (i = 1,...,
With respect to n), an inference value y (t) _Mi (i = 1,..., 12) is obtained using a previously created month model. As a result, the inference value y (t) is calculated based on Expression (7), and the inference value of the entire model obtained by integrating the state models is obtained.

[0032]

(Equation 7)

As described above, since a model recognizing the kind of complicated input / output relation can be automatically or easily generated by the judgment of the operator, the man-hour for model creation can be reduced and the complicated input / output state can be reduced. Since the output relation can be categorized, the description of the model is improved. For example, the monthly model, which is the state model of the lowest layer such as the annual model, has a range of each input / output space limited to the range of the moon. improves. In addition, only the input / output relation data (input / output variables) of the state model in the lowermost layer need be stored in the memory, and modeling can be performed with a small capacity memory. When there is new input data, the recognition model learning means 12 performs learning based on the above recognition model and updates it.

Next, the details of the functions of the causal relation model generation means 21 constituting the situation prediction model production section A, the inference means 25 constituting the situation prediction section C, and the final predicted value determination means 16 will be described. The inference means 25 handles the causal relationship between input factors such as the outside air temperature and the discomfort index of past case data and the result several hours later, for example, room temperature, and X = <x1, x2, as an input space. ..., x
n> and Y = <y> as the output space, events X1 (t) and X2 occurring at time t
(T),..., Xn (t) are output after α hours, and output Y (t +
α), that is, 入 出力 X1
(T), X2 (t), ..., Xn (t), Y (t +
α)｝ (t = 1,..., N), and infers data in which each input / output variable changes continuously.

Before inference is performed by the inference means 25, first, a case base is created for past case data by the causality model means 21. That is, an input space is first discretized and divided into a finite number of input events, and one case is generated by integrating input / output data belonging to the same input phenomenon. At this time, the condition part of the case, that is, the data such as the outside air temperature and the discomfort index is obtained by discretized input data ｛X1,
Xn}, and the conclusion of the case, that is, the room temperature after α hours, is the barycentric value Y of the output data, the number n of times the same input event has occurred, and the barycentric value ΔY /
ΔX1,..., ΔY / ΔXn, that is, ｛Y, n, ΔY /
ΔX1,..., ΔY / ΔXn}.

For example, as input data, the outside air temperature X1
(° C.) and the discomfort index X2 (%) are observed, and the maximum value (max) and the minimum value (min) of these data during each time t (t = 1,..., N) in a certain period are obtained. each, X1 (max) = 30.0, X1 (min) = 20.0 X2 (max) = 80.0, X2 (min) = 70.0 , and the and the data at time t = t ₁ X1 (t ₁ ) = 25.6, X2 (t ₁ ) = 78.7, and the discretization number for discretizing this input space is “10” (division between the maximum value and the minimum value is divided into 10). Then, the discretized data is, for example, X1 = 6, X2 =
9 and is symbolized by one event {6,9}. Here, the room temperature Y (t at time t = t ₁ + α)
_{If (1} + α) is 25.0, a causal relationship of {6, 9} → 25.0 is obtained.

When the event at time t = t ₁ +1 is X1 (t ₁ ) = 25.8, X2 (t ₁ ) = 78.79 and the room temperature Y (t ₁ + α + 1) = 25.5, The input events belong to the same {6, 9}, and {6, 9} → 25.25 [= {25.0 + 25.5} /
2, where n = 2] and the case data is compressed. As a result, the capacity of the memory required for the case base can be significantly reduced as compared with the conventional example. Further, each partial differential value described above is
This is the amount of change in output with respect to the amount of change in each input variable. In this case, since each input / output variable is continuous data, the partial differential value ΔY / ΔXi (t) can be calculated by equation (8).

[0038]

(Equation 8)

As described above, after a case base is created for a past case (existing case), inference about a new case is performed by the inference means 25 next. First, let the condition part of the new case be {Xi ^* } (i = 1, 2,..., N),
The existing case is defined as {Xi, Y, n, ΔY / ΔXi} (i = 1,
2,..., N). Here, the condition part of the new case is
At the same time as the input, it is discretized and symbolized as described above.

Here, when inferring a new case, first, the similarity of the existing case to the new case is determined (this similarity corresponds to the concept of a neighborhood system in phase). The similarity of the existing case to the new case is determined by the following definition. That is, the similarity 0 is | Xi ^* −Xi | = 0 (i = 1,
2,..., N) The similarity 1 is | Xi ^* −Xi | ≦ q _Xi (i = 1,
2,..., N) Similarity 2 is | Xi ^* −Xi | ≦ q _Xi +1 (i = 1,
2,..., N) The similarity 3 is | Xi ^* −Xi | ≦ q _Xi +2 (i = 1,
2,..., N). Here, q _Xi is called a threshold,
It is a digit value determined from variance of Xi (condition part of existing case) with respect to allowable accuracy of Y (conclusion part of existing case) from existing case data.

As an example, when X1 is the outside air temperature, X2 is the discomfort index, and the room temperature after Y = α hours is considered, when an error between the room temperature Y and the predicted value is attempted within two degrees, an input is made. Discretization is performed so that one digit becomes twice as in the case of space. If the discretization number is, for example, "20", Y ^* (conclusion part of a new case) belonging to the same digit for each of digits 1 to 20 is collected from existing case data, clustered, and belong to the same class. The variance of the conditional parts X1 and X2 of the new case is obtained. The variance of each cluster of the condition parts X1 and X2 of the new case belonging to the digit i of Y (conclusion part of the existing case) is calculated as a discrete value, and the digit value q is obtained by averaging with the number of clusters.
i _X1 and qi _X2 are obtained. Digit values qi _X1 and qi _X2 are averaged according to the following equation. That is, q _X1 = Σqi _X1 / 20, q _X2 = Σqi _X2 / 20

Here, the closest digit value is q _X1
= 2, q _X2 = 3 (actually, q _X1 = 2.123...,
q _X2 = 3.456...). However, if Y needs to be within 1 degree, q _X1 = 1 and q _X2 = 2, and the digit value q _Xi differs depending on the required precision. That is, the distance between the existing case and new case for each variable Xi is the phase of the condition part is smaller than q _Xi is closer, the conclusion part of the new case at that time is required for the conclusion of the existing case Considered to be within precision.

Next, the inference means 25 searches for a similar case to the new case. That is, the existing case of the optimal case is extracted as a similar case in the order of higher similarity to the new case. As the optimum number of cases, for example, the number of cases that makes the best inference from a simulation based on existing cases is selected. If there are more existing cases with the same degree of similarity than the optimal case number, the degree of influence of each variable Xi on Y, that is, the correlation coefficient R
_The priority is set for each variable according to the magnitude of _Xi and extracted.

Next, the inference means 25 determines the importance of the similar case with respect to the new case. In this case, a distance is defined in the input space and the phase between cases is considered. Here, a distance L as shown in Expression (9) is introduced as an example.

[0045]

(Equation 9)

Here, Φi is the weight of the distance in the variable Xi. Then, the importance Wj at the time of inference of the extracted m similar cases of the number of optimal cases is defined using Expression (10).
That is,

[0047]

(Equation 10)

Using the m similar cases extracted in this way, an inference value Y ^* for a new case Xi ^* (i = 1, 2,..., N) is calculated using Expression (11) and integrated. Become

[0049]

[Equation 11]

Here, Lij is the distance of the i-th case from the input data on the j-input variable axis, yi is the conclusion value of the i-th similar existing case, and ΔY / ΔXj is j of the i-th similar existing case. The percentage of the change that the second change gives to the conclusion value is shown.

Next, the credibility judgment of the inference result is performed based on the calculation result of the credibility calculating means 26. That is, the credibility of the inference result is determined using the similarity of the similar case used in the inference to the new case. For example, if the similarity used in the inference is similar to the new case, the highest similarity is defined as the credibility of the inference result.
The inference result with the highest similarity “1” is determined to have credibility “1”. In this case, the credibility “0” has the highest credibility, and the credibility decreases as the number increases. Then, the inference result obtained by the inference means 25 and the attribution result of the attribution recognizing means 15 are weighted and averaged by the final predicted value determining means (predicted value integrating means) 16 and displayed on the result display section D, while the authenticity is verified. The authenticity result of the sex calculating means 26 is also displayed on the result display section D.

When performing case-based learning, a new case is learned by the causal relationship model learning means 22 to update the case base. Such case-based updating is performed by the following procedure. However, those marked with * indicate new cases. That is, assuming that the number of events of the same condition part up to the previous time is n, the number of events is increased by one to n + 1, the output value Y is (Y × n + Y ^* ) / (n + 1), and the partial differential value ΔY / ΔX1 is (ΔY / ΔX1 × n
+ ΔY / ΔX1 ^* ) / (n + 1).

FIG. 10 is a graph showing the result of inference based on actual recording data. As a model, 19
In addition to using the recorded data every hour from June to November 1989, as the condition part of the case data, the outside temperature, the discomfort index, the change in the outside temperature, the change in the heat load, and the heat load were used. As the conclusion of the case data, the heat load after 4 hours is used. And classify case data every month,
The monthly case-based model was automatically generated by the causal relation model generation unit 21 in the state prediction model generation unit A.

Next, in order to recognize the attribution to each month,
When the recognition model generation unit 11 in the state prediction model generation unit A performed feature extraction for recognition on each condition part of the case data, that is, from outside temperature to heat load, the outside air temperature as shown in FIG. The discomfort index and the heat load were selected, and an attribute recognition model using these was automatically generated. Then, using these 1989 cases (knowledge), the membership recognizing unit 15 and the inference unit 25 in the situation prediction unit C provide 199
Load prediction was performed from August 27 to September 30 of the year 0. That is, an experiment was performed to predict a load (situation) occurring four hours after the external temperature to the heat load value, which is a condition part of the 1990 case data. The reason for selecting this period is that it is difficult to maintain the prediction accuracy with the conventional prediction model due to severe climate change.

According to the experimental results, the relative error of the conventional model (= | prediction error | / measured value) is 15.5%, while the relative error of the apparatus of this embodiment is 8.9%. ,
The prediction accuracy is approximately doubled, and the accuracy is improved. In addition, the temperature difference between September 6 and 7 is different by as much as 10 degrees, and the apparatus maintains good accuracy despite rainfall and rapid climate change. The reason for this is that September 6 is close to the state of June and October of the previous year, and this information was obtained from the credibility information displayed together with the prediction result.

Next, FIG. 11 is a diagram showing another embodiment of the present state prediction device, and is a functional block diagram showing a main part of the air conditioning load prediction device for predicting the heat load of the air conditioning system.
In FIG. 1, the air-conditioning load prediction device includes an input unit 10 for inputting input variables such as outside air temperature as past case data.
0, delay time determining means 101, input status symbolizing means 10
2. Input / output relation modeling means 103 corresponding to the causal relation model generation means 21 described above, and inference means 25 described above
Consists of The air-conditioning load predicting apparatus includes, as input variables (that is, condition parts of case data), the outside air temperature, the discomfort index, and the heat load shown in FIG. In addition to using the data of each quantity, and using the heat load data generated several hours later as the output variable (conclusion part of the case data), when a new situation occurs and the input variable is input, the heat This is to predict the load.

Here, the delay time determination means 101 receives the past case data from the input variable input means 100, the outside air temperature, the discomfort index, the change amount of the outside air temperature, the heat load, and the change amount of the heat load. When inputting the input variable data of the above, and further inputting the heat load data, which is an output variable actually generated in the air conditioning system several hours later, when the input time of this output variable is used as a reference, Determine how many hours ago the data should be. That is, for example,
The outside temperature data is the value one hour before the reference time,
The discomfort index data is determined two hours before, for each of the five input variable data, and the determined input variables are sent to the input situation symbolizing means 102 together with the output variables.

When input data of these input variables and output variables is input, the input situation symbolizing means 102 symbolizes the input variable data as described above, compresses the data, and sends the data to the input / output relation modeling means 103. send. The input / output relationship model generation means 103 creates an input / output relationship model by associating the symbolized input variables and output variables. Here, when the inference means 25 inputs an input variable such as a new outside air temperature, the inference means 25 infers an air-conditioning load predicted value of the air-conditioning system several hours later based on the input / output relation model generated by the input / output relation modeling means 103. And outputs the inference result. In this way, a model for predicting the air-conditioning load several hours later is generated by simultaneously considering each data of the outside air temperature, the change amount of the outside air temperature, the discomfort index, the heat load, and the change amount of the heat load as input factors. Therefore, it is possible to accurately grasp the phenomenon, and as a result, it is possible to perform accurate load prediction.

As described above, according to the present invention, in the field of control and operation based on state prediction, a large number of factors of the state after several hours can be taken into consideration, so that a complicated phenomenon can be described. Since an appropriate model can be automatically selected in response to the change, it is possible to maintain the prediction accuracy even in the case where the attribute of the sudden change or the state is unclear. In particular, in the case of air conditioning load prediction, it is effective for a situation such as a rapid climate change or a change of season. As a result, when performing operations and controls based on the predicted values, accurate scheduling, parameter setting, and the like can be performed. That is, in the air-conditioning load prediction, the start and stop (scheduling) of a large number of heat supply devices such as refrigerators can be performed at accurate timing, and appropriate parameters can be set for the air-conditioning controller of the building. it can. Further, since the credibility of the predicted value can be displayed together with the accurate prediction, it is possible to prevent the erroneous use of the predicted value, and therefore, a great effect is obtained in a dangerous control field. As a result, the operator can perform more accurate operations with a sense of security and margin, and can reduce the burden on the operator.

[0060]

As described above, according to the present invention,
When predicting the state, it is possible to describe complex phenomena by being able to consider many factors of the state several hours after the input factor, and the appropriate model according to the situation is automatically selected from the classified models Therefore, for example, even when the load suddenly changes or the attribute of the state is unclear, the prediction accuracy can be maintained. Further, since the credibility of the predicted value can be calculated together with the accurate predicted value, it is possible to prevent the erroneous use of the predicted value, and an excellent effect is obtained in a dangerous control field. Also, when predicting the heat load of the air conditioning system, the air conditioning load after several hours is taken into consideration simultaneously as input factors, such as the outside air temperature, the change amount of the outside air temperature, the discomfort index, the heat load, and the change amount of the heat load. Since a model to be predicted is generated and inference is performed based on the model, it is possible to accurately grasp the phenomenon, and as a result, it is possible to perform accurate load prediction.

[Brief description of the drawings]

FIG. 1 is a functional block diagram showing one embodiment of a system to which a state prediction device according to the present invention is applied.

FIG. 2 is a functional block diagram of the device.

FIG. 3 is a diagram showing a state of a monthly range or an annual range with respect to each input / output variable in the device.

FIG. 4 is a diagram showing an example in which the annual range of each input / output variable is subdivided in the above device.

FIG. 5 is a diagram showing a distribution situation of input / output variables for each month in the above-mentioned device.

FIG. 6 is a diagram showing an example in which the distribution of input / output variables for each month is normalized in the above-described apparatus.

FIG. 7 is a diagram showing a distribution status of a membership function for each month in the device.

FIG. 8 is a diagram showing a barycentric value of a membership function for each month in the above device.

FIG. 9 is a graph showing the status of input / output variables used in the device.

FIG. 10 is a diagram showing a state of a prediction result in the device.

FIG. 11 is a functional block diagram of an apparatus showing another embodiment of the present invention.

FIG. 12 is a graph showing the status of input / output variables used in another embodiment device.

[Explanation of symbols]

 2 Data Classification Unit 11 Recognition Model Generation Means 12 Recognition Model Learning Means 15 Attribution Recognition Means 16 Final Predicted Value Determination Means 21 Causal Relation Model Generation Means 22 Causal Relation Model Learning Means 25 Inference Means 26 Credibility Calculation Means 30 Load Prediction Value Display Unit 50 credibility display unit 100 input variable input unit 101 delay time determination unit 102 input status symbolization unit 103 input / output relation modeling unit A state prediction model generation unit B model learning unit C status prediction unit D result display unit

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G06F 17/00 F24F 11/02 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A state prediction device for predicting a state occurring several hours later due to an input factor based on a relation with the input factor, comprising: means for classifying case data indicating the input factor according to a predetermined standard; Means for inputting case data of each type classified according to, extracting factors for recognition and generating a recognition model based on each extracted factor, and inputting case data of each type classified according to a predetermined criterion. A means for generating a causal relationship model between the input factor and the state several hours later, a means for calculating the attribution of a new situation to each class using a recognition model when a new situation occurs, Means for calculating a predicted value in a new situation using a causal relationship model when a certain situation occurs, means for calculating a final inferred value by weighted averaging of the attribution of each class and the predicted value, , Means for calculating the credibility of the value, means for learning the new situation and the state that has actually occurred several hours later based on the recognition model, and means for learning the new situation and the state that has actually occurred several hours later. Means for learning based on a causal relation model. 2. A state estimating device for estimating a state occurring several hours later due to an input factor based on a relationship with the input factor. Using the data on the change in outside temperature, the discomfort index, the heat load and the change in heat load, and using the data on the heat load occurring several hours later as the output factor due to this input factor, predicting the heat load several hours later An input factor determining means for determining how many hours before the input factor should be used for the input factor, and an input factor symbolizing means for converting the input factor determined by the input factor determining means into symbolized data, Modeling means for generating a model by associating the input factor with the output factor based on the symbolization data; and performing inference using the model when a new input factor is input. State predicting apparatus characterized by comprising a Cormorant inference means.