JP5546981B2 - Output value prediction method, apparatus, and program for the method - Google Patents

Description

  The present invention relates to an output value prediction technique for predicting an output value from given data, and more particularly to an output value prediction technique that can use data having no data value for all data items.

  In various fields, future predictions are often made when determining future actions. In particular, in products manufactured through various manufacturing processes in a relatively large-scale manufacturing plant, such as manufacturing of steel products and chemical products, for example, depending on input amount, operation input amount, time passage, etc. In many cases, the output value in each manufacturing process and the output value of the final process directly connected to the product change every moment. For this reason, in order to control the output value, the prediction of the output value is important. For example, a method of performing such prediction by using past performance data has been devised, and for example, proposed in Patent Document 1.

  In the dephosphorization control method disclosed in Patent Document 1, the phosphorus concentration at the end of the blowing of each charge in which the iron and various auxiliary materials are blown into the converter or the pot is put into the target range. A dephosphorization control method for storing a new blowing condition in a newly implemented charge is defined as a new blowing vector, and a plurality of blowing conditions in each charge performed in the past are defined as a plurality of actual blowing vectors. Defining an actual blown vector that is similar to the new blown vector and whose end point phosphorus concentration is within a reference range from the actual blown vectors, and each of the selected actual blown vectors An approximate model for estimating the dephosphorization auxiliary material input amount of the newly implemented charge is created from the blowing conditions and the actual dephosphorization auxiliary material input amount, and the charge of the newly implemented charge is calculated using the prepared approximate model. Prolapse It estimates the auxiliary raw material input, based on the estimated dephosphorization auxiliary raw material input, is intended to determine the dosage of dephosphorization adjuncts charge of implementing the new.

  In the dephosphorization control method disclosed in Patent Document 1, the approximate model is selected from among the actual blown vectors that are actual blown vectors that are similar to the new blown vector and have an end point phosphorus concentration within a reference range. It is selected and created from each blowing condition and actual dephosphorization auxiliary raw material input amount of the selected actual blowing vector. Therefore, if there are not enough actual blown vectors similar to the new blown vector in the past actual blown vectors, the accuracy of the approximate model is lowered and sufficient prediction accuracy cannot be obtained. In particular, in the case of mass production of small varieties, a lot of similar data can be selected and used from a large amount of past performance data. In some cases, this situation is likely to occur. In addition, there are circumstances in which all data items are not always measured for cost reduction and operational efficiency improvement, and such circumstances are likely to occur.

  Therefore, in order to increase the past performance data that can be used, it is conceivable to use performance data including data items with missing data values. For this purpose, it is necessary to supplement the data values in the performance data. A technique for complementing such data values is proposed in Patent Document 2, for example.

  This missing data complementing method disclosed in Patent Document 2 detects missing data in which some features are missing from the sample data group accumulated in the database, and the missing data and normal data without any missing data. Deficiency detection division processing step of dividing the data into the normal data similar to the deficient data using a predetermined similarity scale, and data of features corresponding to the deficient features of the deficient data in the obtained normal data Complement data includes a complement processing step of substituting and complementing the missing data in the missing data. More specifically, normal data that minimizes the Manhattan distance with respect to the missing data is obtained, and a feature value of the feature corresponding to the missing feature of the missing data in the obtained normal data is obtained. The feature value of normal data is set as a candidate for completion and substituted for the missing value of missing data.

JP 2008-7828 A JP 2002-215646 A

  By the way, when the missing data complementing method disclosed in Patent Document 2 is used for a method of predicting an output value by using past performance data, the missing data is missing with one normal data that minimizes the Manhattan distance. Missing values in the data will be complemented. The degree of similarity between normal data and missing data only needs to match the degree of similarity between the feature value and missing value of normal data, but this is not always the case and the missing value may not be properly complemented, and the prediction accuracy is It may be lower.

  The present invention has been made in view of the above-described circumstances, and its object is to obtain a predicted value of output with higher prediction accuracy when using past performance data including data items with missing data values. The present invention provides an output value prediction method, an output value prediction device, and an output value prediction method program.

As a result of various studies, the present inventor has found that the above object is achieved by the present invention described below. That is, an output value prediction method according to an aspect of the present invention is an output value prediction method used in an output value prediction device, and the output value prediction device includes a plurality of past performance data including a plurality of data items. A plurality of data items that have data values in all of the plurality of data items in the past performance data. Are copied to the number corresponding to the number of past measured data of all items, and for each of the copied partial item unmeasured past actual data of the partial item unmeasured past actual data, the unmeasured data item a data value, a complementary step to compensate the data values of the data items corresponding to the unmeasured data item in the plurality of all items measured historical performance data, said output values predicting device, before For each of the plurality of all item measurement past performance data and each of the plurality of copy partial item unmeasured past performance data each supplemented with the data value in the complementing step, a weight calculation step for obtaining a weight, and the output value prediction device, wherein a plurality of weights calculated in the weight calculating step, wherein a plurality of all items measured historical performance data, based on the data value to a plurality of copying and some items unmeasured historical performance data supplemented respectively the complementary step, An output value prediction step that predicts an output value for the prediction target data to be predicted and obtains a prediction value, and the weight calculation step sets each weight of the plurality of all item measurement past performance data as a predetermined constant, Each weight of the plurality of copy partial item unmeasured past performance data, the predetermined constant is the number of the plurality of all item measurement past performance data In that said values and to Rukoto of division. An output value prediction apparatus according to another aspect of the present invention includes a part including an unmeasured data item having no data value among the plurality of data items in a plurality of past performance data including a plurality of data items. Item unmeasured past performance data is copied to a number corresponding to the number of all item measured past performance data having data values for all of the plurality of data items in the past performance data, For each copied partial item unmeasured past performance data of the measured past performance data, the data value of the unmeasured data item corresponds to the unmeasured data item in the plurality of all item measured past performance data a complementary part to supplement the data value of the data item, the plurality of copying some items measuring non supplemented each data value in the plurality of all items measured historical performance data and the complementary steps For each of the past performance data, and the weight calculator for determining a weight, a plurality of weights calculated by the weight calculator, a plurality of all items measured historical performance data, a plurality supplemented with data values respectively the complementary portion based in part for replication and item unmeasured historical performance data, and an output value prediction unit for obtaining the prediction value by predicting an output value for the predicted target data to be predicted, the weight calculation section, the plurality of all the items Each weight of the measured past performance data is set as a predetermined constant, and each weight of the plurality of copy partial item unmeasured past performance data is divided by the number of the plurality of all items measured past performance data. and wherein the value and be Rukoto. Then, another output value prediction program according to an aspect of the present invention, a computer having a storage unit for storing a plurality of historical performance data and the prediction target data comprising a plurality of data items, in order to predict the output value, said read plurality of the plurality comprising data item part item unmeasured historical performance data including the unmeasured data items that do not have a data value in said plurality of data items in the past record data from the storage unit, the Each of the plurality of data items in the past performance data is copied to the number corresponding to the number of all the item measurement past performance data having data values, and each of the copied items of the partial item unmeasured past performance data is copied. For each copy item unmeasured past performance data, the data value of the unmeasured data item is changed to the unmeasured data in the plurality of all item measurement past performance data. And complementary means for supplementing the data value of the data item corresponding to the data item, the plurality of all items measured historical performance data and the complementary means respectively the data value of the plurality of copying some items unmeasured historical performance data supplemented respectively A weight calculation means for obtaining a weight, the plurality of weights obtained by the weight calculation means, the plurality of all-item measurement past performance data, and a plurality of copy partial items each supplemented with a data value in the complementing step based on the unmeasured historical performance data, an output value prediction program to function as an output value predicting means for obtaining a predicted value by predicting an output value for the predicted target data to be predicted, the weight calculation means, Each weight of the plurality of all-item measured past performance data is set as a predetermined constant, and each of the plurality of copy partial item unmeasured past performance data And wherein the values and to Rukoto obtained by dividing said predetermined constant in the plurality of all items measuring the number of historical performance data.

  In the output value prediction method, the output value prediction apparatus, and the output value prediction program having such a configuration, the data value of the unmeasured data item includes data items corresponding to all the unmeasured data items of the all-item measurement past performance data. Data values are used. For this reason, since the number of usable past performance data further increases, the output value prediction method, the output value prediction device, and the output value prediction program having such a configuration include past performance data including data items with missing data values. When using, the predicted value of the output value can be obtained with higher prediction accuracy.

Further, in another aspect, in the output value prediction method described above, the output value prediction step, the plurality of weights the output value prediction model representing a relationship between the output variables and input variables obtained by the weight calculation step A model calculation step obtained by a weighted multiple regression method using, and a prediction value calculation step for obtaining the prediction value by substituting the prediction target data into the output value prediction model obtained in the model calculation step. characterized in that it obtain.

  In the output value prediction method having such a configuration, each weight of the plurality of all-item measured past performance data is a predetermined constant, and each weight of the plurality of copy partial item unmeasured past performance data is the predetermined weight. Is a value obtained by dividing the constant by the number of all past item measurement data of the plurality of items. In this way, the plurality of all-item measured past performance data that do not need to be supplemented with data values and the plurality of copy partial item unmeasured past performance data supplemented with data values are differentiated. This output value prediction method can obtain a predicted value of an output value with higher prediction accuracy when using past performance data including a data item with a missing data value.

Further, in another aspect, in the output value prediction method described above, the output value prediction step, said plurality of copying a part of the data value supplemented respectively in the previous SL multiple full items measured historical performance data and the complementary steps For each of the item unmeasured past performance data, a similarity calculation step for calculating a similarity indicating a difference from the prediction target data, the plurality of weights determined in the weight calculation step, and the similarity calculated in the similarity calculation step Each of the plurality of similarities, and the data items corresponding to the predicted values in the plurality of copy partial item unmeasured past performance data respectively supplemented with the plurality of all item measured past performance data and the data value in the complementing step. based on the data value, and wherein the obtaining Bei the predicted value calculation step of obtaining the predicted value. Preferably, the predicted value calculation step includes the plurality of all-item measurement past performance data and the plurality of weights obtained in the weight calculation step and the plurality of similarities obtained in the similarity calculation step. The predicted value is obtained by weighting each data value of the data item corresponding to the predicted value in the plurality of copy partial item unmeasured past performance data respectively supplemented with the data value in the complementing step.

  In the output value prediction method having such a configuration, each weight of the plurality of all-item measured past performance data is a predetermined constant, and each weight of the plurality of copy partial item unmeasured past performance data is the predetermined weight. Is a value obtained by dividing the constant by the number of all past item measurement data of the plurality of items. In this way, the plurality of all-item measured past performance data that do not need to be supplemented with data values and the plurality of copy partial item unmeasured past performance data supplemented with data values are differentiated. This output value prediction method can obtain a predicted value of an output value with higher prediction accuracy when using past performance data including a data item with a missing data value. In the output value prediction method having such a configuration, for each of the plurality of all item measurement past performance data and each of the plurality of copy partial item unmeasured past performance data each supplemented with a data value in the complementing step, Similarity with prediction target data is obtained, the plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the plurality of all-item measurement past performance data The predicted value is obtained based on each data value of the data item corresponding to the predicted value in the plurality of copy partial item unmeasured past performance data each supplemented with the data value in the complementing step. Since the similarity is used when the predicted value is obtained in this way, the output value predicting method having such a configuration is further improved when using past performance data including data items with missing data values. The predicted value of the output value can be obtained with high prediction accuracy.

According to another aspect, in the output value prediction method described above, the output value prediction step includes the plurality of all-item measurement past performance data and the plurality of copy partial items each supplemented with a data value in the complementing step. For each unmeasured past performance data, obtain the model predicted value by substituting the data value of the data item corresponding to the input variable into the given output value predictive model that represents the relationship between the input variable and the output variable. the prediction error calculating step of obtaining a difference between the model predicted value and the data value of the data item corresponding to the output variable as a prediction error, before Symbol plurality of all items measured historical performance data and the data values in the complementary step, respectively for each of the plurality of copying some items unmeasured historical performance data supplemented obtains a similarity that represents the difference between the predicted target data Based on the similarity calculation step, the plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the plurality of prediction errors obtained in the prediction error calculation step A prediction value prediction error calculation step for obtaining a prediction error in a prediction value of the prediction target data; a value obtained by substituting the prediction target data into the given output value prediction model; and a prediction value prediction error calculation to obtain Bei the predictive value calculating step by obtaining the sum of the prediction error in the prediction value of the prediction target data obtained in step obtains the predicted value and said. Preferably, in the prediction value prediction error calculation step, the plurality of weights obtained in the weight calculation step and the plurality of similarities obtained in the similarity calculation step are obtained in the prediction error calculation step. A prediction error in the prediction value of the prediction target data is obtained by weighting a plurality of prediction errors.

  In the output value prediction method having such a configuration, each weight of the plurality of all-item measured past performance data is a predetermined constant, and each weight of the plurality of copy partial item unmeasured past performance data is the predetermined weight. Is a value obtained by dividing the constant by the number of all past item measurement data of the plurality of items. In this way, the plurality of all-item measured past performance data that do not need to be supplemented with data values and the plurality of copy partial item unmeasured past performance data supplemented with data values are differentiated. This output value prediction method can obtain a predicted value of an output value with higher prediction accuracy when using past performance data including a data item with a missing data value. In the output value prediction method having such a configuration, for each of the plurality of all item measurement past performance data and each of the plurality of copy partial item unmeasured past performance data each supplemented with a data value in the complementing step, Similarity with the prediction target data is obtained, and the plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the prediction error calculation step Based on the plurality of prediction errors, a prediction error in a prediction value of the prediction target data is obtained, and the prediction value is obtained. Since the similarity is used when the predicted value is obtained in this way, the output value predicting method having such a configuration is further improved when using past performance data including data items with missing data values. The predicted value of the output value can be obtained with high prediction accuracy.

According to another aspect, in the output value prediction method described above, the output value prediction step includes the plurality of all-item measurement past performance data and the plurality of copy partial items each supplemented with a data value in the complementing step. For each unmeasured past performance data, obtain the model predicted value by substituting the data value of the data item corresponding to the input variable into the given output value predictive model that represents the relationship between the input variable and the output variable. the prediction error calculating step of obtaining a difference between the model predicted value and the data value of the data item corresponding to the output variable as a prediction error, before Symbol plurality of all items measured historical performance data and the data values in the complementary step, respectively for each of the plurality of copying some items unmeasured historical performance data supplemented obtains a similarity that represents the difference between the predicted target data A similarity score calculation step, for each of the plurality of copying some items unmeasured historical performance data of the data value supplemented respectively the complementary step, the influence degree calculation step of obtaining the degree of influence representing the weight of the unmeasured data items The plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, the plurality of influences obtained in the influence degree calculation step, and obtained in the prediction error calculation step A prediction value prediction error calculating step for obtaining a prediction error in a prediction value of the prediction target data based on the plurality of prediction errors, and substituting the prediction target data into the given output value prediction model. And the prediction error in the prediction value of the prediction target data obtained in the prediction value prediction error calculation step. In wherein the obtaining Bei the predicted value calculation step of obtaining the predicted value.

  In the output value prediction method having such a configuration, since the degree of influence is used, the degree of influence of the complemented unmeasured data item on the predicted value is considered. For this reason, the output value prediction method having such a configuration can obtain a predicted value of an output value with higher prediction accuracy when using past performance data including a data item in which a data value is missing.

According to another aspect, in the above-described output value prediction methods, the output value prediction step includes the plurality of all-item measurement past performance data and the plurality of copies partially supplemented with data values in the complementing step, respectively. By substituting the data value of the data item corresponding to the input variable for a given output value prediction model that represents the relationship between the input variable and the output variable for each of the item unmeasured past performance data, the model predicted value is determined, the prediction error calculation step of obtaining a difference between the data values of the data items corresponding to the output variable and the model predictive value as the prediction error, the previous SL data values in a plurality of all items measured historical performance data and the complementary steps for each of said plurality of copying some items unmeasured historical performance data supplemented respectively, the similarity representing the difference between the predicted target data The plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the prediction errors obtained in the prediction error calculation step. Based on a prediction value prediction error calculation step for obtaining a prediction error in a prediction value of the prediction target data, and a value obtained by substituting the prediction target data into the given output value prediction model and the prediction value prediction A prediction value calculation step of obtaining the prediction value by obtaining a sum of a prediction error in the prediction value of the prediction target data obtained in the error calculation step, and the weight calculation step includes the plurality of all item measurement past Each weight of the performance data is set to a predetermined constant, and each weight of the plurality of copy partial item unmeasured past performance data is set to the unmeasured data. If the predetermined condition for determining that the value should be compensated for by the eyes is not satisfied, 0 is set. If the predetermined condition is satisfied, the predetermined constant is set to the copy number satisfying the predetermined condition. It is a value obtained by dividing the number of part item unmeasured past performance data.

  In the output value prediction method having such a configuration, the copy part item unmeasured past actual result data that does not satisfy the predetermined condition is removed from the calculation of the predicted value, so that the unmeasured data is obtained by the data value that cannot be taken as the data value. The possibility that the data value of the item is supplemented is reduced. For this reason, the output value prediction method having such a configuration can obtain the predicted value of the output value with higher prediction accuracy.

  Moreover, in another one aspect | mode, in these above-mentioned output value prediction methods, several past performance data provided with several data items are divided into the said partial item unmeasured past performance data and the said all item measurement past performance data. It further comprises a separation step.

  With the output value prediction method having such a configuration, even when all past measured data and some past unmeasured past performance data are mixed in past performance data, these can be separated, and a supplementary step is performed. Can be implemented properly.

  An output value prediction method, an output value prediction apparatus, and an output value prediction program according to the present invention obtain a predicted value of an output value with higher prediction accuracy when using past performance data including a data item having a missing data value. be able to.

It is a block diagram which shows the structure of the output value prediction apparatus in 1st Embodiment. It is a flowchart which shows operation | movement of the output value prediction apparatus in 1st Embodiment. It is a flowchart which shows the complementation operation | movement of the output value prediction apparatus in 1st Embodiment. It is a figure for demonstrating the prediction object data in the output value prediction apparatus of 1st Embodiment, and the past performance data memorize | stored in the past performance data database part. It is a figure for demonstrating the all item measurement past performance data memorize | stored in the all item measurement data database in the output value prediction apparatus of 1st Embodiment, and the one part unmeasured past performance data memorize | stored in a missing data database. . It is a block diagram which shows the structure of the output value prediction apparatus in 2nd Embodiment. It is a flowchart which shows operation | movement of the output value prediction apparatus in 2nd Embodiment. It is a figure for demonstrating the prediction object data in the output value prediction apparatus of 2nd Embodiment, and the past performance data memorize | stored in the past performance data database part. It is a figure for demonstrating the all item measurement past performance data memorize | stored in the all item measurement data database in the output value prediction apparatus of 2nd Embodiment, and the one part unmeasured past performance data memorize | stored in a missing data database. . It is a block diagram which shows the structure of the output value prediction apparatus in 3rd Embodiment. It is a flowchart which shows operation | movement of the output value prediction apparatus in 3rd Embodiment. It is a figure for demonstrating the past performance data memorize | stored in the past performance data database part in the output value prediction apparatus of 3rd Embodiment. It is a figure for demonstrating all the item measurement past performance data memorize | stored in the all item measurement data database in the output value prediction apparatus of 3rd Embodiment, and the one part unmeasured past performance data memorize | stored in a missing data database. . It is a figure for demonstrating the complementation operation | movement of the output value prediction apparatus in 3rd Embodiment. It is a block diagram which shows the structure of the output value prediction apparatus in 4th Embodiment. It is a block diagram which shows the structure of the output value prediction apparatus in 5th Embodiment.

  Hereinafter, an embodiment according to the present invention will be described with reference to the drawings. In addition, the structure which attached | subjected the same code | symbol in each figure shows that it is the same structure, The description is abbreviate | omitted suitably. Further, in this specification, when referring generically, it is indicated by a reference symbol without a suffix, and when referring to an individual configuration, it is indicated by a reference symbol with a suffix.

(First embodiment)
FIG. 1 is a block diagram illustrating a configuration of an output value prediction apparatus according to the first embodiment. The output value prediction apparatus Sa in the present embodiment predicts the output Y _p for the prediction target data X _0j to be predicted based on the past performance data (X _j , Y), thereby predicting the predicted value y of the prediction target data X _{0j. This} is a device for determining _p0 . More specifically, the output value prediction device Sa sets a plurality of data values of the unmeasured data items in the partial unmeasured past performance data (X _j , Y) including unmeasured data items that do not have data values. All item measured past performance data (X _j , Y) is supplemented with the data value of the data item corresponding to the unmeasured data item, and the plurality of all item measured past performance data (X _j , Y) and the unmeasured data item The predicted value y _p0 is obtained by predicting the output Y _p for the prediction target data to be predicted based on a plurality of copy partial item unmeasured past performance data (X _j , Y) each supplementing the data value of .

The past performance data (X _j , Y) includes a plurality of data items, and the data values of these data items are values acquired in the past. In numerical experiments, a value prepared in advance by the user may be used. The plurality of data items are composed of a predetermined output Y and a predetermined factor X _j that can be numerically related to the output Y. The all item measurement past result data is past result data (X _j , Y) having data values for all of the plurality of data items. Partial item unmeasured past performance data (X _j , Y) is past performance data (X _j , Y) including at least an unmeasured data item having no data value among the plurality of data items.

  For example, as shown in FIG. 1, the output value prediction apparatus Sa includes a complementing unit 2a, a model calculation unit 3a, and a prediction value calculation unit 4a. In addition, in the present embodiment, the separation unit 1, An input unit 5 and an output unit 6 are provided.

The input unit 5 is, for example, various commands such as a command to start output value prediction program for predicting output Y _p by a method of this embodiment from the previously given data, and, for example, historical performance data (X _{j, Y)} And a device that inputs various data necessary for predicting the output Y _p such as the prediction target data X _0j to the output value prediction device Sa, such as a keyboard and a mouse. The output unit 6 is a device that outputs a command or data input from the input unit 5 and an output value (predicted value) _yp0 predicted by the output value prediction device Sa. For example, a CRT display, LCD, organic A display device such as an EL display and a plasma display, and a printing device such as a printer.

Separation unit 1, historical performance data _(X j, Y) stores store past actual data _(X j, Y) of all items measured historical performance data _(X j, Y) and some items unmeasured It is divided into past performance data (X _j , Y), and each is stored and stored. The separating unit 1 stores, for example, a past performance data database unit (past performance data DB unit) 11 that stores and stores past performance data (X _j , Y), and a past performance data stored in the past performance data DB unit 11. Data separation unit 12 that divides data (X _j , Y) into all-item measured past performance data (X _j , Y) and some item unmeasured past performance data (X _j , Y), and data separation unit 12 All item measured past performance data (data missing portion database portion (missing data DB portion) 13) that stores and stores the partial item unmeasured past performance data (X _j , Y) and the data separator 12 ( X _j , Y), and an all-item measurement data database unit (all-item measurement data DB unit) 14 that stores and stores them.

The complement unit 2a uses the data values of the unmeasured data items based on the all-item measured past performance data (X _j , Y) in order to use the unmeasured past performance data (X _j , Y) for some items for output value prediction. To compensate. Complementing unit 2a, some items unmeasured historical performance data (X _{j, Y),} and copied to the number corresponding to the number of the plurality of all the items past performance data (X _{j, Y),} a plurality of the these copied one For each part item unmeasured past performance data (copy partial item unmeasured past performance data) (X _j , Y), the data value of the unmeasured data item is set to the plurality of all item measured past performance data (X _j , Y), the data complementing unit 21 that supplements the data value of the data item corresponding to the unmeasured data item, the plurality of all-item measured past performance data (X _j , Y), and the plurality of copied partial items not yet A data weight calculation unit 22 for obtaining a weight (data weight t) for each piece of measured past performance data (X _j , Y) is provided. In this embodiment, the data weight calculation unit 22 sets each data weight t of the plurality of all-item measurement past performance data (X _j , Y) to a predetermined constant (for example, 1) and the plurality of copy ones. A value obtained by dividing each data weight t of the part item unmeasured past performance data (X _j , Y) by the number of the plurality of all item measurement past performance data (in the above example, 1 / (all item measurement) The number of past performance data)).

Model calculation portion 3a, historical performance data (X _{j, Y)} stored in the separation unit 1 based on, and requests necessary to obtain the prediction value y _p0 of the prediction target data X _0j. In this embodiment, the model calculation unit 3a obtains an output value prediction model y = f (a _j , X _j ) representing the relationship between the input variable X _j and the output variable Y by the weighted multiple regression method, For example, a weight calculation unit 31 for obtaining a weight (multiple regression weight) W in the weighted multiple regression method based on a plurality of data weights t obtained by the data weight calculation unit 22 of the complement unit 2a, and an output value prediction model y = f And a parameter calculation unit 32 for obtaining a coefficient parameter a _j of (a _j , X _j ).

Prediction value calculation unit 4a, the output value prediction model identified by the model calculation unit 3a (a _j, X _j) to determine the predicted value _{y p0} of the prediction target data _{X 0j} by substituting the predicted target data _{X 0j} Is.

  The output value prediction apparatus Sa having such a configuration can be configured by, for example, a computer, more specifically, a server computer, a personal computer such as a notebook computer or a desktop computer, or the like. In the output value prediction device Sa, the separation unit 1, the complementing unit 2a, the model calculation unit 3a, the prediction value calculation unit 4a, the input unit 5 and the output unit 6 may be configured integrally, or may be configured separately. May be. For example, the separation unit 1 includes a hard disk or the like, and the complementing unit 2a, the model calculation unit 3a, and the predicted value calculation unit 4a include, for example, a microprocessor such as a CPU, a RAM (Random Access Memory), and a ROM (Read Only Memory). ) And the like, and these are functionally configured, whereby the output value predicting device Sa is integrally configured by one computer. Further, for example, the separation unit 1 is configured to include a database server computer, and the complementing unit 2a, the model calculation unit 3a, and the predicted value calculation unit 4a are configured to include a computer, whereby the output value prediction device Sa is 2 Consists of separate computers.

  Note that the output value prediction device Sa may further include an external storage unit as necessary. The external storage unit is, for example, a flexible disk, a CD-ROM (Compact Disc Read Only Memory), a CD-R (Compact Disc Recordable), a DVD-R (Digital Versatile Disc Recordable), and a Blu-ray Disc (registered trademark). For example, a flexible disk drive, a CD-ROM drive, a CD-R drive, a DVD-R drive, a Blu-ray disk drive, and the like are devices that read data from and / or write data to / from a recording medium.

Here, when the output value prediction program or the like is not stored, the output value prediction device Sa is installed so as to install the output value prediction program from the recording medium on which the output value prediction program or the like is recorded via the external storage unit. May be configured. In addition, the output value predicting device Sa is configured to input the data via the external storage unit from a recording medium that records data such as past performance data (X _j , Y) and data for predicting an output value. It may be configured.

  Next, the operation of the output value prediction apparatus Sa will be described. FIG. 2 is a flowchart showing the operation of the output value prediction apparatus in the first embodiment. FIG. 3 is a flowchart showing the complementing operation of the output value prediction apparatus according to the first embodiment. FIG. 4 is a diagram for explaining prediction target data and past performance data stored in the past performance data database unit in the output value prediction apparatus of the first embodiment. FIG. 5 is a diagram for explaining all item measurement past performance data stored in the all item measurement data database and some item unmeasured past performance data stored in the missing data database in the output value prediction apparatus of the first embodiment. FIG.

For example, when an output command is received from the input unit 5 by a user operation, the output value prediction device Sa executes an output value prediction program. By executing the output prediction program, the complementing unit 2a, the model calculation unit 3a, and the predicted value calculation unit 4a are functionally configured in the computer. Then, the output value prediction device Sa predicts the output value (predicted value) y _p0 from the prediction target data X _p0 based on the past performance data (X _j , Y) by the following operation.

In the prediction of the output Y _p, in the past record data DB 11 in the separation unit 1 of the output value prediction unit Sa, for example, historical performance data (X in tabular form (table format) as shown in FIG. 4 (A) _j , Y) is stored in advance, and the prediction target data _Xp0 is stored as shown in FIG. 4B by being input from the input unit 5. Here, the predetermined output Y is a set of N elements (output elements) y _j , that is, Y = {y _j | j = 1, 2,... The factor X related to the output Y is a set of N elements (factor elements) x _ji , that is, X = {x _ji | j = 1, 2,..., N}.

The past performance data table 1000 shown in FIG. 4A includes an output field (first data item) 101 for registering the actually measured output value y _j and the factor X related to the predetermined output Y. Each field of the factor data field 101 for registering the data x _ji is provided, and a record is provided for each past record data (X _j , Y). That is, the past performance data table 1000 has N records. Note that 0 is attached to the prediction target data _Xp0 as a subscript to distinguish it from the past performance data ( _Xj , Y). In addition, past performance data (X _j , Y) are assigned 1 to N as first subscripts in order to distinguish N pieces of data in an identifiable manner. Then, the first to nth data items that are n elements in the factor X related to the predetermined output Y are distinguished from the prediction target data _Xp0 and the past performance data ( _Xj , Y) in an identifiable manner. Therefore, 1 to n are added as second subscripts (the right side of the subscripts). The prediction target data X _p0 is X _p0 = [x ₀₁ , x ₀₂ ,..., X _0N ], and the past performance data (X _j , Y) is X _ji = [x _j1 , x _{j2 ,.} X _jn ]. For example, y ₃ represents the third output value in the past performance data (X _j , Y), and for example, x ₂₃ represents the second third value in the past performance data (X _j , Y). it represents the value of the data item, also for example, x ₀₄ represents the value of the fourth data item in the prediction target data X _p0.

As described above, the factor X related to the predetermined output Y is configured to include a plurality of n elements (data items, factor elements). Therefore, the factor data field 102 corresponds to the number n of elements. A factor data item subfield is provided. In the example shown in FIG. 4, the predetermined output Y involves at least n elements (first to nth data items). Therefore, the factor data item subfield is a factor data item subfield (second to n + 1) for registering the first to nth data x _{j1 to} x _jn respectively corresponding to each element x _ji of the factor X. (Th data item) 102-1 to 102-n. The past performance data (X _j , Y) is data obtained by actual measurement or the like at different times (time points) t _j in the past under different conditions in the past. In the example shown in FIG. It consists of data. It predicted target data X _p0 is data of the object to be predicted, for example, and the measured data values x _0i until prediction time t _0, the operation value x _0i and the input, and operation date and time x _0i, for simulation The prepared data value x _0i or the like.

In the present embodiment, the past performance data table 1000 includes past performance data (partially unmeasured past performance data) (X _j , Y) including unmeasured data items having no data value in the data items. Is present. In the example shown in FIG. 4A, the first past performance data (X _j , Y) is unmeasured past performance data for some items, and the first factor data item subfield (second data item) ) 102-1 is missing its data value and does not have that data value.

In the case where such past performance data (X _j , Y) is stored in the past performance data DB unit 11 of the separation unit 1, the output value prediction device Sa firstly separates the separation unit 1 as shown in FIG. by the data separation section 12, historical performance data _(X j, Y), and divided into all items measured historical performance data _(X j, Y) and some items unmeasured historical performance data _(X j, Y) and, respectively, All item measurement data DB unit 14 and missing data DB unit 13 are stored (S11).

For example, the data separation unit 12 of the separation unit 1 outputs the output field 101 and the first to nth factor data item subfields 102-1 to 102-n (first to nth) for each record in the past performance data table 1000. It is determined whether or not a value (data value) is stored in (n + 1th data item), and as a result of this determination, the output field 101 and the first to nth factor data item subfields 102-1 to 102-n When all values (data values) are stored, it is determined that the record is all-item measurement past performance data (X _j , Y). On the other hand, as a result of this determination, the output field 101 and the first to first data A value (data value) is stored in at least one of the n factor data item subfields 102-1 to 102-n. When no is the record determines that some items unmeasured historical performance data (X _{j, Y).} Then, the data separation unit 12 stores the past performance data (X _j , Y) of the record determined to be all item measurement past performance data (X _j , Y) in the all item measurement data DB unit 14, and performs all item measurement. is stored in the data DB 14 stores some items unmeasured historical performance data _(X j, Y) historical performance data of the judgment record and _(X j, Y) of the missing data DB unit 13, the missing data DB Store in the unit 13.

For example, in the example shown in FIG. 4A, the first past performance data (x _1j , y ₁ ) from which the data value of the first data item X ₁ is missing is the unmeasured past performance data for some items. 5B, some item unmeasured past performance data table 1003 stores the second to second past performance data (x _2j , y ₂ ) to (x _nj , y _n ). This is item measurement past result data, and is stored in the all item measurement past result data table 1002 shown in FIG. Similar to the past performance data table 1000 shown in FIG. 4A, this all-item measurement past performance data table 1002 is related to the output field 121 for registering the actually measured output value y _j and the predetermined output Y. Each of the factor data fields 122 (122-1 to 122-n) for registering the data x _ji of the factor X to be recorded, and further, the data weight field 123 for registering the data weight t _j is provided, and all items measurement past results A record is provided for each data (X _j , Y). Also, the part item unmeasured historical performance data table 1003, similar to the historical performance data table 1000 shown in FIG. 4 (A), the output field 131 for registering the measured output values y _j, and, of the predetermined Each field of the factor data field 132 (132-1 to 132-n) for registering the data x _ji of the factor X related to the output Y is provided, and further, the data weight field 123 for registering the data weight t _j is provided. The number of records is obtained by multiplying the number of part item unmeasured past performance data (X _j , Y) by the number of all item measurement past performance data (X _j , Y). At this stage of step S11, the data weight t is calculated in the next step S12, so that the data value is not registered.

Next, the output value prediction apparatus Sa uses the complementing unit 2a to convert the data values of the unmeasured data items in the partial unmeasured past performance data (X _j , Y) to the plurality of all-item measured past performance data (X _j , Y) are supplemented with the data values of the data items corresponding to the unmeasured data items, and the plurality of copies are respectively supplemented with the data values of the plurality of all-item measured past performance data (X _j , Y) and the unmeasured data items. The data weight t of each piece of data of some item unmeasured past performance data (X _j , Y) is calculated (S12).

More specifically, for example, as shown in FIG. 5B, the data complementing unit 21 of the complementing unit 2a converts some item unmeasured past performance data (X _j , Y) into a plurality of all item past performances. The data (X _j , Y) is copied to the number corresponding to the number of data, and each of the copied partial item unmeasured past performance data (copy partial item unmeasured past performance data) (X _j , Y) The data value of the unmeasured data item is supplemented with the data value of the data item corresponding to the unmeasured data item in the plurality of all-item measured past performance data (X _j , Y). In FIG. 5B, for each of a plurality of copied partial unmeasured past performance data (X _j , Y), the data values of the first data item X ₁ are all item measured past performance data (X _2). , Y) to (X _N , Y) are respectively supplemented with data values X _{21 to} X _N1 of the first data item. Then, the data weight calculation unit 22 of the complementing unit 2a sets the data weight t to a predetermined constant (for example, 1) for each of the plurality of all item measurement past performance data (X _j , Y), and all the item measurement past performances. Registered in the data weight field 123 of the data table 1002, and for each of the plurality of partial item unmeasured past performance data (X _j , Y), the data weight t is the predetermined constant and the plurality of all item measurement past performance data. A value obtained by dividing by the number of (X _j , Y) (in the above example, 1 / (number of all item measured past performance data = N−1)), and data weight field 133 of some item unmeasured past performance data table 1003. Register with. That is, in the example shown in FIG. 4 and FIG. 5, since the number of all item measured past performance data (X _j , Y) is N−1, some item unmeasured past performance data (X _j , Y) The data weight t _i is expressed by Equation 1.

Here, t i _(k) is copied a plurality of part items unmeasured historical performance data (X _{j, Y)} k-th part items unmeasured historical performance data (X _{j, Y)} in the data it is the weight _{t i.}

In this example, the sum of the data weights t _i (k) in the copied past unmeasured past item data is defined so as to be the predetermined constant (1 in this example). The total data weights t _i in the plurality of past unmeasured past data of some items are made equal to the respective data weights t of the all past measured past data (X _j , Y).

In this embodiment, in order to obtain a more accurate predicted value y _p0 , a relatively high-precision estimated value (as a data value of an unmeasured data item in some item unmeasured past performance data (X _j , Y)) ( If there is a comparatively highly reliable estimated value), it is supplemented with this estimated value instead of supplementing with all-item measured past performance data.

That is, step S12 operates according to the flow shown in FIG. 3 in more detail. In FIG. 3, the data complementing unit 21 of the complementing unit 2a first lacks a variable that does not have a high-precision estimated value in step S21 with respect to some item unmeasured past performance data (X _j , Y). Determine whether or not. That is, it is determined whether or not a highly accurate estimated value exists for an unmeasured data item. As a result of this determination, when a variable without a high-precision estimated value is missing (when a high-precision estimated value does not exist for an unmeasured data item) (Yes), as described above, in step S24. , Some item unmeasured past performance data (X _j , Y) is copied to the number corresponding to the number of all items past performance data (X _j , Y), and these copied items are not measured For each of the past performance data (copy partial item unmeasured past performance data) (X _j , Y), the data value of the unmeasured data item in the plurality of all item measured past performance data (X _j , Y) The data weight of the data item corresponding to the unmeasured data item is supplemented, and then, in step S25, the data weight calculation unit 22 determines that each of the plurality of copy partial item unmeasured past performance data is as described above. ,formula Calculates the data weight _t i (k) in accordance with, subsequently, step S23 is executed.

On the other hand, as a result of the determination in step S21, if there is no missing variable that does not have a highly accurate estimated value (when a highly accurate estimated value exists for an unmeasured data item) (No), the complementing unit 21 , The data value of the unmeasured data item of the partial unmeasured past performance data (X _j , Y) is supplemented with the estimated value, and the data weight calculator 22 replaces the data weight t _i with the predetermined constant (above In the example, it is 1).

  Subsequently, in step S23, it is determined whether or not the complementing process and the data weight calculation process have been completed for all unmeasured past performance data of some items. If the determination result indicates that all have been completed (Yes) ), The process ends. On the other hand, if all the processes are not completed (No), the process returns to step S21.

Returning to FIG. 2, next, the model calculation unit 3a uses the weight calculation unit 31 based on the plurality of data weights t obtained by the data weight calculation unit 22 of the complementing unit 2a to output value prediction model y = f ( The weight W of the weighted multiple regression method when a _j , X _j ) is determined by the weighted multiple regression method is calculated (step S13), and the output value prediction model y = f (a _j , A coefficient parameter a _i of X _j ) is obtained (step S14). The predicted value calculation unit 4a, the model calculation unit output value prediction model identified by 3a (a _j, X _j), the predicted value of the predicted target data _{X 0j} by substituting the predicted target data _{X 0j y p0} Is obtained (step S15).

More specifically, the output value prediction model _{y = f (a j, X} j) , for example, the input variable X = the coefficient parameter _{a j [x 1, x 2} , ···, x n] linear When expressed by a polynomial and obtained by a weighted multiple regression method, the output variable Y, the input variable X, and the data weight t are expressed by Expression 2, and the weight W is expressed by Expression 3. The coefficient parameter a _j is expressed by Equation 4. Therefore, by substituting the prediction target data X _0j into the output value prediction model y = f (a _j , X _j ) using the coefficient parameter a _j represented by this equation 4, as shown in equation 5, the prediction The value y _p0 is determined.

Then, the predicted value y _p0 of the prediction target data X _0j obtained in this manner is output to the output section 6.

As described above, in the output value prediction apparatus Sa of this embodiment, the output value prediction method executed by the output value prediction apparatus Sa, and the output value prediction program implemented in the output value prediction apparatus Sa, the data value of the unmeasured data item is set. The data values of the data items corresponding to all the unmeasured data items of the all item measurement past performance data (X _j , Y) are used. For this reason, since the number of usable past performance data (X _j , Y) further increases, the output value prediction device Sa, the output value prediction method, and the output value prediction program having such a configuration have data values. When past performance data (X _j , Y) including missing data items is used, a predicted value of an output value can be obtained with higher prediction accuracy.

In the output value prediction apparatus Sa, the output value prediction method, and the output value prediction program of the present embodiment, each data weight t of the plurality of all-item measurement past performance data (X _j , Y) is a predetermined constant. (1 in the above example), and each data weight of the plurality of copy partial item unmeasured past performance data (X _j , Y) is the number of the plurality of all item measurement past performance data. The value divided by. In this way, the plurality of all-item measured past performance data (X _j , Y) that do not need to be supplemented with data values, and the plurality of copy partial item unmeasured past performance data (X _j , Y) supplemented with data values, Are differentiated by making the values of the data weights t different from each other, the output value prediction device Sa, the output value prediction method, and the output value prediction program having such a configuration lack data values. When using past performance data (X _j , Y) including data items, the predicted value of the output value can be obtained with higher prediction accuracy.

Next, another embodiment will be described.
(Second Embodiment)
FIG. 6 is a block diagram illustrating a configuration of an output value prediction apparatus according to the second embodiment. FIG. 7 is a flowchart showing the operation of the output value prediction apparatus in the second embodiment. FIG. 8 is a diagram for explaining prediction target data and past performance data stored in the past performance data database unit in the output value prediction apparatus of the second embodiment. FIG. 9 is a diagram for explaining all item measurement past performance data stored in the all item measurement data database and partial item unmeasured past performance data stored in the missing data database in the output value prediction apparatus of the second embodiment. FIG.

The output value prediction device Sa in the first embodiment outputs an output value based on a plurality of all-item measured past performance data (X _j , Y) and a plurality of copy partial item unmeasured past performance data (X _j , Y). predictive model _{y = f (a j, X} j) identified, the identified output value prediction model _{y = f (a j, X} j) prediction target data _{X 0j} by substituting the predicted target data _{X 0j} in of was calculated predicted value y _p0, the output value prediction apparatus Sb of the second embodiment, a plurality of all items measured historical performance data (X _{j, Y)} and a plurality of copying some items unmeasured historical performance data (X _j , Y) with each data weight t, a plurality of all-item measured past performance data (X _j , Y), and a plurality of copy partial item unmeasured past performance data (X _j , Y) with prediction target data X _0j Each similarity and multiple items overmeasured Each data value (data value of output variable Y) of the data item corresponding to the predicted value _yp0 in the past record data ( _Xj , Y) and a plurality of copy partial item unmeasured past record data ( _Xj , Y) based on, and requests the predicted value _{y p0} of the prediction target data _{X 0j.} The output value prediction device Sb in the second embodiment is so-called just-in-time modeling.

  Such an output value prediction device Sb of the second embodiment is configured to include, for example, a complementing unit 2a, a model calculation unit 3b, and a predicted value calculation unit 4b, as shown in FIG. , A separation unit 1, an input unit 5, and an output unit 6.

  The separation unit 1, the complement unit 2a, the input unit 5 and the output unit 6 in the output value prediction device Sb of the second embodiment are respectively the separation unit 1, the complement unit 2a, and the output value prediction device Sa of the first embodiment. Since it is the same as that of the input part 5 and the output part 6, the description is abbreviate | omitted.

Model calculation portion 3b, historical performance data (X _{j, Y)} stored in the separation unit 1 based on, and requests necessary to obtain the prediction value y _p0 of the prediction target data X _0j. In this embodiment, the model calculation unit 3b predicts each of the plurality of all-item measured past performance data (X _j , Y) and each of the plurality of copy partial item unmeasured past performance data (X _j , Y). A distance calculation unit 33 for _obtaining a distance d _i to the target data X _0j , the plurality of all-item measured past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y ), A similarity calculation unit 34 that calculates the similarity w _i with the prediction target data X _0j based on the distance d _i obtained by the distance calculation unit 33.

The predicted value calculation unit 4a uses the data of the complementing unit 2a for each of the plurality of all-item measured past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y). The data weight t obtained by the weight calculation unit 22, the similarity w _i obtained by the similarity calculation unit 34 of the model calculation unit 3b, a plurality of all-item measurement past performance data (X _j , Y), and a plurality of copies part item unmeasured historical performance data (X _{j, Y)} based on the data values of the data items corresponding to the predicted value y _p0 in (data values of the output variable Y), the predicted target data X predictive value of _0j y _p0 Is what you want.

Next, the operation of the output value prediction device Sb will be described. For example, when an output command is received from the input unit 5 by a user operation, the output value prediction device Sb executes an output value prediction program. By executing the output prediction program, the complementing unit 2a, the model calculation unit 3b, and the predicted value calculation unit 4b are functionally configured in the computer. Then, the output value prediction device Sb predicts the output value (predicted value) y _p0 from the prediction target data X _p0 based on the past performance data (X _j , Y) by the following operation.

In the prediction of the output Y _p, in the past record data DB 11 in the separation unit 1 of the output value prediction unit Sb, for example, historical performance data (X in tabular form (table format) as shown in FIG. 8 (A) _j , Y) is stored in advance, and the prediction target data _Xp0 is stored as shown in FIG. 8B by being input from the input unit 5. The past performance data table 2000 shown in FIG. 8A includes an output field 201 for registering an actually measured output value y _j , a similarity factor data field 202 for registering a similarity variable z _ji , and past performance data (X _j , distance field 203 for registering the distance _{d i} for the predicted target data _{X 0j} of Y), and the past actual data _(X j which is determined based on the distance _{d i,} Y) predicted target data _{X 0j} for the similarity w of It is configured with a similarity field 204 for registering _i , and has a record for each past performance data (X _j , Y). The similar factor data field 202 includes similar factor data item subfields 201-1 to 201-m corresponding to the number m of elements of the similar factor Z.

The similarity variable z _ji used for calculating the similarity w _i (distance d _i ) is a variable that is empirically known that the result (output) is similar if this data item is similar, All or part x of the factor X related to the predetermined output Y, and may coincide with the Z which is the whole or part of the factor X related to the predetermined output Y, and They may be mismatched or partially matched (partially mismatched).

In the present embodiment, the past performance data table 2000 includes past performance data including unmeasured data items having no data values in the first to nth similar factor data items (partial item unmeasured past performance data). ) (Z _j , Y). In the example shown in FIG. 8A, the first past performance data (Z _j , Y) is part item unmeasured past performance data, and the data value of the first similarity factor data item Z ₁ is Missing and does not have that data value. Therefore, the distance d _i and the similarity w _i are also missing data values, do not have that data value.

In the case where such past performance data (Z _j , Y) is stored in the past performance data DB unit 11 of the separation unit 1, the output value predicting device Sb first performs the separation unit 1 as shown in FIG. 7. In the same manner as the process of step S11 shown in FIG. 2 described above, the past data (Z _j , Y) is converted into the all-item measured past performance data (Z _j , Y) and some items not measured. The data is divided into past performance data (Z _j , Y) and stored in the all item measurement data DB unit 14 and the missing data DB unit 13 (S31).

For example, in the example shown in FIG. 8A, the first past performance data (z _1j , y ₁ ) from which the data value of the first similar factor data item Z ₁ is missing is partly unmeasured past performance data. The second to second past record data (z _2j , y ₂ ) to (z _nj , y _n ) are stored in the partial item unmeasured past record data table 2003 shown in FIG. 9B. Is the all item measurement past result data, and is stored in the all item measurement past result data table 2002 shown in FIG. Similar to the past performance data table 2000 shown in FIG. 8A, the all-item measurement past performance data table 2002 is similar to the output field 221 for registering the actually measured output value y _j and the similarity variable z _ji. The factor data field 222 (222-1 to 222-m), the distance field 223 for registering the distance d _i to the prediction target data X _0j of the past performance data (Z _j , Y), and the distance d _i are obtained. was historical performance data _(Z j, Y) is configured to include a similarity field 224 for registering the similarity _{w i} for the predicted target data _{X 0j} of, further comprising a data weight field 225 that registers the data weight _{t j} , All items measurement past performance data (X _j , Y) is provided for each record. In addition, the partial item unmeasured past result data table 2003 is similar to the past result data table 2000 shown in FIG. 8A. The output field 231 for registering the actually measured output value y _j and the similarity variable z _ji The similarity factor data field 232 (232-1 to 232-m) for registering the distance, the distance field 233 for registering the distance d _i for the prediction target data X _0j of the past performance data (Z _j , Y), and the distance d _i A data field for registering a data weight t _j , which includes a similarity field 234 for registering the degree of similarity w _i of the past performance data (Z _j , Y) obtained based on the prediction target data X _0j Field 235 is provided, and the number of all past measured data (Z _j , Y) is included in the number of past measured data (Z _j , Y) of some items. It has a number of records multiplied by a number. In step S31, the distance d _i , similarity w _i , and data weight t are calculated in the next steps S32 to S34, and the data values are not registered.

Next, the output value predicting device Sb uses the complementing unit 2a to store the data of the unmeasured data item in the partial unmeasured past performance data (Z _j , Y) in the same manner as the process of step S12 illustrated in FIG. values, a plurality of all items measured historical performance data (Z _{j, Y)} supplemented with the data values of the data items corresponding to the unmeasured data items in, the plurality of all items measured historical performance data (Z _{j, Y)} and A data weight t of each data of a plurality of copy partial item unmeasured past performance data (Z _j , Y) each supplementing the data value of the unmeasured data item is calculated (S32).

Next, the model calculation unit 3b uses the distance calculation unit 33 to store the plurality of all-item measured past performance data (Z _j , Y) and the plurality of copy partial item unmeasured past performance data (Z _j , Y). For each, a distance d _i to the prediction target data X _0j is calculated (step S33). This distance d _i is defined so as to represent the relationship between the two data, and for example, Euclidean distance, weighted Euclidean distance, normalized Euclidean distance, or the like is used. In the present embodiment, the weighted distance d _i defined by Equation 6 is used. Expression 6-1 is used to calculate the distance d _i to the prediction target data X _0j for all item measurement past performance data (Z _j , Y). Expression 6-2 is a partial copy item This is used when calculating the distance d _i to the prediction target data X _0j for the unmeasured past performance data (Z _j , Y). k indicates the kth copy partial item unmeasured past performance data (Z _j , Y). Note that Formula 6-3 may be used instead of Formula 6-1.

Here, h _j (j = 1 to m) is a weight at the weighted distance d _i , and is an adjustment parameter that determines which similarity factor data item z _j is important. The weight h _j can be determined by using a known method, and can be obtained by, for example, multiple regression analysis as disclosed in Japanese Patent No. 393441, and for example, Japanese Patent No. 3912215. As disclosed in the document, each data item can be obtained by normalizing it with its statistical value (for example, average value or standard deviation).

Next, the model calculation unit 3b uses the similarity calculation unit 34 to measure the plurality of all item measurement past performance data (Z _j , Y) and the plurality of copy partial item unmeasured past performance data (Z _j , Y). For each of the _above , the similarity w _i to the prediction target data X _0j is obtained based on the distance d _i obtained by the distance calculation unit 33 (step S34). The similarity w _j is an index representing how similar the two data are. For example, as the weighted Euclidean distance d _j is smaller, the similarity w _j is larger and takes a positive value. Defined. The similarity w _j is defined as in Expression 7, for example. Expression 7-1 is used when calculating the similarity w _i to the prediction target data X _0j for all item measurement past performance data (Z _j , Y). This is used when calculating the similarity w _i to the prediction target data X _0j for the item unmeasured past performance data (Z _j , Y). k indicates the kth copy partial item unmeasured past performance data (Z _j , Y).

Here, α is an adjustment parameter.

The predicted value calculation unit 4b then complements each of the plurality of all-item measured past performance data (Z _j , Y) and the plurality of copy partial item unmeasured past performance data (Z _j , Y). Data weight t _j obtained by the data weight calculation unit 22, the similarity w _i obtained by the similarity calculation unit 34 of the model calculation unit 3 b, a plurality of all-item measurement past performance data (X _j , Y) and a plurality based copying some items unmeasured historical performance data (X _{j, Y)} to each data value of the data item corresponding to the predicted value y _p0 in (data values y _j of the output variable _Y), the predicted target data X _{0j Is} calculated (step S35).

More specifically, as shown in Expression 8, the predicted value calculation unit 4b determines that the plurality of all item measurement past performance data (Z _j , Y) and the plurality of copy partial item unmeasured past performance data (Z _j , Y), a plurality of data weights t _j obtained by the data weight calculating unit 22 of the complementing unit 2a and each degree of similarity w _i obtained by the similarity calculating unit 34 of the model calculating unit 3b. All item measured past performance data (X _j , Y) and a plurality of copy partial item unmeasured past performance data (X _j , Y), each data value (output variable Y) of the data item corresponding to the predicted value _yp0 The predicted value y _p0 of the prediction target data X _0j is obtained by weighting the data value y _j ).

Then, the predicted value y _p0 of the prediction target data X _0j obtained in this manner is output to the output section 6.

As described above, in the output value prediction device Sb, the output value prediction method executed by the output value prediction device Sb, and the output value prediction program implemented in the output value prediction device Sb, the data value of the unmeasured data item is set. , Data values of data items corresponding to all the unmeasured data items of all item measurement past performance data (Z _j , Y) are used. For this reason, since the number of usable past performance data (Z _j , Y) further increases, the output value prediction apparatus Sb, the output value prediction method, and the output value prediction program having such a configuration have data values. When past performance data (Z _j , Y) including missing data items is used, a predicted value of an output value can be obtained with higher prediction accuracy.

In the output value prediction device Sb, the output value prediction method, and the output value prediction program of the present embodiment, each data weight t _j of the plurality of all item measurement past performance data (Z _j , Y) is a predetermined value. A constant (1 in the above example) is used, and each data weight t _j of the plurality of copy partial item unmeasured past performance data (Z _j , Y) is the predetermined constant as the plurality of all item measurement past performances. The value is divided by the number of data. In this way, the plurality of all-item measured past performance data (Z _j , Y) that do not need to be supplemented with data values, and the plurality of copied partial item unmeasured past performance data (Z _j , Y) supplemented with data values, Are differentiated by making the values of the data weights t _j different from each other, the output value prediction device Sb, the output value prediction method, and the output value prediction program having such a configuration are missing data values. When the past performance data (Z _j , Y) including the data item is used, the predicted value of the output value can be obtained with higher prediction accuracy.

In the output value prediction device Sb, the output value prediction method, and the output value prediction program of the present embodiment, the similarity value w _j is used when obtaining the predicted value _yp0. The prediction method Sb, the output value prediction method, and the output value prediction program, when using past performance data (Z _j , Y) including data items with missing data values, output values with higher prediction accuracy. A predicted value can be obtained.

Next, another embodiment will be described.
(Third embodiment)
FIG. 10 is a block diagram illustrating a configuration of an output value prediction apparatus according to the third embodiment. FIG. 11 is a flowchart showing the operation of the output value prediction apparatus in the third embodiment. FIG. 12 is a diagram for explaining past performance data stored in a past performance data database unit in the output value prediction apparatus of the third embodiment. FIG. 13 is a diagram for explaining all item measurement past performance data stored in the all item measurement data database and some item unmeasured past performance data stored in the missing data database in the output value prediction apparatus of the third embodiment. FIG. FIG. 14 is a diagram for explaining a complementary operation of the output value prediction apparatus according to the third embodiment.

The output value prediction device Sa in the first embodiment outputs an output value based on a plurality of all-item measured past performance data (X _j , Y) and a plurality of copy partial item unmeasured past performance data (X _j , Y). predictive model _{y = f (a j, X} j) identified, the identified output value prediction model _{y = f (a j, X} j) prediction target data _{X 0j} by substituting the predicted target data _{X 0j} in Although calculated predicted value _{y p0,} the output value prediction unit Sc in the third embodiment, the predicted value of the predicted target data _{X 0j} by a given output value prediction model _{y = f (a j, X} j) y p0 And the calculated predicted value y _p0 as a temporary predicted value (provisional predicted value) y _p0, and by _adding the prediction error y _pj of the prediction target data X _{0j to} the temporary predicted value y _p0 , Final predicted value y of the prediction target data _{X0j Pre} is obtained. Here, the prediction error _{y pj} prediction target data _{X 0j} is the predicted value of the past record data _(X j, Y) by a given output value prediction model _{y = f (a j, X} j) ( model predictive value) seeking y _pj, the obtained historical performance data _(X j, Y) model predicted value _{y pj} and historical performance data _(X j, Y) data values of the data items corresponding to the model prediction value _{y pj} in (the output of variable Y data values _{y p)} and the difference between the historical performance data _(X j of, calculated as the prediction error △ _j of Y), the prediction error △ _j and historical performance data (X this past record data _(X j, Y) _j , Y) and the degree of similarity w _j between the prediction target data X _0j .

The given output value prediction model y = f (a _j , X _j ) may be obtained in advance by a multiple regression method, a PLS (Partial Least Square) regression method, or the like, and may be given to the output value prediction device Sc. The output value prediction device Sc of the present embodiment calculates the given output value prediction model y = f (a _j , X _j ) at that time point before calculating the prediction value (final prediction value). Based on past performance data (X _j , Y), the weight multiple regression method is used.

  Such an output value prediction device Sc of the third embodiment includes, for example, a complementing unit 2a, a model calculation unit 3c, a prediction value calculation unit 4c, and a model prediction error calculation unit 7, as shown in FIG. Furthermore, in this embodiment, the separation part 1, the input part 5, and the output part 6 are provided.

  The separating unit 1, the complementing unit 2a, the input unit 5 and the output unit 6 in the output value predicting device Sc of the third embodiment are respectively the separating unit 1 and the complementing unit 2a in the output value predicting device Sa of the first embodiment. Since it is the same as that of the input part 5 and the output part 6, the description is abbreviate | omitted.

The model calculation unit 3c obtains what is necessary for _obtaining the final predicted value y _pre of the prediction target data _X0j based on the past performance data ( _Xj , Y) stored in the separation unit 1. It is. In this embodiment, the model calculation unit 3c predicts each of the plurality of all-item measured past performance data (X _j , Y) and each of the plurality of copy partial item unmeasured past performance data (X _j , Y). a distance calculation unit 33 for determining the distance d _j of the target data X _0j, the plurality of all items measured historical performance data (X _{j, Y)} and said plurality of copying some items unmeasured historical performance data (X _j, Y ), The similarity calculation unit 34 for _obtaining the similarity w _j to the prediction target data X _0j based on the distance d _j obtained by the distance calculation unit 33, and the plurality of all-item measurement past performance data (X _j, Y) and said plurality of copying some items unmeasured historical performance data (X _{j, Y} for each), prompted by data weight t _j and the similarity calculation unit 34 obtained by the data weight calculating unit 22 Was based on the similarity w _j, the predicted target used when obtaining the prediction error y _pj data X _0j, all items measured historical performance data (X _{j, Y)} and copying some items unmeasured historical performance data (X a prediction error weight calculating unit 35 for obtaining a weight (prediction error weight) ω _j of the prediction error Δ _j in _j , Y).

The model prediction error calculation unit 7 includes past performance data (X _j , Y) (the plurality of all item measurement past performance data (X _j , Y)) and the plurality of copy partial item unmeasured past performance data (X _j , Y)), the model predicted value y _pj of the past performance data (X _j , Y) is obtained from the given output value prediction model y = f (a _j , X _j ), and the obtained past performance data ( X _j, Y) model predicted value _{y pj} and historical performance data _(X j, Y) difference of the past record of the data values of the data items corresponding to the predicted value _{y pj} (data value _{y p} of the output variable Y) in the This is obtained as a prediction error Δ _j of data (X _j , Y). The model prediction error calculation unit 7 obtains the given output value prediction model y = f (a _j , X _j ) by the weighted multiple regression method in the past actual data (X _j , Y) (the above-mentioned) Output value prediction model y = f (a _j , X _j ) based on a plurality of all-item measured past performance data (X _j , Y) and the plurality of copied partial item unmeasured past performance data (X _j , Y)) ) For calculating the coefficient parameter a _j , past performance data (X _j , Y) (the plurality of all item measurement past performance data (X _j , Y), and the plurality of copied partial items unmeasured past) actual data _(X j, Y) for each), the model prediction value y of the parameter calculation unit 71 output values are identified by the prediction model _{y = f (a j, X} j) by historical performance data _(X j, Y) model predictive value meter to determine the _pj A part 72, the obtained historical performance data _(X j, Y) model predicted value _{y pj} and historical performance data _(X j, Y) data values of the data items corresponding to the model prediction value _{y pj} in (output variables And a model error calculation unit 73 that obtains a difference from the Y data value y _p ) as the prediction error Δ _j of the past performance data (X _j , Y).

The predicted value calculation unit 4c includes past performance data ( _Xj , Y) (the plurality of all item measurement past performance data ( _Xj , Y) and the plurality of copy partial item unmeasured past performance data ( _Xj , Y). )) For each data weight t _j obtained by the data weight calculator 22, each similarity w _j obtained by the similarity calculator 34, and past performance data (X _j) obtained by the model error calculator 73. Y) Based on each prediction error Δ _j of Y), a prediction error Δ ₀ in the final prediction value y _pre of the prediction target data X _0j is obtained, and the output value prediction model y = f (a _j , X _j ) to obtain the sum of the provisional prediction value y _p0 obtained by substituting the prediction target data X _0j into the prediction target data X _0j and the prediction error Δ ₀ in the final prediction value y _pre of the prediction target data X _0j so And requests a serial final prediction value y _pre. More specifically, the predicted value calculation unit 4c is obtained by the model error calculation unit 73 using each data weight t _j obtained by the data weight calculation unit 22 and each similarity w _j obtained by the similarity calculation unit 34. The prediction error Δ ₀ in the final prediction value y _pre of the prediction target data X _0j is obtained by weighting each prediction error Δ _j of the past performance data (X _j , Y).

Next, the operation of the output value prediction device Sc will be described. For example, when an output command is received from the input unit 5 by a user operation, the output value prediction device Sc executes the output value prediction program. By executing this output prediction program, the complementing unit 2a, model calculation unit 3b, prediction value calculation unit 4b, and model prediction error calculation unit 7 are functionally configured in the computer. Then, the output value prediction device Sc predicts the output value (predicted value) y _pre from the prediction target data X _p0 based on the past performance data (X _j , Y) by the following operation.

In the prediction of the output Y _p, in the past record data DB 11 in the separation unit 1 of the output value prediction unit Sc, for example, historical performance data (X j in tabular form (table format) as shown in FIG. _12, Y ) Is stored in advance, and is input from the input unit 5 to store the prediction target data _Xp0 . The past performance data table 3000 shown in FIG. 12 includes an output field 301 for registering the actually measured output value y _j , a factor data field 302 for registering the data x _ji of the factor X related to the predetermined output Y, and the past performance. data _(X j, Y) model predictive value field 303 for registering the model prediction value _{y pj} of the prediction error field 304 for registering the prediction error △ _j past actual data _(X j, Y), the similarity variable _{z ji} The similar factor data field 305 for registering, the distance field 306 for registering the distance d _i to the prediction target data X _0j of the past result data (X _j , Y), and the past result data obtained based on the distance d _{i (X} j, Y) configured with a similarity field 307 for registering the similarity _{w i} for the predicted target data _{X 0j} of It is provided with a record for each historical track record data _(X j, Y). The factor data field 302 includes factor data item subfields 302-1 to 302-n corresponding to the number n of elements of the factor X. The similar factor data field 305 includes similar factor data item subfields 305-1 to 305-m corresponding to the number m of elements of the similar factor Z.

In the present embodiment, the past performance data table 3000 includes past performance data (partially unmeasured past performance data) (X _j , Y) including unmeasured data items having no data value in the data items. Is present. In the example shown in FIG. 12, the first and second (NL) pieces of past performance data (X _j , Y) are unmeasured past performance data for some items, and the first past performance data. The output field (first data item) 301 of (X _j , Y) is missing the data value, does not have the data value, and the second factor data item subfield (third data item). Item) 302-2 has a missing data value, does not have the data value, and the first and second factor data item subfields of the second past performance data (X _j , Y) The second and third data items) 302-1 and 302-2 are missing their data values and do not have those data values. For this reason, the model predicted value y _pj and the prediction error Δ _{j of} the past performance data (X _j , Y) are missing, and do not have the data value.

In the case where such past performance data (X _j , Y) is stored in the past performance data DB unit 11 of the separation unit 1, the output value predicting device Sc, first, as shown in FIG. In the same manner as the process of step S11 shown in FIG. 2 described above, the past data (X _j , Y) is converted into the all-item measured past performance data (X _j , Y) and some items not measured. The data is divided into past performance data (X _j , Y) and stored in all item measurement data DB unit 14 and missing data DB unit 13 (S41).

For example, in the example shown in FIG. 11, the (N−L) number of past performance data (X _j , Y) such as the first and second, etc. are some item unmeasured past performance data, and FIG. Are stored in the unmeasured past performance data table 3002 for some items, and the remaining L past performance data (X _j , Y) are all item measurement past performance data, and all item measurements shown in FIG. Stored in the past performance data table 3001. The fields 311 to 317 of the all item measurement past performance data table 3001 and the fields 321 to 327 of the partial item unmeasured past performance data table 302 are the fields 301 to 307 of the past performance data table 3000 shown in FIG. It is the same. At the stage of step S41, historical performance data _(X j, Y) model predicted value _{y pj} of the prediction error △ _j, the distance _{d i} and the similarity _{w i} is calculated in the following step, the data The value is not registered.

Next, the output value prediction device Sc uses the complementing unit 2a to store the data of the unmeasured data items in the partial unmeasured past performance data (X _j , Y) as in the process of step S12 shown in FIG. 2 described above. values, a plurality of all items measured historical performance data (X _{j, Y)} supplemented with the data values of the data items corresponding to the unmeasured data items in, the plurality of all items measured historical performance data (X _{j, Y)} and A data weight t _j of each of the plurality of copy partial item unmeasured past performance data (X _j , Y) each supplementing the data value of the unmeasured data item is calculated (S42). As described in the first embodiment, each data weight t _j of all item measurement past performance data (X _j , Y) is set to a predetermined constant (for example, 1) equal to each other, and some item unmeasured past performance data (X _{j, Y)} each data weight t of _j, the predetermined constant the plurality of all items measured historical performance data (X _{j, Y)} in the value (the example divided by the number of t _{i (k)} = 1 / (number of all item measurement past result data = L, Σt _i (k) = 1, Σ is the sum of t _i (k) from 1 to L)). FIG. 14 shows an all-item measurement past performance data table 303 (FIG. 14A) for registering all-item measurement past performance data (X _j , Y) after execution of step S42, and some items not yet stored. A partial item unmeasured past performance data table 3004 (FIG. 14B) for registering measurement past performance data (X _j , Y) is shown. These all-item measured past performance data table 303 and some item unmeasured past performance data table 3004 are respectively the same fields 331 to 337; 341 to fields 301 to 307 of the past performance data table 3000 shown in FIG. 347, and further includes a data weight field 338; 348 for registering the data weight t _j and a prediction error weight field 339; 349 for registering the prediction error weight ω _j . In FIG. 14B, the nth factor data item subfield 342-n of the Nth past performance data (N-Lth partial item unmeasured past performance data) ( _Xj , Y) is shown. The data value of the expanded complementation data is shown in Fig. 5, but the other complements are also complemented in the same way.

Next, the model prediction error calculation unit 7 uses the parameter calculation unit 71 to calculate the coefficient parameter of the output value prediction model y = f (a _j , X _j ) based on the past performance data (X _j , Y) at that time. By obtaining a _j , the output value prediction model y = f (a _j , X _j ) is identified by the weighted multiple regression method, and the model predicted value calculation unit 72 uses the plurality of all-item measurement past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y)) by the output value prediction model y = f (a _j , X _j ) identified by the parameter calculation unit 71 A model predicted value y _pj of the past performance data (X _j , Y) is obtained (step S43).

Next, the model prediction error calculation unit 7 uses the model prediction value y _{pj of} the past performance data (X _j , Y) obtained by the model prediction value calculation unit 72 and the past performance data (X _j , The difference between the data value of the data item corresponding to the model predicted value y _pj in Y) (the data value y _p of the output variable Y) and the prediction error Δ _j (= y _j − of the past performance data (X _j , Y)). y _pj ) (step S44).

Next, the model calculation unit 3c causes the distance calculation unit 33 to store the plurality of all-item measured past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y). For each, a distance d _i to the prediction target data X _0j is calculated (step S45). This distance d _i, in the present embodiment, weighted distance d _i, which is defined by the equation 6 is used.

Next, the model calculation unit 3c uses the similarity calculation unit 34 to measure the plurality of all item measurement past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y). _Is calculated based on the distance d _i obtained by the distance calculation unit 33 for the similarity w _i to the prediction target data X _0j (step S46). The similarity w _j is defined as in Expression 7 in this embodiment.

Next, the model calculation unit 3c uses the prediction error weight calculation unit 35 to measure the plurality of all item measurement past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y). ) For each, a prediction error weight ω _j is calculated (step S47).

This prediction error weight ω _j is defined by the product of the similarity w _j and the data weight t _j as shown in Equation 9. That is, the prediction error weight ω _j for each of the plurality of all-item measured past performance data (X _j , Y) is defined by Equation 9-1, and the plurality of copy partial item unmeasured past performance data (X _j , The prediction error weight ω _j for each of Y) is defined by Equation 9-2.

Then, the predicted value calculation unit 4c outputs past performance data (X _j , Y) (the plurality of all item measurement past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y)), each model error Δ _j obtained by the model error calculation unit 73, each prediction error weight ω _j obtained by the prediction error weight calculation unit 35, and the past obtained by the model error calculation unit 73. Based on each prediction error Δ _j of the actual data (X _j , Y), a prediction error Δ ₀ in the final prediction value y _pre of the prediction target data X _0j is obtained (step S48), and is identified by the parameter calculation unit 71. output value prediction model y was _{= f (a j, X j} ) to final prediction value _{y pr} of the predicted target data _{X 0j} provisional obtained by substituting a predicted value _{y p0} predicted target data _{X 0j} Determining the predicted value y _pre by calculating the sum of the prediction error △ ₀ in. More specifically, as shown in Expression 10, the predicted value calculation unit 4 c uses the past performance data (X) obtained by the model error calculation unit 73 with each prediction error weight ω _j obtained by the prediction error weight calculation unit 35. _j, by weighting each prediction error △ _j of Y), determining the prediction error △ ₀ in the final predicted value _{y pre} prediction target data _{X 0j.} Then, the predicted value calculation unit 4c obtains the final predicted value by _obtaining the sum of the temporary predicted value y _p0 and the predicted error Δ ₀ of the final predicted value y _pre as shown in Expression 11. _Find y _pre .

Then, the final predicted value y _{pre of} the prediction target data X _0j obtained in this way is output to the output unit 6.

Thus, in the output value prediction device Sc of this embodiment, the output value prediction method executed by the output value prediction device Sc, and the output value prediction program implemented in the output value prediction device Sc, the data value of the unmeasured data item is set. , Data values of data items corresponding to all the unmeasured data items of all item measurement past performance data (Z _j , Y) are used. For this reason, since the number of usable past performance data (Z _j , Y) further increases, the output value prediction device Sc, the output value prediction method, and the output value prediction program having such a configuration have data values. When past performance data (Z _j , Y) including missing data items is used, a predicted value of an output value can be obtained with higher prediction accuracy.

In the output value prediction device Sc, the output value prediction method, and the output value prediction program of the present embodiment, each data weight t _j of the plurality of all-item measurement past performance data (Z _j , Y) is a predetermined value. A constant (1 in the above example) is used, and each data weight t _j of the plurality of copy partial item unmeasured past performance data (Z _j , Y) is the predetermined constant as the plurality of all item measurement past performances. The value is divided by the number of data. In this way, the plurality of all-item measured past performance data (Z _j , Y) that do not need to be supplemented with data values, and the plurality of copied partial item unmeasured past performance data (Z _j , Y) supplemented with data values, Are differentiated by making the values of the data weights t _j different from each other, the output value prediction device Sc, the output value prediction method, and the output value prediction program having such a configuration are missing data values. When the past performance data (Z _j , Y) including the data item is used, the predicted value of the output value can be obtained with higher prediction accuracy.

In the output value prediction device Sc, the output value prediction method, and the output value prediction program of the present embodiment, a plurality of all item measurement past performance data (Z _j , Y) and a plurality of copy partial item unmeasured past results For each of the data (Z _j , Y), the similarity w _j to the prediction target data X _0j is obtained, and the data weight t _j , the similarity w _j and the past performance data (Z _j , Y) based on the prediction error △ _j, the prediction error △ _p0 is obtained in the final predicted value _{y pre} prediction target data _{X 0j,} the final prediction value _{y pre} is calculated. Since the similarity w _j is used when the final predicted value y _pre is obtained in this way, the output value predicting device Sc, the output value predicting method, and the output value predicting program of the present embodiment are data When using past performance data (Z _j , Y) including a data item whose value is missing, the predicted value of the output value can be obtained with even higher prediction accuracy.

In the above embodiment, the prediction error △ _j past actual data _(X j, Y) is, in the past actual data _(X j, Y) model predicted value _{y pj} and historical performance data of the _(X j, Y) data values of the data items corresponding to the model prediction value y _pj have been the difference between (the data value y _p of the output variable _Y), if it is defined as shown in equation 12, as shown in equation 13 May be defined.

Next, another embodiment will be described.
(Fourth embodiment)
In the output value prediction devices Sa to Sc of the first to third embodiments, each data weight of a plurality of all item measured past performance data (X _j , Y) and a plurality of copy partial item unmeasured past performance data (X _j , Y) are different values from each other, and they are treated differently, but they are treated in the same way even if the number of unmeasured data items to be complemented is different. The output value predicting apparatus Sd of the fourth embodiment varies the handling according to the number of unmeasured data items to be complemented.

  FIG. 15 is a block diagram illustrating a configuration of an output value prediction apparatus according to the fourth embodiment. Such an output value prediction apparatus Sd in the fourth embodiment includes, for example, a complementing unit 2a, a model calculation unit 3d, a prediction value calculation unit 4c, and a model prediction error calculation unit 7, as shown in FIG. Furthermore, in this embodiment, the separation part 1, the input part 5, and the output part 6 are provided.

  The separation unit 1, the complement unit 2a, the input unit 5 and the output unit 6 in the output value prediction device Sd of the fourth embodiment are respectively the separation unit 1, the complement unit 2a, and the output value prediction device Sa of the first embodiment. Since it is the same as that of the input part 5 and the output part 6, the description is abbreviate | omitted. And since the predicted value calculation part 4c in the output value prediction apparatus Sd of this 4th Embodiment is the same as the predicted value calculation part 4c in the output value prediction apparatus Sc of 3rd Embodiment, the description is abbreviate | omitted.

The model calculation unit 3d obtains what is necessary to obtain the final predicted value y _pre of the prediction target data _X0j based on the past performance data ( _Xj , Y) stored in the separation unit 1. It is. In this embodiment, the model calculation unit 3d predicts each of the plurality of all-item measurement past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y). a distance calculation unit 33 for determining the distance d _j of the target data X _0j, the plurality of all items measured historical performance data (X _{j, Y)} and said plurality of copying some items unmeasured historical performance data (X _j, Y ), A similarity calculation unit 34 for _obtaining a similarity w _j to the prediction target data X _0j based on the distance d _j obtained by the distance calculation unit 33, and the plurality of copy partial item unmeasured past results data (X _{j, Y)} for each of the complementary influence degree calculation unit 36 for obtaining the complementary impact Q _j is an index indicating the degree of influence of complementary data item has on the final prediction value y _pre, de Data weight t _j obtained at data weight calculating unit _22, the similarity w _j obtained by the similarity calculation unit _34, and, based on the determined at complementing influence degree calculation part 36 complementary impact Q _j, the predicted The prediction error Δ _j of all item measured past performance data (X _j , Y) and copy partial item unmeasured past performance data (X _j , Y), which is used when _obtaining the prediction error y _pj of the target data X _0j A prediction error weight calculation unit 37 for obtaining a weight (prediction error weight) ω _j .

The complementary influence calculation unit 36 obtains an unmeasured data item complementary influence q _j for each unmeasured data item to be complemented defined by Expression 14, for example, and, for example, as shown in Expression 15, unmeasured data item complementary impact q _j obtained by adding the amount of unmeasured data items complemented obtains a complementary impact Q _j, and, on the basis of the obtained complement impact Q _j complement and requests the data weight _{r j} due to the influence of _{Q j.} Expression 15 is an expression representing the complementary influence Q _j when the data items Y, X _k , and X _l are complemented in the i-th copied partial item unmeasured past performance data. Complementary influence Q _j by data weight (complementary impact data weight) r _j is the total item measured historical performance data (X _{j, Y),} for example, is a 1, the copying part item unmeasured historical performance data (X _j , Y) is defined as r _j = 1 / exp (β × Q _j ), for example, where β is the adjustment parameter.

Here, in Expression 14, the influence of the element j of each factor X on the prediction error Δ _j is obtained by partial differentiation of the output value prediction model y = f (a _j , X _j ) with the element j (note that If an unmeasured data item remains, the average value is used (for example, in the case of the output variable Y, the average value of the output variable Y), and the degree of variation of the unmeasured data item is a standard deviation. Is defined by seeking in

Then, as shown in Expression 16, the prediction error weight calculation unit 37 obtains a prediction error weight ω _j defined by the product of the similarity w _j , the data weight t _j, and the complementary influence degree Q _j . That is, the prediction error weight ω _j for each of the plurality of all-item measured past performance data (X _j , Y) is defined by Equation 16-1, and the plurality of copy partial item unmeasured past performance data (X _j , The prediction error weight ω _j for each of Y) is defined by Equation 16-2.

The output value predicting apparatus Sd in the fourth embodiment calculates the complementary influence degree Q _j by the complementary influence degree calculation part 36 before the prediction error weight calculation part 37 calculates the prediction error weight ω _j. An operation similar to the operation of the output value prediction apparatus Sc in the third embodiment described with reference to FIG.

The output value prediction apparatus Sd in this embodiment, the output value prediction method executed by the output value prediction apparatus Sd, and the output value prediction program implemented in the output value prediction apparatus Sd are the output value prediction in the third embodiment. The same effect as the output value prediction method executed by the device Sc, the output value prediction device Sc, and the output value prediction program implemented in the output value prediction device Sc is achieved. In the output value predicting apparatus Sd, the output value predicting method, and the output value predicting program according to the present embodiment, the complementation influence degree Q _j is used. The degree of influence of the determined unmeasured data item on the final predicted value y _pre is taken into account. For this reason, the output value prediction device Sd, the output value prediction method, and the output value prediction program according to the present embodiment can obtain the predicted value of the output value with even higher prediction accuracy.

Next, another embodiment will be described.
(Fifth embodiment)
In the output value prediction devices Sa to Sd in the first to fourth embodiments, a plurality of copy partial item unmeasured past performance data (X _j , Y) in which the data value of the unmeasured data item is supplemented by the data complementing unit 21 are stored. All are used, but unmeasured data with data values (impossible data values) that cannot be obtained with the data values of the remaining data items other than the unmeasured data items in the past measured data (X _j , Y) of some items The data value of the item may be supplemented. Therefore, the output value prediction device Se in the fifth embodiment uses the copy partial item unmeasured past performance data (X _j , Y) in which the data value of the unmeasured data item is supplemented by such a data value that cannot be obtained, By removing the prediction value y _p0 or the final prediction value y _{pre from} the calculation, the possibility is reduced.

  FIG. 16 is a block diagram illustrating a configuration of an output value prediction apparatus according to the fifth embodiment. Such an output value prediction device Se in the fifth embodiment includes, for example, a complementing unit 2b, a model calculation unit 3c, a prediction value calculation unit 4c, and a model prediction error calculation unit 7, as shown in FIG. Furthermore, in this embodiment, the separation part 1, the input part 5, and the output part 6 are provided.

  The separation unit 1, the input unit 5 and the output unit 6 in the output value prediction device Se of the fifth embodiment are respectively the separation unit 1, the input unit 5 and the output unit 6 in the output value prediction device Sa of the first embodiment. Since it is the same, the description is abbreviate | omitted. The model calculation unit 3c, the prediction value calculation unit 4c, and the model prediction error calculation unit 7 in the output value prediction device Se of the fifth embodiment are respectively the model calculation unit 3c in the output value prediction device Sc of the third embodiment. Since it is the same as the predicted value calculation unit 4c and the model prediction error calculation unit 7, the description thereof is omitted.

The complementing unit 2b uses the data values of the unmeasured data items based on the all-item measured past performance data ( _Xj , Y) in order to use the unmeasured past performance data ( _Xj , Y) of some items for output value prediction. The copy partial item unmeasured past performance data (X _j , Y) in which the data value of the unmeasured data item is supplemented by the irretrievable data value is used as the predicted value y _p0 or the final prediction It is removed from the calculation of the value y _pre . Complementing portion 2b, some items unmeasured historical performance data (X _{j, Y),} and copied to the number corresponding to the number of the plurality of all the items past performance data (X _{j, Y),} a plurality of the these copied one For each part item unmeasured past performance data (copy partial item unmeasured past performance data) (X _j , Y), the data value of the unmeasured data item is set to the plurality of all item measured past performance data (X _j , Y), a data complementing unit 21 that compensates with the data value of the data item corresponding to the unmeasured data item, and a copy partial item unmeasured past record in which the data value of the unmeasured data item is supplemented with the irretrievable data value data _(X j, Y) to remove from the calculation of the predicted value _{y p0} or final prediction value _{y pre,} wherein the plurality of all items measured historical performance data _(X j, Y) and a plurality of copying some items unmeasured Removed by actual data (X _{j, Y)} based on, as the predetermined condition for determining that the value to compensate the raw measurement data items, determine the range of values △ _missing data values can be taken by the unmeasured data items For each data of the model error tolerance calculation unit 24 and the plurality of all-item measured past performance data (X _j , Y) and the plurality of copy partial item unmeasured past performance data (X _j , Y), And a data weight calculator 25 for obtaining a weight (data weight t).

More specifically, the model error tolerance range calculation unit 24, the similarity w _{i (distance} d _i) in the similarity variable z _ji used to calculate the part item unmeasured historical performance data (X _j, Y) and distance D _i between all item measurement past performance data (X _j , Y) is obtained.

The distance D _i is a normalized variable of the l-th all items measured past results ( _i.e. , the similarity variable in the i-th partial items unmeasured past results data (X _j , Y) normalized). When the similarity variable in X _j , Y) is z _lk , it is defined by, for example, Expression 17-1. Further, for example, the distance D _i is defined by Expression 17-2.

Then, the model error allowable range calculation unit 24 extracts all item measurement past performance data (X _j , Y) in which the distance D _i is equal to or less than a predetermined threshold, and this extracted all item measurement past performance data (X _j , the prediction error △ _total look of Y), and based on the average value ^~ △ _total and its standard deviation sigma _{△ whole} city, and requests a range of data values is possible values △ _missing unmeasured data items. The range of the value Δ _missing that the data value of the unmeasured data item can take is defined by, for example, Expression 18.

Here, γ is an adjustment parameter.

In the present embodiment, the data weight calculation unit 25 sets each data weight t _j of the plurality of all-item measurement past performance data (X _j , Y) to a predetermined constant (for example, 1), Each of the data weights t _j of the past measured data (X _j , Y) of the unmeasured partial item of “0” is set to 0 when the expression 18 is not satisfied, and the predetermined constant is satisfied when the expression 18 is satisfied. Is divided by the number of past unmeasured past performance data of the copy part item satisfying the expression 18 (in the example, 1 / (number of unmeasured past performance data of the copy partial item satisfying the expression 18)). Is. In this way, when the equation 18 is not satisfied, by setting each data weight t _j of the copy partial item unmeasured past performance data (X _j , Y) to 0, the unmeasured data by the irretrievable data value The copy partial item unmeasured past performance data (X _j , Y) supplemented with the data value of the item is removed from the calculation of the predicted value y _p0 or the final predicted value y _pre , thereby reducing the possibility. .

The output value predicting device Se in the fifth embodiment operates in substantially the same manner as the output value predicting device Sc in the third embodiment, but the value Δ _missing range that the data value of the unmeasured data item can take. the calculation, since the model-predicted value Y _pj and the prediction error △ _j is needed, before the calculation of the range the unmeasured data data value of the item can take a value △ _missing, the model predicted value Y _pj and the prediction error △ Calculation of _j is executed. That is, the output value predicting device Se in the fifth embodiment includes each step in the order of the step S41, the step S43, the step S44, the step S42, the step S45, the step S46, the step S47, and the step S48. Next, the predicted value calculation unit 4c is obtained by substituting the prediction target data _X0j into the output value prediction model y = f (a _j , X _j ) identified by the parameter calculation unit 71. The prediction value y _pre is obtained by calculating the sum of the provisional prediction value y _p0 and the prediction error Δ ₀ in the final prediction value y _pre of the prediction target data X _0j , and the prediction target thus obtained The final predicted value y _{pre of the} data X _0j is output to the output unit 6. Here, in step S42 in the fifth embodiment, the data weight tj is obtained by the data weight calculator 25 as described above.

The output value prediction device Se in this embodiment, the output value prediction method executed by the output value prediction device Se, and the output value prediction program implemented in the output value prediction device Se are the output value prediction in the third embodiment. The same effect as the output value prediction method executed by the device Sc, the output value prediction device Sc, and the output value prediction program implemented in the output value prediction device Sc is achieved. In the output value predicting device Se, the output value predicting method, and the output value predicting program according to the present embodiment, the copy partial item unmeasured past performance in which the data value of the unmeasured data item is supplemented by the irretrievable data value. Since the data (X _j , Y) is removed from the calculation of the predicted value y _p0 or the final predicted value y _pre , the possibility that the data value of the unmeasured data item is supplemented by the impossible data value is reduced. . For this reason, the output value prediction device Sd, the output value prediction method, and the output value prediction program according to the present embodiment can obtain the predicted value of the output value with even higher prediction accuracy.

  The output value prediction devices Sa to Se of the first to fifth embodiments can be applied to various cases, and are particularly preferably applied to the following cases. For example, in products manufactured through various manufacturing processes in a relatively large-scale manufacturing plant, such as manufacturing of steel products and chemical products, for example, according to the input amount, operation input amount, time passage, etc. In many cases, the output value in each manufacturing process and the output value of the final process directly connected to the product change every moment. In such a case, detecting all the factors that affect the output value and obtaining the detected value is often omitted because it requires time, labor, and cost. For example, in the converter operation, in the case of measuring the concentration and temperature of each component such as phosphorus and carbon after completion of blowing, the measurement is often omitted. In such a case, the output value prediction devices Sa to Se of the first to fifth embodiments can use the past performance data lacking data values, and the past performance data lacking such data values. Even in the case where is used, it is possible to obtain a predicted value with higher prediction accuracy, and thus it is preferably applied to such a case.

  In order to express the present invention, the present invention has been properly and fully described through the embodiments with reference to the drawings. However, those skilled in the art can easily change and / or improve the above-described embodiments. It should be recognized that this is possible. Therefore, unless the modifications or improvements implemented by those skilled in the art are at a level that departs from the scope of the claims recited in the claims, the modifications or improvements are not covered by the claims. To be construed as inclusive.

Sa, Sb, Sc, Sd, Se Output value prediction device 1 Separation unit 2a, 2b Complement unit 3a, 3b, 3c, 3d Model calculation unit 4a, 4b, 4c Prediction value calculation unit 7 Model prediction error calculation unit

Claims (11)

An output value prediction method used in an output value prediction apparatus,
The output value prediction device includes a partial item unmeasured past performance data including an unmeasured data item having no data value in the plurality of data items in a plurality of past performance data including a plurality of data items. The plurality of data items in the past result data are copied to the number corresponding to the number of all the item measured past result data having data values, and the partial item unmeasured past result data is copied. For each copy partial item unmeasured past performance data, the data value of the unmeasured data item is supplemented with the data value of the data item corresponding to the unmeasured data item in the plurality of all item measured past performance data Steps,
A weight calculating step for obtaining a weight for each of the plurality of all-item measured past performance data and the plurality of copy partial item unmeasured past performance data each supplemented with a data value in the complementing step; ,
The output value predicting device includes the plurality of weights obtained in the weight calculation step, the plurality of all item measurement past performance data, and a plurality of copy partial items not measured past each supplemented with a data value in the complementing step. An output value prediction step for predicting an output value for the prediction target data to be predicted based on the actual data and obtaining a predicted value ,
In the weight calculation step, each weight of the plurality of all item measured past performance data is set to a predetermined constant, and each weight of the plurality of copy partial item unmeasured past performance data is set to the predetermined constant. output value prediction method to all items measured values and to Rukoto divided by the number of historical performance data and features. The output value prediction step includes :
A model calculation step of obtaining the input and output variables and the weighted regression method the output value prediction model using the plurality of weights calculated by the weight calculating step which represents a relationship,
The output value prediction model which is obtained by the model calculation step, the output value prediction according to claim 1, wherein the obtaining Bei the predicted value calculation step of obtaining the prediction value by substituting the prediction target data Method. The output value prediction step includes :
For each of the previous SL plurality of copying some items unmeasured historical performance data of the data value supplemented by a plurality of all items measured historical performance data and the complementary step, obtaining the degree of similarity that represents the difference between the predicted target data A similarity calculation step;
The plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the plurality of all-item measured past performance data and the plurality of data values supplemented in the complementing step, respectively. based of the respective data values of the data items corresponding to the predicted value in the copying part item unmeasured historical performance data, according to claim 1, wherein the obtaining Bei the predicted value calculation step of obtaining the predicted value Output value prediction method. The predicted value calculating step includes the plurality of weights obtained in the weight calculating step and the plurality of similarities obtained in the similarity calculating step, the plurality of all-item measured past performance data and the data value in the complementing step. 4. The predicted value is obtained by weighting each data value of a data item corresponding to the predicted value in the plurality of copy partial item unmeasured past performance data each supplementing each of the copy items. Output value prediction method. The output value prediction step includes:
A given output representing a relationship between an input variable and an output variable for each of the plurality of all-item measured past performance data and each of the plurality of copied partial item unmeasured past performance data each supplemented with a data value in the complementing step A model prediction value is obtained by substituting the data value of the data item corresponding to the input variable into the value prediction model, and the difference between the model prediction value and the data value of the data item corresponding to the output variable is determined as a prediction error. A prediction error calculation step obtained as :
For each of the previous SL plurality of copying some items unmeasured historical performance data of the data value supplemented by a plurality of all items measured historical performance data and the complementary step, obtaining the degree of similarity that represents the difference between the predicted target data A similarity calculation step;
Based on the plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the plurality of prediction errors obtained in the prediction error calculation step, A prediction value prediction error calculation step for obtaining a prediction error in the prediction value;
Obtaining a sum of a value obtained by substituting the prediction target data into the given output value prediction model and a prediction error in a prediction value of the prediction target data obtained in the prediction value prediction error calculation step output value prediction method according to claim 1, wherein the obtaining Bei the predicted value calculation step of obtaining the predicted value in. The prediction value prediction error calculation step includes the plurality of weights obtained in the weight calculation step and the plurality of similarities obtained in the similarity calculation step, and the plurality of prediction errors obtained in the prediction error calculation step. The output value prediction method according to claim 5, wherein a prediction error in the prediction value of the prediction target data is obtained by weighting. The output value prediction step includes:
A given output representing a relationship between an input variable and an output variable for each of the plurality of all-item measured past performance data and each of the plurality of copied partial item unmeasured past performance data each supplemented with a data value in the complementing step A model prediction value is obtained by substituting the data value of the data item corresponding to the input variable into the value prediction model, and the difference between the model prediction value and the data value of the data item corresponding to the output variable is determined as a prediction error. A prediction error calculation step obtained as :
For each of the previous SL plurality of copying some items unmeasured historical performance data of the data value supplemented by a plurality of all items measured historical performance data and the complementary step, obtaining the degree of similarity that represents the difference between the predicted target data A similarity calculation step;
For each of the plurality of copying some items unmeasured historical performance data of the data value supplemented respectively the complementary step, the influence degree calculation step of obtaining the degree of influence representing the weight of the unmeasured data item,
The plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, the plurality of influences obtained in the influence calculation step, and obtained in the prediction error calculation step Based on the plurality of prediction errors, a prediction value prediction error calculation step for obtaining a prediction error in a prediction value of the prediction target data;
Obtaining a sum of a value obtained by substituting the prediction target data into the given output value prediction model and a prediction error in a prediction value of the prediction target data obtained in the prediction value prediction error calculation step output value prediction method according to claim 1, wherein the obtaining Bei the predicted value calculation step of obtaining the predicted value in. The output value prediction step includes:
A given output representing a relationship between an input variable and an output variable for each of the plurality of all-item measured past performance data and each of the plurality of copied partial item unmeasured past performance data each supplemented with a data value in the complementing step A model prediction value is obtained by substituting the data value of the data item corresponding to the input variable into the value prediction model, and the difference between the model prediction value and the data value of the data item corresponding to the output variable is determined as a prediction error. A prediction error calculation step obtained as :
For each of the previous SL plurality of copying some items unmeasured historical performance data of the data value supplemented by a plurality of all items measured historical performance data and the complementary step, obtaining the degree of similarity that represents the difference between the predicted target data A similarity calculation step;
Based on the plurality of weights obtained in the weight calculation step, the plurality of similarities obtained in the similarity calculation step, and the plurality of prediction errors obtained in the prediction error calculation step, A prediction value prediction error calculation step for obtaining a prediction error in the prediction value;
Obtaining a sum of a value obtained by substituting the prediction target data into the given output value prediction model and a prediction error in a prediction value of the prediction target data obtained in the prediction value prediction error calculation step A predicted value calculation step for obtaining the predicted value at
The weight calculation step sets each weight of the plurality of all-item measured past performance data as a predetermined constant, and supplements each weight of the plurality of copy partial item unmeasured past performance data with the unmeasured data item. If the predetermined condition for determining the power value is not satisfied, it is set to 0. If the predetermined condition is satisfied, the predetermined constant is set to the copy partial item not satisfying the predetermined condition. The output value prediction method according to claim 1, wherein the output value is divided by the number of past measurement data. 9. A separation step of dividing a plurality of past performance data having a plurality of data items into the partial unmeasured past performance data and the all-item measurement past performance data. The output value prediction method according to any one of the above. In the past performance data, some item unmeasured past performance data including unmeasured data items having no data value in the plurality of data items in the plurality of past performance data including a plurality of data items. A plurality of all item measurement data having a data value in all of the plurality of data items are copied in accordance with the number of past performance data, and the copied part of the past measurement data is not measured. For each past performance data, a complementing unit that supplements the data value of the unmeasured data item with the data value of the data item corresponding to the unmeasured data item in the plurality of all-item measurement past performance data;
A weight calculation unit for obtaining a weight for each of the plurality of all item measurement past performance data and each of the plurality of copy partial item unmeasured past performance data each supplemented with a data value in the complementing step;
Wherein said plurality of weights calculated by the weight calculator, the plurality of all items measured historical performance data, based on the data value to a plurality of copying and some items unmeasured historical performance data supplemented respectively by the completion unit, An output value prediction unit that predicts an output value for prediction target data to be predicted and obtains a predicted value ;
The weight calculation unit sets each weight of the plurality of all item measured past performance data as a predetermined constant, and sets each weight of the plurality of copy partial item unmeasured past performance data as the plurality of predetermined constants. output value predicting device comprising a value and to Rukoto divided by the total entry number of measurements of the past record data. A computer comprising a storage unit for storing a plurality of historical performance data and the prediction target data comprising a plurality of data items, in order to predict the output value,
Said read plurality of the plurality comprising data item part item unmeasured historical performance data including the unmeasured data items that do not have a data value in said plurality of data items in the past record data from the storage unit, the Each of the plurality of data items in the past performance data is copied to the number corresponding to the number of all the item measurement past performance data having data values, and each of the copied items of the partial item unmeasured past performance data is copied. Complement means for supplementing the data value of the unmeasured data item with the data value of the data item corresponding to the unmeasured data item in the plurality of all-item measured past performance data for each of the copied partial item unmeasured past performance data When,
Weight calculation means for obtaining a weight for each of the plurality of all item measurement past performance data and each of the plurality of copy partial item unmeasured past performance data each supplemented with a data value by the complementing means;
Wherein a plurality of weights calculated by the weight calculation means, wherein a plurality of all items measured historical performance data, based on the data value to a plurality of copying and some items unmeasured historical performance data supplemented respectively the complementary step, an output value prediction program to function as an output value predicting means for obtaining a predicted value by predicting an output value for the predicted target data to be predicted,
The weight calculating means sets each weight of the plurality of all-item measured past performance data to a predetermined constant, and sets each weight of the plurality of copy partial item unmeasured past performance data to the plurality of predetermined constants. output value predicting program for all items measured values and to Rukoto divided by the number of historical performance data and features.

Images