JP2010231447A - Output value prediction method and device, and program for the method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an output value prediction method, an output value prediction device and an output value prediction program, for evaluating a factor of an influence on variation of a prediction value and determining the variation of the prediction value. <P>SOLUTION: In the output value prediction method, when calculating the variation of an output value y of prediction target data X<SB>0</SB>based on a similarity w<SB>j</SB>between the prediction target data X<SB>0</SB>and past actual data (X, y), and the past actual data (X, y), a degree of influence of the factor x<SB>i</SB>on the variation in a predetermined output y is calculated as an A-th weight a<SB>i</SB>for the factor x<SB>i</SB>, a predetermined distance d<SB>j</SB>between the prediction target data X<SB>0</SB>and the past actual data (X, y) is calculated by use of the calculated A-th weight a<SB>i</SB>, and the similarity w<SB>j</SB>is calculated based on the calculated predetermined distance d<SB>j</SB>. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an output value prediction technique for predicting an output value from given data, and in particular, an output value prediction technique for obtaining a variation in a predicted output value in consideration of the magnitude of variation in an output value given by an input. About.

  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.

  Such prediction is generally performed by analyzing a factor related to a prediction target and statistically analyzing past performance data of the factor.

  For example, the steel material estimation apparatus disclosed in Patent Document 1 includes a material storage unit that accumulates material component results, operation results, and material results for each product manufactured in the past, and a product storage device from among a large number of input variables. An input variable that restricts an input variable according to the rule by using an input variable limiting rule storage unit that stores a rule for selecting an input variable having a large influence on the material, and input material component information and operation information. Defining a distance function using a weighting factor as an influence of the input value on the output value for calculating the distance between each data in the material storage means and the input value using the limited input variable. And a material estimation calculating means for extracting data close to the input value based on the distance calculated using the distance function, calculating an estimated value of the material from the extracted data, and outputting the estimated value.In the material estimation apparatus having such a configuration, it is possible to prevent an estimation error caused by the difference between the model structure and the target structure, and to improve the estimation accuracy in all areas of the input space.

  Further, for example, the result prediction apparatus disclosed in Patent Document 2 includes a result database storing past condition values and results obtained by the conditions, and a condition space defined by conditions stored in the result database. Means for calculating the influence coefficient for the result of each condition in the vicinity of the requirement condition for which the result is to be predicted, and the axis of the condition space is converted based on the obtained influence coefficient, and in the converted condition space, the result database Means for calculating the distance between the value of the past condition stored in the request condition and the requirement condition, means for calculating the similarity between the value of each condition and the requirement condition based on the obtained distance, and Means for creating a prediction formula in the vicinity of the required condition based on the obtained similarity, and means for calculating a result for the required condition based on the obtained prediction formula. In the result prediction apparatus having such a configuration, even if the target is a complex / non-linear target whose influence on the result of each condition changes depending on the position of the condition space, any required condition can be set without inputting a special rule. Can be predicted with high accuracy.

  Also, for example, the converter blowing control method disclosed in Patent Document 3 is a molten steel at the end of blowing of each charge in each charge in which blowing is performed in a state where pig iron and various auxiliary materials are charged into the converter at least. It is a converter blowing control method for making the temperature coincide with the end point target temperature, and at least the amount, temperature, and component of the pig iron in the newly implemented charge, the input amount of various auxiliary materials, the end point target temperature, and the end point target component A new blowing vector is defined as a new blowing vector, and the actual heat margin is calculated by converting the actual blowing rate of each heating charge and the actual amount of heat-increasing material and coolant introduced in the past. The blowing condition results for each charge stored in the blowing performance database storing the above are defined as actual blowing vectors, and a place similar to the new blowing vector from the plurality of actual blowing vectors. Select a number of actual blowing vectors, create an approximation model for estimating the thermal margin of the newly implemented charge from each blowing condition and each actual thermal margin of the selected predetermined number of actual blowing vectors, Using the created approximate model, the thermal margin of the newly implemented charge is estimated, and based on the estimated thermal margin, the molten steel temperature at the end of the newly implemented charge blowing is set as the end point target temperature. It determines the amount of coolant or heating material to be matched. In the converter blowing control method with such a configuration, by defining the entire charge item newly implemented this time as one vector, a truly similar actual charge is selected from a large number of past actual charges. It is possible to calculate a high-accuracy thermal margin from this actual charge, and to ensure that the actual end point temperature matches the end point target temperature.

Japanese Patent No. 393441 JP 2004-355189 A Japanese Patent No. 3912215

  By the way, in the technologies disclosed in Patent Document 1 to Patent Document 3, all predict the predicted value from one point of data. For this reason, it is good if this predicted value is correct, but if this predicted value deviates from the true value, the operations and judgments performed based on this predicted value will be incorrect and appropriate. Can not get the correct output value.

  In particular, depending on the direction of deviation of the true value with respect to the predicted value, that is, the true value is likely to shift upward relative to the predicted value, or the true value is likely to shift downward relative to the predicted value. Therefore, when the operation or determination performed based on the predicted value is different, it is difficult to perform an appropriate operation or determination only with the predicted value. For example, in the manufacturing process of steel products, in the case where it is sufficient if it is not more than a certain standard value, such as impurities in molten steel, the impurity may be added if there is a high possibility that the true value is shifted upward from the predicted value. While it is necessary to perform an operation for removing, if there is a high possibility that the true value is shifted downward with respect to the predicted value, it is not necessary to perform the operation for removing impurities. For example, in the case of a steel product manufacturing process, in the case of the molten steel temperature in the ladle transported from the molten steel processing facility to the continuous casting facility, if the true value is likely to shift downward relative to the predicted temperature, In addition, since there is an operational risk such as casting stoppage due to solidification of the molten steel, the operating conditions for avoiding a decrease in the molten steel temperature are selected, while the true value is likely to shift upward with respect to the predicted temperature. In such a case, so-called breakout in continuous casting is likely to occur, so that the casting speed is adjusted.

  Moreover, in the said patent document 1, not only a predicted value but the prediction error is calculated. The reliability of the predicted value can be understood from this prediction error. However, if the operation or judgment performed based on the predicted value differs depending on the direction of deviation of the true value from the predicted value, an appropriate value depends on the predicted value and the predicted error. It is difficult to make operations and judgments.

  Therefore, when predicting the output value from the given data, it is desirable to obtain the variation of the predicted value in order to more appropriately perform the operation and determination based on the predicted value.

Here, the output value generally depends on a plurality of factors, but each factor does not necessarily contribute equally to variations in the output value. That is, the degree of influence of each factor on the output value is not uniform. More specifically, an example, the output value y is in relation to the primary factor x _1, depending on the value x _i1 factors x ₁ Notwithstanding any other factors x excluding the factors x ₁ output one variation of the value y is large, the output value y is in relation to other factors x _2, factors x ₂ Notwithstanding any other factors x excluding the factor x ₂ value x of the output value in accordance with _j2 Variation may be small. In such cases, when obtaining the variation of the output value y, it should be emphasized factors x ₁ than factor x _2. That is, in the case of obtaining the variation of the predicted value based on the similarity between the predicted target data and the past actual data, when calculating the similarity, it should be emphasized factors x ₁ than factor x _2.

  In the techniques disclosed in Patent Document 1 to Patent Document 3, when the predicted value is obtained, there is no disclosure or suggestion about the variation in the predicted value. There is no suggestion about evaluation.

  Further, although the Patent Document 2 describes an influence coefficient, this influence coefficient is for the result of each condition in the vicinity of the requirement condition for which the result is to be predicted. It is a coefficient for each condition in a line format that approximates the relationship with the result. Therefore, the influence coefficient in Patent Document 2 does not relate to the dispersion of results, that is, the dispersion of predicted values.

  The present invention has been made in view of the above-described circumstances, and an object of the present invention is to provide an output value prediction method, an output value, and an output value prediction method that can evaluate the influence on the variation in the predicted value and obtain the variation in the predicted value. To provide a prediction device and an output value prediction 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, the output value prediction method according to one aspect of the present invention includes prediction target data to be predicted, and a plurality of past performance data acquired in the past that includes a predetermined output and a quantifiable factor related to the output. A first calculation step of calculating a similarity for each of the plurality of past performance data based on a factor of the prediction target data and a factor of the past performance data; and a plurality of similarities calculated by the first calculation step And a second calculation step of calculating a variation in the output value of the prediction target data based on the past performance data, wherein the first calculation step includes a predetermined distance between the prediction target data and the past performance data. Calculating the distance for each of the plurality of past performance data based on the factors of the prediction target data and the past performance data A similarity calculation step of calculating the similarity between the prediction target data and the past performance data for each of the plurality of past performance data based on the distance calculated in the distance calculation step, The distance calculating step includes a weight calculating step for calculating the factor with the degree that the factor contributes to the magnitude of variation in the predetermined output as the Ath weight, and uses the Ath weight calculated in the weight calculating step. And calculating the predetermined distance. An output value prediction apparatus according to another aspect of the present invention includes a measured data storage unit that stores a plurality of past performance data acquired in the past, which includes a predetermined output and a factor that can be quantified related to the output; A first calculator that calculates the degree of similarity between the prediction target data to be predicted and the past performance data for each of the plurality of past performance data based on the factor of the prediction target data and the factor of the past performance data; A second calculation unit that calculates variations in output values of the prediction target data based on a plurality of similarities calculated by the first calculation unit and the past performance data, and the first calculation unit includes the prediction The predetermined distance between the target data and the past performance data is calculated based on the factors of the prediction target data and the past performance data. A distance calculation unit that calculates the similarity, and the similarity that calculates the similarity between the prediction target data and the past performance data for each of the plurality of past performance data based on the distance calculated by the distance calculation unit A degree calculating unit, and the distance calculating unit includes a weight calculating unit that calculates the factor as an A-weight to which the factor contributes to the magnitude of variation in the predetermined output. The predetermined distance is calculated using the calculated Ath weight. An output value prediction program according to another aspect of the present invention is an output value prediction program for causing a computer to execute the prediction target data to be predicted, a predetermined output, and a numerical value related to the output. A first calculation for calculating a similarity between a plurality of past performance data acquired in the past based on a factor of the prediction target data and a factor of the past performance data. And a second calculation step of calculating a variation in the output value of the prediction target data based on the plurality of similarities calculated in the first calculation step and the past performance data, the first calculation step Is based on a factor of the prediction target data and a factor of the past performance data, based on a predetermined distance between the prediction target data and the past performance data. A distance calculation step for calculating each of the plurality of past performance data, and the similarity between the prediction target data and the past performance data based on the distance calculated in the distance calculation step. A similarity calculation step for calculating each of past performance data, wherein the distance calculation step calculates a weight for calculating the factor with the degree that the factor contributes to the magnitude of variation in the predetermined output as an Ath weight. A step, wherein the predetermined distance is calculated using the A-th weight calculated in the weight calculation step.

  In the output value prediction method, the output value prediction apparatus, and the output value prediction program having such a configuration, the output value variation of the prediction target data is calculated based on the similarity between the prediction target data and the past performance data and the past performance data. In this case, the degree to which the factor contributes to the magnitude of the variation in the predetermined output is calculated as the Ath weight for the factor, and the calculated Ath weight is used to calculate the predetermined target data and the past performance data. A distance is calculated, and the similarity is calculated based on the calculated predetermined distance. Therefore, the output value predicting method, the output value predicting apparatus, and the output value predicting program having such a configuration can obtain the variation in the predicted value by evaluating the influence on the variation in the predicted value. For this reason, in the output value prediction method, the output value prediction device, and the output value prediction program having such a configuration, the similarity between the prediction target data and the past performance data is more appropriately evaluated, and the accuracy of the prediction value variation is improved. To do.

  Here, the degree to which the factor contributes to the magnitude of the variation in the predetermined output is large when the factor greatly contributes to the magnitude of the variation in the predetermined output, and the factor is the magnitude of the variation in the predetermined output. If it does not contribute much to, it will be a small value. That is, when the range of variation in the predetermined output (range of the range) is relatively dependent on the value of the factor, the factor greatly contributes to the magnitude of variation in the predetermined output, and the factor is the predetermined output. The magnitude of the influence on the magnitude of variation in the output is large, and if the range of variation in the predetermined output (range) is relatively independent of the factor value, the factor is the magnitude of variation in the predetermined output. In this case, the influence of the factor on the magnitude of the variation in the predetermined output is a small value. In other words, the magnitude of the A-th weight is determined according to the magnitude of the influence exerted on the predetermined output variation by the factor.

  In the above-described output value prediction method, the weight calculation step may be performed when the magnitude of the variation in the predetermined output is a first output variable and the variable related to the factor is a first input variable. Generating a first model representing a relationship between the first input variable and the first output variable based on the past performance data, and calculating the A weight based on the first model.

  According to this configuration, it is possible to calculate the Ath weight with higher accuracy by generating the first model.

  The first model is represented by, for example, a function equation having two variables or three or more variables (a regression equation such as a single regression equation or a multiple regression equation), and the first input variable and the first output variable. Is obtained by regression calculation. Examples of the regression calculation include a least square method and a partial least square method.

  Further, in the output value prediction method described above, the weight calculating step calculates a B-weight according to the number of the plurality of past performance data, and generates the first model using the B-weight. Features.

  According to this configuration, the B-th weight corresponding to the number of past past record data is calculated, and the first model is generated. Therefore, according to this configuration, it is possible to reduce the influence of the error included in the factor on the A-th weight, to generate the first model with higher accuracy, and as a result, to improve the first A with higher accuracy. The weight can be calculated.

  In the above-described output value prediction methods, the weight calculation step generates the first model based on residual past performance data obtained by removing predetermined past performance data from the plurality of past performance data. And

  According to this configuration, the first model is generated based on the remaining past performance data obtained by removing predetermined past performance data from the plurality of past performance data. For this reason, according to this configuration, it is possible to increase resistance to abnormal data.

  In the above-described output value prediction method, the distance calculating step determines whether the factor contributes to the magnitude of variation in the predetermined output and the degree to which the factor contributes to the predetermined output. A weight calculation step for calculating the factor as a weight is provided, and the predetermined distance is calculated using the A-th weight calculated in the weight calculation step.

According to this configuration, the A-th weight a _i can be obtained in consideration of the influence of the factor on the variation in the predetermined output and the influence of the factor on the value of the predetermined output. The accuracy of the variation of the is improved.

  In these output value prediction methods, the second calculation step uses the input variable when the predetermined output is an output variable and a part or all of the factor is an input variable. When a second model representing a relationship between an output variable and the input variable is generated, a value obtained by giving an input value of the input variable to the second model and the output corresponding to the input value of the input variable A parameter calculating step of calculating an error parameter that is a difference from an output value of the variable for each of the plurality of past actual data based on an input variable and an output variable of the past actual data; and the input variable and the error parameter Generating a third model representing a relationship between the output variable and the input variable, and a factor corresponding to the input variable among the factors of the prediction target data A predicted value calculation step of calculating each of the plurality of error parameters calculated by the parameter calculation step using the output value of the prediction target data as a predicted value by giving a value and a value of the error parameter to the third model; A variation calculating step of calculating variations in output values of the prediction target data based on the plurality of similarities calculated in the similarity calculating step and the plurality of predicted values calculated in the predicted value calculating step. It is characterized by that.

Even if the predetermined output is predicted based on some or all of the possible factors related to the predetermined output, such as fluctuations, disturbances, factors that cannot be clarified at the present time, etc. An error α exists between the predicted value and the true value due to an uncertain element such as or an uncertain element such as a modeling error. That is, the predetermined output y is expressed by the second model; y = f (Z, Θ) according to a part Z or all Z of the factor X that can be assumed and is numerically related to the predetermined output y. Even if it is modeled by, an error α exists between the predicted value and the true value due to an uncertain element that cannot express the output y only by Z and Θ. Here, Θ is a coefficient of Z and includes a term of Z ⁰ (constant term).

  This uncertain factor is a cause of variation. In the output value prediction method having the above-described configuration, the error α related to the uncertain factor is set as the error parameter α, and the error parameter α is based on the past performance data. The third model is calculated for each of the plurality of past performance data, and the error parameter α is added, and the output value of the prediction target data is calculated for each of the plurality of error parameters as a predicted value by the third model. The And the dispersion | variation in the output value of prediction object data is calculated based on several similarity and several predicted value.

  Therefore, in the output value prediction method having the above-described configuration, the variation in the predicted value is calculated with higher accuracy. As a result, it is possible to consider the variation in the predicted value when performing an operation or a determination based on the predicted value.

  Here, the factors that can be quantified include not only physical quantities that can be measured by a measuring instrument, but also, for example, each individual of an operation group that executes a process, each individual of equipment that is used to execute a process, and the like. Such quantification of each individual is executed by, for example, being set to 1 when involved in the process and set to 0 when not involved in the process. For example, when there are four teams A, B, C, and D, and the team A is involved, the data of the team A is 1 and the data of the other teams B to D is 0.

  In such an output value prediction method, preferably, for example, the variation is a histogram, and the variation calculation step includes calculating the similarity to the plurality of prediction values calculated by the prediction value calculation step. Corresponding to the first step that respectively corresponds to the plurality of similarities calculated by the step, the second step that divides the range including at least the plurality of predicted values into a finite number of sections, and the predicted values included in the sections And a third step of calculating the frequency of the section for each of the plurality of sections by adding all the similarities. According to this configuration, the variation is indicated by the histogram, and the appearance frequency of the predicted value can be easily known. In such an output value prediction method, preferably, for example, the variation is a probability density, and the variation calculation step further sets the scale of the frequency so that the area of the histogram becomes 1. A fourth step of adjusting. According to this configuration, the variation is indicated by the probability density, and the appearance probability of the predicted value can be easily known. The variation is not only a histogram and standard deviation, but also, for example, (average value ± (constant) × (standard deviation)), upper / lower limit width of predicted value (range of predicted value), and fixed amount from both ends in the range of predicted value It may be a range of remaining prediction values from which (a certain ratio) is removed.

  In these output value prediction methods described above, the factor includes a plurality of elements, and includes at least time as the element.

  According to this configuration, it is possible to obtain a predicted value of an output value in a process in which the output changes from time to time, and to obtain a variation in the predicted value.

  Moreover, in the above-described output value prediction methods, the predetermined output is a molten steel temperature in a ladle or in a tundish in a process from a converter steelmaking process through a molten steel treatment process to a continuous casting process. Features.

  According to this configuration, it is possible to predict the molten steel temperature in the ladle or in the tundish in the process from the converter steelmaking process through the molten steel treatment process to the continuous casting process, and to obtain a variation in this predicted value. It is possible to provide an output value prediction method that can be performed.

  Further, in the above-described output value prediction methods, the predetermined output is a molten steel component or a molten steel temperature corresponding to an integrated amount of blown-in oxygen in the converter process.

  According to this configuration, it is possible to provide an output value prediction method capable of predicting a molten steel component or a molten steel temperature according to the integrated amount of blown-in oxygen in the converter process, and obtaining variations in the predicted value. It becomes.

  In the above-described output value prediction methods, the predetermined output is a steel material temperature of the steel material according to a heating time or an integrated amount of heating heat in a steel material heating furnace process.

  According to this configuration, the steel material temperature of the steel material according to the heating time or the integrated amount of the heating heat amount in the steel heating furnace process is predicted, and the variation of the predicted value can be obtained. Provision is possible.

  In addition, the above-described output value prediction methods further include a presentation step of presenting the variation.

  According to this configuration, for example, a user such as an operator can know the variation, and when performing an operation or determination based on the predicted value, the variation in the predicted value can be taken into consideration.

  The output prediction method, the output prediction apparatus, and the output prediction program according to the present invention can obtain the variation of the predicted value by evaluating the influence on the variation of the predicted value.

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 figure which shows the data memorize | stored in an actual measurement data storage part. It is a figure for demonstrating the Euclidean distance of prediction object data and each past performance data. Given as an example to describe the method of calculating the first A weight a, a diagram showing the relationship between the first data item x ₁ and an output value y. Is a graph showing the relationship between the average value of the degree of variation sigma _y ^k and x _i1 output value y _i in the k-section. As an example, the relationship between the first data item x ₁ and an output value y, and a diagram showing the relationship between the average value of the degree of variation sigma _y ^k and x _i1 output value y _i in the k-section. As another example, the relationship between the first data item x ₁ and an output value y, and is a diagram showing the relationship between the average value of the degree of variation sigma _y ^k and x _i1 output value y _i in the k-section . As another example, a diagram showing the relationship between the average value and the weight w ¹ of the degree of variation sigma _y ^k and x _i1 output value y _i in the k-section. Given as another example to describe the method of calculating the first A weight a, a diagram showing the relationship between the first and second data items x _1, x ₂ and the output value y. It is a graph showing the relationship between the average value and the average value of x _i2 of the output value y _i of the degree of variation sigma _y ^k and x _i1 in the k-section. It is a graph showing the relationship between the average value and the average value and the B weight b of x _i2 of the degree of variation sigma _y ^k and x _i1 output value y _i in the k-section. It is a figure which shows the data memorize | stored in an intermediate data storage part. It is a figure which shows the data memorize | stored in a predicted value memory | storage part. It is a figure for demonstrating the calculation procedure of the dispersion | variation in a predicted value. It is a figure for demonstrating the method of calculating | requiring the probability density curve shown in FIG.15 (C) from the histogram shown in FIG.15 (B). It is a figure which shows the relationship between the temperature fall amount of an object, and elapsed time. It is a figure which shows the model of past performance data. It is a figure for demonstrating the calculation procedure of the dispersion | variation in a predicted value. It is a figure which shows the relationship between weighted distance and similarity. It is a figure which shows the relationship between weighted distance and similarity. It is a figure which shows the model of past performance data. It is a figure which shows the probability density in each estimated value. It is a block diagram which shows the structure of an output value prediction system.

  DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, an embodiment of the 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.

(First embodiment)
First, the configuration of the output value prediction apparatus S in the first embodiment will be described. FIG. 1 is a block diagram illustrating a configuration of an output value prediction apparatus according to the first embodiment. In FIG. 1, the output value prediction apparatus S includes an arithmetic control unit 1, an input unit 2, a presentation unit 3, and a storage unit 4.

  The input unit 2 outputs various commands such as a command for starting an output value prediction program for predicting an output value by using the method of the present invention from previously given data, and various data necessary for predicting the output value. A device that inputs to the prediction device S, such as a keyboard or a mouse. The presenting unit 3 is a device that presents (outputs) the command and data input from the input unit 2 and the output value (predicted value) predicted by the output value predicting apparatus S. For example, the presenting unit 3 includes a CRT display, LCD, A display device such as an organic EL display and a plasma display, and a printing device such as a printer.

  The storage unit 4 is functionally measured data storage that stores a plurality of past performance data acquired in the past and prediction target data for which an output value is to be predicted, which is functionally composed of a predetermined output and a quantifiable factor related to the output. Unit 41, intermediate data storage unit 42 for storing intermediate data generated in the process of predicting an output value based on past performance data from prediction target data, and prediction based on past performance data from prediction target data A prediction value storage unit 43 that stores (calculated) output values (prediction values), and a variation storage unit 44 that stores variation in prediction values, and various programs such as an output value prediction program; It is a device that stores various data such as data necessary for execution and data generated during the execution. The storage unit 4 is, for example, a volatile storage element such as a RAM (Random Access Memory) serving as a so-called working memory of the arithmetic control unit 1, a ROM (Read Only Memory), or a rewritable EEPROM (Electrically Erasable Programmable Read Only Memory). And the like, and a hard disk for storing various programs and various data.

  The arithmetic control unit 1 includes, for example, a microprocessor and its peripheral circuits, and functionally includes a distance calculation unit 11, a similarity calculation unit 12, a parameter calculation unit 13, and a predicted value calculation unit 14. And a variation calculating unit 15 for controlling the input unit 2, the presenting unit 3 and the storage unit 4 according to the function according to the control program.

  The distance calculation unit 11 functionally includes a weight calculation unit 111, and uses the A-th weight calculated by the weight calculation unit 111 to calculate a predetermined distance between the prediction target data and the past performance data of the prediction target data. Based on the factors and the factors of the past performance data, each of the plurality of past performance data is calculated.

  The weight calculation unit 111 calculates the degree of contribution of the factor to the magnitude of variation in the predetermined output as the Ath weight. More specifically, the weight calculation unit 111 is based on a plurality of past performance data when the magnitude of variation in the predetermined output is a first output variable and the variable related to the factor is a first input variable. A first model representing the relationship between the first input variable and the first output variable is generated, and the A-th weight is calculated based on the first model. This first model is represented by, for example, a function equation having two variables or three or more variables (a regression equation such as a single regression equation or a multiple regression equation), and is regressed from the first input variable and the first output variable. Calculated by calculation. Examples of the regression calculation include a least square method and a partial least square method (PLS, Partial Least Square).

  Here, the degree to which the factor contributes to the magnitude of the variation in the predetermined output is large when the factor greatly contributes to the magnitude of the variation in the predetermined output, and the factor is the magnitude of the variation in the predetermined output. If it does not contribute much to, it will be a small value. That is, when the range of variation in the predetermined output (range of the range) is relatively dependent on the value of the factor, the factor greatly contributes to the magnitude of variation in the predetermined output, and the factor is the predetermined output. The magnitude of the influence on the magnitude of variation in the output is large, and if the range of variation in the predetermined output (range) is relatively independent of the factor value, the factor is the magnitude of variation in the predetermined output. In this case, the influence of the factor on the magnitude of the variation in the predetermined output is a small value. In other words, the magnitude of the A-th weight is determined according to the magnitude of the influence exerted on the predetermined output variation by the factor.

  Note that the weight calculation unit 111 may calculate a B-th weight according to the number of past past record data, and generate the first model using the B-th weight. According to this configuration, the influence of the error included in the factor on the Ath weight can be reduced, the first model can be generated more accurately, and the Ath weight can be calculated more accurately. It becomes possible to do.

The weight calculation unit 111 sets the degree of contribution of the factor to the magnitude of variation in the predetermined output and the degree of contribution of the factor to the magnitude of the absolute value of the predetermined output for the factor A. It may be calculated. According to this configuration, the A-th weight a _i can be obtained in consideration of the influence of the factor on the variation in the predetermined output and the influence of the factor on the absolute value of the predetermined output. The accuracy of value variation is improved.

  The similarity calculation unit 12 calculates the similarity between the prediction target data and the past performance data for each of the plurality of past performance data based on the plurality of distances calculated by the distance calculation unit 11. .

The parameter calculation unit 13 uses the input variable to indicate the relationship between the output variable and the input variable when the predetermined output y is the output variable and the part Z or all Z of the factor X is the input variable. Model; y = f (Z, Θ), where Θ is a coefficient of Z and includes the term of Z ⁰ (constant term), by giving the input value of the input variable to the second model An error parameter α, which is the difference between the obtained value and the output value of the output variable corresponding to the input value of the input variable, is calculated for each of a plurality of past result data based on the input variable and output variable of the past result data It is.

  The predicted value calculation unit 14 generates a third model representing the relationship between the output variable and the input variable using the input variable and the error parameter α; y = (Z, Θ, α), and among the factors of the prediction target data For each of the plurality of error parameters α calculated by the parameter calculation unit 13 using the value of the factor corresponding to the input variable and the value of the error parameter α as the predicted value by giving the value of the error parameter α to the third model. Is to be calculated.

  The variation calculating unit 15 calculates the variation in the output value of the prediction target data based on the plurality of similarities calculated by the similarity calculating unit 12 and the plurality of predicted values calculated by the predicted value calculating unit 14. is there.

  The arithmetic control unit 1, the input unit 2, the presentation unit 3, and the storage unit 4 are connected by a bus 5 so that signals can be exchanged with each other.

  The arithmetic control unit 1, the input unit 2, the presentation unit 3, the storage unit 4, and the bus 5 can be configured by, for example, a computer, more specifically, a personal computer such as a notebook type or a desktop type.

  Note that the output value prediction apparatus S may further include an external storage unit (not shown) 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 program is installed in the storage unit 4 via the external storage unit from the recording medium on which the output value prediction program or the like is recorded. The output value prediction device S may be configured. Alternatively, the output value prediction device S may be configured such that data such as past performance data and data for predicting an output value is recorded on a recording medium via an external storage unit.

Next, the operation of the output value prediction apparatus S in the first embodiment 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 diagram illustrating data stored in the actual measurement data storage unit. FIG. 4 is a diagram for explaining the Euclidean distance between the prediction target data and each past performance data. 5 was cited as an example to describe the method of calculating the first A weight a, a diagram showing the relationship between the first data item x ₁ and an output value y. FIG. 6 is a diagram illustrating the relationship between the variation degree σ _y ^k of the output value y _i in the k-th section and the average value of x _i1 . 6A shows the relationship in a tabular form, and FIG. 6B shows the relationship in a graph. FIG. 7 shows, as an example, a relationship between the first data item x ₁ and the output value y, and a relationship between the variation degree σ _y ^k of the output value y _i in the k-th section and the average value of x _i1 . It is. FIG. 7A is a diagram showing the relationship between the first data item x ₁ and the output value y, and FIG. 7B shows the degree of variation σ _y ^k and x of the output value y _i in the k-th section. It is a figure which shows the relationship with the average value of _i1 . FIG. 8 shows, as another example, the relationship between the first data item x ₁ and the output value y, and the relationship between the variation degree σ _y ^k of the output value y _i in the k-th interval and the average value of x _i1. FIG. FIG. 8A is a diagram showing the relationship between the first data item x ₁ and the output value y, and FIG. 8B shows the degree of variation σ _y ^k and x of the output value y _i in the k-th section. It is a figure which shows the relationship with the average value of _i1 . FIG. 9 is a diagram showing the relationship between the variation value σ _y ^k and the average value of x _i1 and the weight w ¹ of the output value y _i in the k-th section as another example. FIG. 9A shows the relationship in a tabular form, and FIG. 9B shows the relationship in a graph. FIG. 10 is a diagram illustrating a relationship between the first and second data items x ₁ and x ₂ and the output value y, which is given as another example for explaining the calculation method of the A-th weight a. FIG. 11 is a diagram illustrating a relationship between the degree of variation σ _y ^k of the output value y _i in the k-th section, the average value of x _i1 , and the average value of x _i2 . FIG. 11A shows the relationship in a tabular form, and FIG. 11B shows the relationship in a graph. FIG. 12 is a diagram illustrating the relationship between the variation degree σ _y ^k of the output value y _i in the k-th section, the average value of x _i1 , the average value of x _i2 , and the _Bth weight b. FIG. 13 is a diagram illustrating data stored in the intermediate data storage unit. FIG. 14 is a diagram illustrating data stored in the predicted value storage unit. FIG. 15 is a diagram for explaining a procedure for calculating a variation in predicted values. FIG. 15A shows the relationship between the similarity w and the output predicted value y ₀ , the horizontal axis is the similarity w, and the vertical axis is the predicted value y ₀ . FIG. 15B shows a histogram of the predicted value y ₀ , the horizontal axis is the weighted frequency Fw, and the vertical axis is the predicted value y ₀ . FIG. 15C shows the probability density distribution of the predicted value y ₀ , the horizontal axis is the probability density P (y ₀ ), and the vertical axis is the predicted value y ₀ .

  For example, when an output command is received from the input unit 2 by a user operation, the output value prediction device S executes an output value prediction program. By executing this output prediction program, the calculation control unit 1 is functionally connected to the distance calculation unit 11 (including the weight calculation unit 111), the similarity calculation unit 12, the parameter calculation unit 13, the predicted value calculation unit 14, and the variation calculation unit 15. Configured. And the output value prediction apparatus S estimates an output value (predicted value) from prediction object data based on past performance data by the following operations.

In the prediction of the output value, the past result data and the prediction target data are stored in advance in the measured data storage unit 41 in the storage unit 4 of the output value prediction apparatus S, for example, in the table format (table format) shown in FIG. Yes. The actual measurement data table 51 shown in FIG. 3 includes an output field 511 for registering the actually measured output value y, and a data field 512 for registering the factor x data relating to the output value y. is provided with a record for each historical track record data (X, y), further comprises a record of the predicted target data X _0. The factor related to the output value y includes a plurality of elements (data items). For this reason, the data field 512 includes data item subfields corresponding to the number of elements. In the example shown in FIG. 3, the output value y involves at least N elements (first to Nth data items). Therefore, the data item subfield, a data item subfield 5121~512N registering each data x ₁ ~x _N first to N-th data item corresponding to each element of the source, respectively. The past performance data (X, y) is data obtained by actual measurement or the like at different times (points) in the past under different conditions in the past. In the example shown in FIG. It is composed of Predicted target data X ₀ is data x ₀ for which you want to predict the output value, for example, and the data x _0, which is measured up to the prediction time t _0, and the value x ₀ of the operation input, and operation date and time x _0, the data x _0, etc., which was prepared for the simulation.

Here, the output value prediction unit S is predicted from the data value x ₀ of the prediction target data X ₀ historical performance data (X, y) output value based on the (predicted value) y _0, the prediction value y ₀ The variation is obtained.

Incidentally, the predicted target data X _0, in order to distinguishably distinguished historical performance data (X, y) and, 0 is assigned as the first subscript (left subscript), historical performance data (X, y ), 1 to M are assigned as first subscripts to distinguish the M pieces of data in an identifiable manner. Then, the predicted target data X ₀ and historical performance data (X, y), the first to N data items that are N elements in the factors involved in the output value y to identifiably distinguished from each other, 1 ˜N are added as second subscripts (the right side of the subscripts). X ₀ = [x ₀₁ , x ₀₂ ,..., X _0N ], and X _j = [x _j1 , x _j2 ,..., X _jN ]. For example, y ₃ represents the third output value in the past performance data (X, y), and for example, x ₂₃ represents the second third data item in the past performance data (X, y). represents the value, also for example, x ₀₄ represents the value of the fourth data item in the prediction target data X _0.

In the case where such a historical record data (X, y) and the predicted target data X ₀ is stored in the measured data storage unit 41, an output value prediction device S, as shown in FIG. 2, initially, historical performance data to assess the association between each data value x of the first to N data items in the prediction target data X ₀ and each data value x of the first to N data items in X, calculating the distance between the data The distance is calculated by the distance calculation unit 11 of the control unit 1, and the calculated distance is stored in the intermediate data storage unit 42 of the storage unit 4 (S11). The distance is defined to represent the relationship between the two data x, and for example, Euclidean distance, normalized Euclidean distance, or the like is used.

More specifically, in the present embodiment, as shown in FIG. 4, the distance calculation unit 11 calculates the Euclidean distance d between the prediction target data X ₀ and the past performance data X in the data item space as each past performance data (X , Y). Since there are N data items, the data item space is an N-dimensional space in this embodiment. In the present embodiment, the Euclidean distance d is a weighted distance, and is obtained by, for example, Expression 1-1. In Expression 1-1, the weighted Euclidean distance d _j between the prediction target data X ₀ and the j-th past performance data X is the i-th data item x _ji in the j-th past performance data (X, y). The square of the difference from the i-th data item x _0i in the prediction target data X ₀ is set to the square of the weight a _i of the i-th data item (a _i ≧ 0, the A-th weight a _i , and the weight a _i regarding the distance). The product is calculated by taking the sum from the first data item to the Nth data item and obtaining the square root of the result.

Further, for example, the Euclidean distance d may be obtained by Expression 1-2. In Formula 1-2, the weighted Euclidean distance d _j between the prediction target data X ₀ and the j-th past performance data X is the i-th data item x _ji in the j-th past performance data (X, y). to the square of the difference between the i-th data item x _0i in the prediction target data X _0, sum multiplied by the absolute value of the a weight a _i of the i data item from the first data item to the N-th data item And taking the square root of the result.

Further, for example, the Euclidean distance d may be obtained by Expression 1-3. In Expression 1-3, the weighted Euclidean distance d _j between the prediction target data X ₀ and the j-th past performance data X is the i-th data item x _ji in the j-th past performance data (X, y). the difference between the i-th data item x _0i in the prediction target data X _0, the absolute value of the multiplying first a weight a _i of the i data item, by taking the sum from the first data item to the N-th data item Is calculated.

Here, the A-th weight a _i is a parameter representing the importance (degree of importance) of the i-th data item among the first to N-th data items in obtaining the Euclidean distance d, and the i-th data item x _i is about contributing to the magnitude of variation in a given output y.

It should be noted that the A-th weight a _i is determined according to the magnitude of the influence of the i-th data item x _{i on} the variation of the output value y. For example, the A-th weight a _i is set so as to increase as the data item having a large influence on the variation of the output value y and decrease as the data item having a small influence on the variation of the output value y. The

More specifically, the A-th weight a _i is calculated by the following procedure. Here, the output value y is generally a multiple correlation that depends on a plurality of factors (first to Nth data items x _{1 to} x _N ) as described above. In the case where y is a single correlation that depends on the first data item x, the calculation method of the Ath weight a _i will be described in detail below. From the generalization of the calculation method of the Ath weight a _i and the ease of explanation, In the case where the output value y is a multiple correlation that depends on the first and second data items x ₁ and x ₂ , a method of calculating the A-th weight a _i will be described below.

In explaining the calculation method of the A-th weight a _i , the relationship between the first data item x ₁ and the output value y is assumed to be the relationship shown in FIG. That is, the output value y is with increasing second through N data items _x 2 ~x regardless of the value of _N depends only on the value of the first data item _{x 1,} the first data item _{x 1} values _{x i1} The variation gradually increases and takes a fan-like distribution.

First, the first, as a first step (step), the first data item x ₁ is divided into sections of a plurality K, historical performance data in each section (k = 1 to K, the first section, second K section) A variation degree σ of the output value y of (X, y) is obtained.

In the first data item x ₁ , a set S ₁ ^k of past performance data (X, y) included in the k-th k-th section (kεK) is expressed by Expression 2.

Here, {(y _i , x _i1 )} is the past performance data (X, y), although the first data item x _i1 and the output value y _i are expressed in reverse (i = 1-M). As shown in FIG. 5, x ₁ ^k is a starting value of the range of the k-th section in the first data item x ₁ , and x ₁ ^k + Δx ₁ is the k-th section of the first data item x ₁ . The end value of the range. Δx ₁ is the width of the range of the k-th section in the first data item x ₁ . Note that the subscript _{j of} the set S _j ^k represents a data item, the superscript k represents a section, and therefore, the set S ₁ ^k has a value x _{i1 of} the first data item x ₁ in the k-th section. historical performance data _{(y i, x i1)} to be within the scope of the ^{_{(x 1 k ≦ x i1 ≦}} x 1 k + △ x 1) is the set of. In the inequality indicating the range of the k-th section, equal signs are included at both ends, but the equal signs may not be included at one end.

Section, the range may be set so as not to overlap with the adjacent section, but the error included in the first data item x ₁ is absorbed, the change in the degree of variation sigma _y ^k is smooth with respect to a change in the interval Therefore, it is preferable that the range is set so as to overlap with an adjacent section. The size of the portion (overlapping portion) overlapping is set appropriately in accordance with the first number of data items x _1, the size of the overlap, is set narrower as the number of the first data item x ₁ is large, the number of the first data item x ₁ is more widely set small.

As the variation degree σ _y ^k , an arbitrary index indicating the degree of variation is adopted. For example, the variation degree σ _y ^k is an output value of past performance data {(y _i , x _i1 )} included in the set S ₁ ^k. and a standard deviation of y _i, historical performance data included in the set ^{_{S 1 k {(y i,}} x i1)} on the lower limit range of the output value _{y i} of the (= upper limit - minimum value) or the like is employed. In addition, in order to strengthen tolerance against abnormal data (in order to suppress a decrease in accuracy of the degree of variation σ _y ^k due to abnormal data), the past performance data {(y _i , x _i1 )} included in the set S ₁ ^k After removing both ends of the distribution in the output value y _i by a certain amount (a certain width, a certain ratio), the standard deviation σ _y ^k may be obtained, and the past performance data included in the set S ₁ ^k {( After removing both ends of the distribution of the output value y _i of y _i , x _i1 )} by a certain amount, the upper and lower limit width σ _y ^k may be obtained.

Subsequently, secondly, as a second step, the past performance data {(y _i , x _i1 )} included in the set S ₁ ^k in each section (k = 1 to K, the first section to the Kth section)} The average value of the first data item x _i1 is obtained, and this value is set to μ _x1 ^k . In order to reduce the calculation amount (information processing amount), x ₁ ^k + (Δx ₁ ) / 2 may be approximately set to μ _x1 ^k .

Here, when obtaining the degree of variation σ _y ^k , as described above, both ends of the distribution in the output value y _i of the past performance data {(y _i , x _i1 )} included in the set S ₁ ^k are fixed amounts. when removing portions, the mu _x1 the degree of variations in determining the ^k sigma _y historical performance data that was removed when calculating the ^{_{k {(y i, x i1}} )} Following removal of, the mu _x1 ^k is Desired.

Preferably, in the first data item _{x 1,} between the start edge value _x ^{1 k} and start edge value _x ^{1 k + 1} of the (k + 1) th interval of the k section, starting values _x ^{1 k + 1} of the (k + 1) th section is the k The relationship is larger than the start value x ₁ ^{k of the} section, and the deviation amount from the adjacent sections (the deviation amount of the start value between the adjacent sections, the step size) Δk ₁ (= x ₁ ^{k + 1} −x ₁ ^k ) is , Expression 3 is established. That is, a relation between the width △ x ₁ in the range of displacement amount △ k ₁ and the section between adjacent sections, the width △ x ₁ in the range of the interval is greater than the shift amount △ k ₁ of the adjacent segment relationship It is said.

By making each of these relationships, a relatively large number of sets of the variation degree σ _y ^k and the average value μ _x1 ^k are generated (the value of K becomes a relatively large value), and the A-th weight a ₁ is set. The amount of data to be obtained is relatively large. Therefore, it becomes possible to make it difficult to be affected by data errors, and resistance to abnormal data is enhanced. Therefore, stably, it is possible to obtain the first A weighting a _1.

In addition, since at least two pieces of data are necessary to obtain the degree of variation σ _y ^k , there is no past performance data {(y _i , x _i1 )} in the k-th section and the k-th case. historical performance data _{{(y i, x i1)} } in the interval when is one, historical performance data _{{(y i, x i1)} } until the predetermined number of 2 or more, the k-th segment in case the end value of the k-th interval x _{1 k} ⁺ △ x ₁ is sequentially incremented by the deviation amount △ k ₁ of the adjacent section (updated). Alternatively, the width Δx ₁ of the section range may be reset so that the past performance data {(y _i , x _i1 )} becomes a predetermined number of 2 or more. As described above, the same applies when both ends of the distribution of the output value y _i of the past performance data {(y _i , x _i1 )} included in the set S ₁ ^k are removed by a certain amount.

As described above, when the variation degree σ _y and the average value μ _x1 thereof are obtained for each of the first to Kth sections, for example, as shown in FIG. 6, for example, the average value μ _x1 and the variation degree σ _y The relationship is obtained. The relationship between the average value μ _x1 and the variation degree σ _y represents a change in the distribution of the output value y (the width of the distribution and the magnitude of the variation) according to the value x _i1 of the first data item x ₁ .

Therefore, thirdly, as a third step, when the average value μ _x1 is the so-called x axis and the variation degree σ _y is the so-called y axis, the relationship between the average value μ _x1 and the variation degree σ _y gradient is obtained in this the obtained slope is the first a weighting a _1.

By executing such a first step to third step, the A weighting a ₁ representing the degree of influence on the variation of the output value y by the first data item x ₁ is obtained.

As an example, when the output value y for each value x _i1 of the first data item x ₁ is distributed as shown in FIG. 7A, the relationship between the average value μ _x1 and the variation degree σ _y is By executing the first to third steps, the substantially linear relationship shown in FIG. 7B is obtained. More specifically, as each value x _{i1 of} the first data item x ₁ increases from 0 to 1, the output value y spreads in a fan shape around the coordinates (0, 0.5), and x _i1 = At 0.8, it expands from about -0.1 to 1.1. When executing the first step to the third step, the deviation amount Δk ₁ = 0.04 from the adjacent section and the width Δx ₁ = 0.2 of the section range were set. Then, a function formula (regression formula, single regression formula, an example of the first model) representing the relationship between the average value μ _x1 and the variation degree σ _y is calculated by the least square method, and its slope is calculated as 0.2335, The intercept was calculated to be 0.006589. Therefore, in the case shown in FIG. 7, the A-th weight a ₁ is 0.2335.

As another example, when the output value y for each value x _i1 of the first data item x ₁ is distributed as shown in FIG. 8A, the average value μ _x1 and the variation degree σ _y With respect to the relationship, the substantially linear relationship shown in FIG. 8B is obtained by executing the first to third steps. More specifically, as each value x _{i1 of} the first data item x ₁ increases from 0 to 1, the output value y spreads in a fan shape around the coordinates (0, 0.5), and x _i1 = At 0.8, it expands from about 0 to 1. The example shown in FIG. 8A has a fan-shaped spread smaller than the example shown in FIG. When executing the first step to the third step, the deviation amount Δk ₁ = 0.04 from the adjacent section and the width Δx ₁ = 0.2 of the section range were set. A functional expression representing the relationship between the average value μ _x1 and the variation degree σ _y was calculated by the least square method, the slope was calculated to be 0.2013, and the intercept was calculated to be 0.001987. Therefore, in the case shown in FIG. 8, the A-th weight a ₁ is 0.2013.

Here, as can be seen from FIG. 8B, the result (◯) shown in FIG. 8B obtained from the data in FIG. 8A and the approximation of the functional expression obtained by the least square method. straight and is at both ends of the range of the first data item x ₁ value, it shown in FIG. 8 (B) compared in FIG. 7 (B) are deviated. This is a relatively large part of the value of the average value mu _x1, the number of data used in determining the relationship between the average value mu _x1 and degree of variation sigma _y is relatively small, the average value mu _x1 is relatively reliable Is affected by low data.

For this reason, when obtaining a functional expression representing the relationship between the average value μ _x1 and the degree of variation σ _y , the average value x _i1 of the first data item x ₁ is averaged in each k-th interval in order to suppress the influence. Even if the weight w ^k (the B weight b ^k , the weight b ^k related to the A weight a ₁ ) corresponding to the number ^nk of past performance data used when obtaining the value μ _x1 and the variation degree σ _y is given. Good. The B-th weight b ^k is defined as, for example, Expression 4-1 and is the number of data ^nk . Alternatively, the B-th weight b ^k is defined as, for example, Expression 4-2, and may be a value obtained by subtracting 1 from the number of data ^nk . Alternatively, the B-th weight b ^k is defined as, for example, Expression 4-3, and may be a square root of a value obtained by subtracting 1 from the number of data n ^k . Note that the number of data n ^k is the value after removal when both ends of the distribution in the output value y _i of the past performance data {(y _i , x _i1 )} included in the set S ₁ ^k are removed by a certain amount. This is the number of data.

When obtaining such a B-th weight b ^k , for example, as shown in FIG. 9 as an example, a relationship among the average value μ _x1 , the variation degree σ _y, and the B-th weight b ^k is obtained. Here, Formula 4-2 was used as the B weight ^{b k.} When obtaining such a B-th weight b ^k , a regression equation representing the relationship between the average value μ _x1 and the degree of variation σ _y is obtained by the weighted least square method shown in Equation 5.

  Here, capital pie Π, capital gamma Γ, and capital W are as shown in Equation 6, respectively.

In this case, the slope in the relationship between the average value μ _x1 and the degree of variation σ _y was calculated as 0.248, and the intercept was calculated as 0.0005751. As can be seen from FIG. 9B, the approximate straight line of this regression equation is better fitted to each data as compared to FIG. 8B. Accordingly, in the case shown in FIG. 8A, the A-th weight a ₁ is 0.248.

Next, in the case where the output value y is a multiple correlation that depends on the first and second data items x ₁ and x ₂ , a method of calculating the A-th weight a _i will be described below.

In explaining the calculation method of the A-th weight a _i , the relationship between the first and second data items x ₁ and x ₂ and the output value y is assumed to be the relationship shown in FIG. That is, the output value y depends only on the values of the first and second data items x ₁ and x ₂ regardless of the values of the _{third to} Nth data items x _{3 to} x _N , and the first and second data items. The distribution is such that the variation gradually increases as x ₁ and x ₂ increase.

In the case of this multiple correlation as well, in the same way as in the case of the single correlation, the relationship between the average value μ _x1 and the average value μ _x2 and the variation degree σ _y is obtained by executing the first to third steps.

Here, in the first step, included in the k-th k-th section (kεK) in the first data item x ₁ corresponding to the set S ₁ ^k in the case of the single correlation represented by Equation 2, A set S ₁ ^{(k, j)} of past performance data {(y _i , x _i1 , x _i2 )} included in the j-th j-th section (j∈J) in the second data item x ₂ is Expressed by Equation 7.

Further, in the second step, Equation 3 in the case of a single correlation becomes Equation 8 in the case of this multiple correlation. That relates to the first data item x _1, between the width △ x ₁ in the range of deviation between adjacent sections △ k ₁ and the interval, width △ x ₁ range of the interval is displaced to an adjacent section The relationship between the interval Δk ₂ and the width Δx ₂ of the interval between the adjacent interval and the width Δx ₂ of the interval for the second data item x ₂ is greater than the amount Δk _1. Δx ₂ is set to be larger than the deviation amount Δk ₂ from the adjacent section.

When the variation degree σ _y and the average value μ _x1 and the average value μ _x2 are obtained for each of the first to Kth sections and each of the first to Jth sections in this way, for example, FIG. As shown in FIG. 11, the relationship between the average value μ _x1 and the average value μ _x2 and the variation degree σ _y is obtained. Here, in the case of _obtaining the relationship between the average value μ _x1 and the average value μ _x2 and the variation degree σ _y , the second data item x _j2 is fixed to the jth interval, and the first data item x _i1 is determined from the first interval. While scanning up to the K-th section, the variation degree σ _{y of the} section is obtained, and then the second data item x _j2 is fixed to the j + 1-th section, and the first data item x _i1 is changed from the first section to the K-th section. The variation degree σ _{y of the} section is obtained while scanning up to. Such a procedure is performed from the first section to the _Jth section of the second data item _xj2 . Thus, by scanning the section of the other data item while fixing the section of one data item, the relationship between the average value μ _x1 and the average value μ _x2 and the variation degree σ _y is obtained. The relationship between the average value μ _x1 and the average value μ _x2 and the variation degree σ _y is the distribution of the output value y (distribution of the distribution) according to the values x _i1 and x _i2 of the first and second data items x ₁ and x ₂ . It represents the change in width and magnitude of variation. In the example shown in FIG. 11, the variation degree σ _y of the output value y increases as the value x _{i1 of} the first data item x ₁ increases, but the output x _j2 increases as the value x _{j2 of} the second data item x ₂ increases. The variation degree σ _y of the value y does not change much.

Therefore, thirdly, as in the third step of simple correlation, the average value μ _x1 is the so-called x axis, the average value μ _x2 is the so-called y axis, and the variation degree σ _y is the so-called z axis. In this case, the variation degree σ _y is expressed by Equation 9, and the slopes a ₁ and a ₂ in the relationship between the average value μ _x1 and the average value μ _x2 and the dispersion degree σ _y are obtained, and the obtained slopes a _{1, a 2} are the first a weighting _a 1, _{a 2} of the first and second data, respectively item _x 1, _{x 2.}

By thus if the multiple correlation also executes a first step to third step as in the case of a single correlation, the A weighting a representative of a degree of influence on the variation of the output value y by the first data item x _{1 1} is determined, the a weighting a ₂ representing the degree of influence on the variation of the output value y of the second data item x ₂ is calculated.

Here, as in the case of the single correlation, as shown in FIG. 12, a B-th weight b corresponding to the number n of past performance data is further calculated, and the average value μ _x1 and the B-th weight b are used. The relationship between the average value μ _x2 and the degree of variation σ _y may be obtained.

In the case of obtaining such a B-th weight b ^{(k, j)} , the average value μ _x1 ^{(k, j)} and the average value μ _x2 ^{(k, j)} and the degree of variation are obtained by the weighted least square method shown in Expression 10. A regression equation representing the relationship with σ _y ^{(k, j)} is obtained.

  Here, capital pie Π, capital gamma Γ, and capital W are as shown in Equation 11, respectively.

Here, the A-th weight a _i is obtained by considering not only the influence of the i-th data item x _{i on} the variation of the output y but also the influence of the i-th data item x _{i on} the value of the output y. In this case, not only the equation 9 but also the parameters α ₁ , α ₂ , and β of the equation 12 are obtained by multiple regression, and the Ath weight a _i is defined as, for example, the equation 13-1, The square root of the sum of the square of a _{i and} the square of the parameter α _i may be obtained. Alternatively, the A-th weight a _i may be defined as, for example, Expression 13-2, and may be obtained by the sum of the absolute value of the parameter a _{i and} the absolute value of the parameter α _i . Alternatively, the A-th weight a _i may be defined as, for example, Expression 13-3, and may be obtained as the square root of the absolute value of the multiplication of the parameter a _i and the parameter α _i . And effects whereby the variation of the output value y according to the i data item x _i, by taking into account the influence of the value of the output y by the i data item x _i, it is possible to obtain the first A weight a _i improves the accuracy of the variation of the predicted value y _0. That is, a data item x that affects the variation of the output y but does not affect the value of the output y, a data item x that does not affect the variation of the output y but affects the value of the output y, and For each of the data item x that affects the output y variation and also affects the output y value, and the data item x that does not affect the output y variation and does not affect the output y value. The A-th weight a _i can be obtained according to the magnitude of the influence.

Note that, when obtaining the relational expression (regression equation), the least square method or the weighted least square method is used, but a partial least square method may be used. As a result, even when so-called multicollinearity exists between variables, the A-th weight a _i can be obtained with high accuracy.

Then, the distance calculation unit 11 stores each distance d calculated using the Ath weight a _i calculated by the weight calculation unit 111 in the intermediate data storage unit 42 of the storage unit 4.

Returning to FIG. 2, next, the output value predicting apparatus S determines the similarity (degree of similarity) w between the two data x in order to evaluate how similar the prediction target data X ₀ is. Each of the 1st to Mth past performance data (X, y) is calculated by the similarity calculation unit 12 of the calculation control unit 1, and each calculated similarity w is stored in the intermediate data storage unit 42 of the storage unit 4. (S12). The degree of similarity w is defined, for example, such that the smaller the weighted Euclidean distance d is, the larger the degree of similarity is and a positive value is taken.

  More specifically, the similarity calculation unit 12 calculates, for example, the similarity w by Expression 14-1. Further, for example, the similarity calculation unit 12 calculates the similarity w by Expression 14-2. For example, the similarity calculation unit 12 calculates, for example, the similarity w using Expression 14-3.

Here, w _j is the similarity of the j-th past performance data X with respect to the prediction target data X ₀ , σ is a normalization parameter, specifically, d _j (j = 1 to M) The standard deviation is c, g, and r are positive real adjustment parameters. D of the superscript bar is an average value of d _j (j = 1 to N).

  Then, the similarity calculation unit 12 stores each calculated similarity w in the intermediate data storage unit 42 of the storage unit 4.

  In addition, when the upper limit value of the similarity score w and / or the lower limit value thereof is provided and the similarity score calculated by any one of the formulas 14-1 to 14-3 exceeds the upper limit value, the similarity score is If the similarity calculated by any one of the formulas 14-1 to 14-3 exceeds the lower limit value, the similarity is replaced with the lower limit value. The degree calculation unit 12 may be configured. By being configured in this way, it is possible to prevent only specific past performance data from excessively increasing in similarity or conversely decreasing. If the degree of similarity of only specific past performance data becomes excessive, if there is an error in the data measurement value, it will be pulled by that error and the wrong variation will be predicted. It will be. For this reason, as described above, setting the upper limit value has an effect of becoming stronger in error.

Further, for example, when a predetermined threshold is provided in advance and the similarity calculated by any one of the formulas 14-1 to 14-3 is equal to or less than the threshold, the similarity is set so that the similarity is replaced with 0. The degree calculation unit 12 may be configured. Alternatively, the similarities calculated by any one of the formulas 14-1 to 14-3 are arranged in ascending order, and the similarities up to a predetermined number (or predetermined ratio) set in advance from the smallest one are replaced with 0. In addition, the similarity calculation unit 12 may be configured. By being constructed as described above, when obtaining the predicted value, it is possible to prevent to consider more than necessary historical performance data X which is not much similar to the predicted target data X _0. Also, it excluded historical performance data X which is not much similar to the predicted target data X ₀ is, the arithmetic processing is not required to be described below, as a result, the arithmetic processing amount reduction (shortening of processing time) becomes possible.

  Next, the output value prediction device S calculates an error parameter α representing an uncertain element by the parameter calculation unit 13 of the arithmetic control unit 1 for each of the first to M-th past performance data (X, y), The calculated error parameters α are stored in the intermediate data storage unit 42 of the storage unit 4 (S13).

  More specifically, the parameter calculation unit 13 obtains a prediction model (third model) for predicting the output value y based on the M past performance data (X, y), and uses the obtained model. Each error parameter α is obtained for each of the first to Mth past performance data (X, y).

  This prediction model is expressed by, for example, the function f in Expression 15. In this case, the parameter calculation unit 13 obtains the coefficient Θ of the function equation 15 based on the M past performance data (X, y), and uses the obtained function equation 15 to obtain the first to M-th past Each error parameter α is obtained for each result data (X, y).

Here, Z is a factor involved in the prediction of the output value y, such as each condition of operation conditions (manufacturing conditions) and each measurement item in each process of the manufacturing process, and includes a plurality of L elements z. (Z _j = [z _j1 , z _j2 ,..., Z _jL ]). Z is constituted by, for example, part or all of data items (X _j = [x _j1 , x _j2 ,..., X _jN ]) that are a plurality of elements in the factors related to the output value y. Z may further include elements other than the data item X. Z is set in advance when the expression 15 of the function f is determined. Θ is a coefficient of the function formula 15, and is obtained by identification calculation based on M pieces of past performance data (X, y). For this identification, an error between the actual value of the output value y _j and the predicted value y ₀ such as the least square method, maximum likelihood estimation method, partial least square method, quadratic programming method, and PSO (Particle Swarm Optimization) is predetermined. A method of determining the minimum (or maximum) under the evaluation function (within a predetermined constraint condition) is used. α is an error parameter representing an indeterminate element, and represents a factor (factor of variation) in which the output y cannot be expressed only by Θ and Z. In the case where the output y is predicted using Θ and Z, corresponding to the error between the predicted value y ₀ and the actual value y.

When the output value y is predicted by a multiple regression equation, for example, Equation 16-1 can be used as the function equation 15, and the error parameter α _j in the j-th past performance data (X _j , y _j ) is used. Is given by Equation 16-2. The model expressed by Expression 16-1 is effective when an uncertain element (variation factor) exists additively.

For example, when the output value y is predicted by a multiple regression equation, the equation 15-1 can be used as the function equation 15, for example, and an error in the _jth past performance data (X _j , y _j ) The parameter α _j is given by Equation 17-2. The model expressed by Expression 17-1 is effective when an uncertain element exists in a multiplicative manner.

For example, when the output value y is predicted by a multiple regression equation, the equation 15-1 can be used as the function equation 15, for example, and an error in the _jth past performance data (X _j , y _j ) The parameter α _j is given by equation 18-2. The model expressed by Expression 18-1 is effective when there is an uncertain element in the influence coefficient of z _j1 .

In the above description, a mathematical expression representing the function f is used. However, an arithmetic program including a table representing the function f, a convergence calculation algorithm, an if-then rule, a fuzzy reasoning, a neural network, a JIT (Just in Time) model, and the like. May be used. Here, when the error parameter α _j cannot be obtained from Z _j , Θ, and y _j by back calculation, the value of the error parameter α _j is obtained by, for example, a binary search method, a carpet bombing method, or PSO (Particle Swarm Method). And an error parameter α _j such that the output value y matches y _j .

  Then, the parameter calculation unit 13 stores the calculated error parameters α in the intermediate data storage unit 42 of the storage unit 4.

The weighted Euclidean distance d, the similarity w, and the error parameter α calculated by the processes S11 to S13 are stored in the intermediate data storage unit 42 in a table format (table format) as shown in FIG. 13, for example. The The intermediate data table 52 shown in FIG. 13 is used for calculating the similarity among the output field 521 for registering the actually measured output value y and the data items corresponding to a plurality of elements in the factors related to the output value y. A data field 522 for calculating the degree of similarity for registering the data x of the selected data item, and the data item used for calculating the predicted value y ₀ among the data items corresponding to a plurality of elements in the factors related to the output value y Output prediction data field 523 for registering data x, weighted distance field 524 for registering weighted Euclidean distance d of the past performance data (X, y), similarity w of the past performance data (X, y) And the error parameter field for registering the error parameter α of the past performance data (X, y). Is configured to include the fields of Rudo 526 includes a record for each historical track record data (X, y), further comprises a record of the predicted target data X _0. The similarity calculation data field 522 includes data item subfields 5221 to 522N corresponding to the data items used for calculating the similarity. Similarly, the output prediction data field 523 includes data item subfields 5231 to 523L corresponding to each data item used for calculation of the predicted value. The output field 521 and the similarity calculation field 522 in the intermediate data table 52 shown in FIG. 13 correspond to the output field 511 and the data field 512 in the actual measurement data table 51 shown in FIG.

Then, the output value prediction device S, by the predictive value calculating unit 14 of the arithmetic and control unit 1, using the prediction model obtained in process S13, the first to the data value of the N data items in the prediction target data X ₀ Based on x _{01 to} x _0N , the predicted value y ₀ is calculated for each error parameter α _j obtained in step S 13, and the calculated predicted value y ₀ is stored in the predicted value storage unit 43 of the storage unit 4. (S14). Here, in the processing S14, the prediction model is changed to the error parameters α _j obtained in the processing S13. For example, in the case where the prediction model is expressed by Expression 15 of the function f described above, the coefficient f obtained in the process S13 and the function f changed to each of the error parameters α _j obtained in the process S13. In Expression 15, each data value x _{01 to} x _0L of the first to Lth data items used for calculation of the predicted value y ₀ among the first to Nth data items in the prediction target data X ₀ is used as Z. Accordingly, the predicted value calculation unit 14 calculates the predicted value y ₀ for each error parameter α _j obtained in step S13. The predicted value _{y 0,} each error parameter alpha _j is because it is the M, the M number of predicted value _y 01 _{~y 0M.}

  In the process S14, as in the process S13, an arithmetic program including a table representing the function f, a convergence calculation algorithm, an if-then rule, a fuzzy inference, a neural network, a JIT model, and the like may be used.

Each such prediction value y ₀₁ ~y _0M calculated by the step S14 is stored in, for example, the predicted value storing unit 43 by the table format table format as shown in FIG. 14. Predicted value data table 53 shown in FIG. 14, the predicted value y ₀ of the first to N data items in the prediction value field 531, the predicted target data X ₀ to register the predicted value y ₀ calculated by the step S14 An output prediction data field 532 for registering the data values x _{01 to} x _0L of the first to L-th data items used for the calculation, an error parameter field 533 for registering the error parameter α calculated in step S13, and , Each field of the similarity field 534 for registering the similarity w in the past performance data (X, y) used for calculation of the error parameter of the parameter field 533, and records for each error parameter α _j. I have. Note that the error parameter field 533 and the similarity field 534 in the predicted value data table 53 shown in FIG. 14 correspond to the error parameter field 526 and the similarity field 525 in the intermediate data table 52 shown in FIG. Here, each prediction value y ₀₁ ~y _0M calculated by the step S14, as shown in FIG. 14, also the similarity w _j corresponding to the predicted value y _0, the predicted value storing section to correspond to each other 43.

Then, the output value prediction device S, the variation calculation unit 15 of the arithmetic and control unit 1, with each of the prediction value _y 01 _{~y 0M} obtained in processing S14, calculates the variance of the predicted value _{y 0,} this calculation storing the dispersion of the predicted value _{y 0} that the variation storage unit 44 of the storage unit 4 (S15).

In the present embodiment, based on the rule of thumb that results in similar in similar conditions, the higher the similarity between the data x ₀ of the prediction target data X ₀ and data x _j of the j-th historical performance data X ( If the degree of similarity w _j is large), the error parameter α ₀ of the prediction target data X ₀ is considered to have high similarity. For this reason, the predicted value y ₀ of the prediction target data X ₀ is considered to be the predicted value y _0j predicted using the error parameter α _j with the similarity w _j .

Based on this concept, and more specifically, the variation calculation unit 15, as shown in FIG. 15 (A), taking the similarity w on the horizontal axis with taking the prediction value y ₀ on the vertical axis, First, for each of M prediction values y _0j corresponding to each of the M error parameters α _j (j = 1 to M) calculated from the M past performance data (X, y), The similarity w _{j is made} to correspond. Then, the variation calculation unit 15 divides in FIG 15 (A) the longitudinal axis y ₀ in at least a range y _0j including each predicted value finite number of a plurality of sections of (class, grade), contained in each interval A weighted frequency F _w is generated by adding all the similarities w _j of the predicted value y _0j , and a histogram representing the variation of the predicted value y ₀ is generated as shown in FIG.

That is, as shown in Equation 19, if the set of j the predicted value _{y 0j} is included in the k-th interval _{(Y k} or _{Y k + 1} less than the section) and _{S k (S k = {j} | Y k ≦ y _0j <Y _{k + 1} }) and the sum of the similarities w _j for j included in the set S _k is the weighted frequency F _w in the k-th interval.

This variation of the predicted value y ₀ as is indicated by the histogram, it is possible to easily know the frequency of occurrence of the predicted value y _0.

As described above, the histogram illustrated in FIG. 15B may be the variation of the predicted value y ₀ , but in the present embodiment, the variation calculation unit 15 further has an area of the histogram illustrated in FIG. Normalize so that This normalized histogram is taken as the probability density of the predicted value y ₀ in the prediction target data X _0, and is regarded as the variation of the predicted value y ₀ . Further, the variation calculation unit 15 may represent the histogram shown in FIG. 15B as a curve as shown in FIG. 15C while maintaining the area at 1. This curve is a probability density of the predicted value y ₀ in the prediction target data X _0, are variations of the predicted value y _0.

Note that the normalization, for example, when dividing the longitudinal axis y ₀ in FIG. 15 (A) into a finite number of intervals, uniform width h = | when a is divided into _| Y k + _{1 -Y} k , According to Equation 20.

  Further, when obtaining the curve shown in FIG. 15C from the histogram shown in FIG. 15B, a known probability distribution such as a lognormal distribution or a Weibull distribution may be used.

  FIG. 16 is a diagram for explaining a method for obtaining the probability density curve shown in FIG. 15C from the histogram shown in FIG. 16A shows a state in which the center points of the histogram are connected by a broken line, FIG. 16B shows the cumulative probability density of FIG. 16A, and FIG. 16C shows FIG. FIG. 16D shows a state where the cumulative probability density shown in B) is smoothed, and FIG. 16D shows the smoothed probability density (probability density curve).

First, as shown in FIG. 16 (A), in the histogram normalized shown in FIG. 15 (B), connecting the center positions of the respective frequencies of (y ₀ direction of the center) in a line. At both ends, a point separated from the end by a half (h / 2) of the section width h is also set to 0 and connected to the broken line. The area surrounded by the broken line is also 1.

Next, as shown in FIG. 16 (B), the cumulative probability density _SwN is obtained from the equation 21-1 from FIG. 16 (A).

Next, as shown in FIG. 16C, the cumulative probability density _SwN of the broken line in FIG. _16B is smoothed by using, for example, Expression 21-2.

  Then, as shown in FIG. 16D, a smoothed probability density (probability density curve) is obtained from the smoothed cumulative probability density shown in FIG. It is done.

Thus, the variation of the predicted value y ₀ is indicated by the probability density, and the appearance probability of the predicted value y ₀ can be easily known.

Further, when calculating the weighted frequency F _w , the similarity w of the M past performance data (X, y) is arranged in descending order (in descending order), and a predetermined value set in advance from the larger one is set. The past performance data (X, y) up to the number (predetermined ratio) may be extracted, and the weighted frequency F _w may be obtained by using only the extracted data. Low historical performance data similarity w (X, y) by pre-removing the arithmetic processing amount of relief for calculating the weighted frequency F _w (shorter processing time) becomes possible. Further, in the case of calculating the similarity w by Equation 14 described above (formula 14-1～ formula 14-3), historical performance data (X, y) is low similarity w between the prediction target data _{X 0} for even, Since the similarity does not become 0, the weighted frequency F _w is affected. For this reason, the width of the weighted frequency F _w shown in FIG. 15B matches the width of M predicted values y ₀ , and when the function f is Equation 16, the width is the prediction target data always constant irrespective of the conditions of X _0. As a result, the base of probability density shown in FIG. 15C may spread more than necessary. However, as described above, by excluding the past performance data (X, y) having a small similarity w, the base of the probability density is prevented from being excessively expanded, and the predicted value y ₀ of the prediction target data X ₀ is reduced. The characteristics of the distribution shape are remarkably expressed.

The output value prediction device S, the arithmetic control unit 1, the variation of the predicted value y ₀ calculated by the variation calculation unit 15 in the processing S15 is presented to the presentation unit 3 (S16), the processing is terminated. This variation of the predicted value y ₀ is presented to the presentation unit 3 as the user is able to know the variation of the predicted value y _0, the prediction in the case of performing the operation or judgment, etc. based on the predicted value y ₀ variation value y ₀ it becomes possible to consider.

By operating the output value predicting apparatus S in this way, by using M error parameters α _j (j = 1 to M) calculated from M past performance data (X, y), a prediction target predicted value _{y 0} of the data _{X 0} is calculated as M, and weighted frequency _{F w} are calculated for the predicted value _{y 0} in accordance with the similarity _{w j} from the predicted target data _{X 0.} Furthermore, the probability density is calculated from the weighted frequency F _w. Therefore, historical performance data (X, y) and the variation of the predicted value y ₀ in the prediction target data X ₀ of similarity is considered the predicted target data X ₀ is obtained with high accuracy. Accordingly, the output value prediction unit S may present variations of the predicted value y _0, and thus also the variation of the predicted value y ₀ if based on the predicted value y ₀ performing operations and judgments, etc. can be considered It becomes.

In the present embodiment, the variation in the output value y of the prediction target data X ₀ based on the similarity w _j between the prediction target data X ₀ and the past performance data (X, y) and the past performance data (X, y). when calculating a degree data item x _i contribute to the magnitude of variation in a given output y is calculated for the data item x _i as the a weight a _i, a first a weight a _i the calculated The predetermined distance d _j between the prediction target data X ₀ and the past performance data (X, y) is calculated, and the similarity w _j is calculated based on the calculated predetermined distance d _j. The Therefore, in the present embodiment, it is possible to determine the variation of the predicted value y ₀ to assess the impact on the variation of the predicted value y ₀ for the data item x _i. For this reason, the similarity w _j between the prediction target data X ₀ and the past performance data (X, y) is more appropriately evaluated, and the accuracy of variation in the predicted value y ₀ is further improved.

Next, another embodiment will be described.
(Second Embodiment)
For example, in the case of products manufactured through various manufacturing processes in a relatively large-scale manufacturing plant, such as the manufacture of steel products and chemical products, for example, according to the input amount, operation input amount, time passage, etc. The output value in the manufacturing process and the output value in the final process directly connected to the product often change every moment. For example, in the manufacturing process of steel products, the relationship between the hot metal temperature in the topped car and the elapsed time, the relationship between the molten steel temperature in the ladle and the elapsed time, and the carbon concentration in the molten steel and the integrated oxygen injection value in the converter blowing. The relationship and the relationship between the molten steel temperature and the blown oxygen integrated value in converter blowing are mentioned.

FIG. 17 is a diagram illustrating the relationship between the temperature drop amount of the object and the elapsed time. FIG. 17 shows the relationship between the temperature drop amount (deviation from the initial temperature) y of the object left in the atmosphere and the elapsed time (temperature measurement time) t plotted with ○ for each past performance data. Assume that it was a result. Here, when predicting the temperature drop amount y (t ₀ ) at the predetermined time t ₀ , the following problem may arise when the probability density is obtained by using past performance data near the time t ₀ . That is, firstly, in the time region where the past performance data is small (or does not exist), there is very little data that can be utilized, and the data that is utilized is only a part of the past performance data. For this reason, it is difficult to predict the distribution of the temperature drop amount y (t ₀ ) of the prediction target data with high accuracy. Secondly, past performance data having a high degree of similarity with the prediction target data is not necessarily near the predetermined time t ₀ , and a past performance having a high degree of similarity with the prediction target data at a location away from the time t _0. If there is data, the past performance data is not used.

Therefore, for such a problem, as shown in FIG. 17 with a part of the past performance data by a thin broken line, a prediction model representing the relationship between the temperature drop amount y and the elapsed temperature t in each past performance data is constructed. By past each past performance data at a predetermined time t ₀ (that is, by determining the amount of temperature drop y (t ₀ ) at the time t ₀ of the constructed prediction model), the past away from the predetermined time t ₀ actual data can also be utilized for the estimation of the probability density in the prediction value y (t _0), it is possible to determine the variation of the predicted value y of the predicted target data (t ₀₎ with higher accuracy.

  In the second embodiment, the output value prediction apparatus S of the first embodiment is applied to the above-described case. Therefore, in the output value prediction apparatus S of the second embodiment, the output value prediction apparatus S of the first embodiment includes parameter calculation processing (S13) for calculating parameters and prediction value calculation processing (S14) for calculating prediction values. Since it is the same as that of the output value prediction apparatus S in 1st Embodiment except the point which performs a process as follows, description of the same point is abbreviate | omitted.

FIG. 18 is a diagram illustrating a model of past performance data. FIG. 19 is a diagram for explaining a procedure for calculating a variation in predicted values. FIG. 19A shows the predicted value y (t ₀ ) at a predetermined time t ₀ , and FIG. 19B shows the relationship between the similarity w and the output predicted value y (t ₀ ). The horizontal axis is the similarity w, and the vertical axis is the predicted value y (t ₀ ).

  In the output value prediction apparatus S of the second embodiment, the past result data and the prediction target data are stored in advance in the table format (table format) in the actual measurement data storage unit 41 of the storage unit 4 as in the first embodiment. ing. In the second embodiment, the past performance data and the prediction target data include the temperature drop amount y, the actual measurement time t when the temperature drop amount y is measured, and the factor data x related to the temperature drop amount y. Is done. The temperature drop amount y corresponds to the output value y, and the actual measurement time t can be regarded as one of the factors in the factor related to the output value y. That is, the factor X related to the output value y includes at least the time t as an element. The origin of the actual measurement time t is the time when the temperature drop amount y = 0, that is, the time of the initial temperature of the object (the time when the measurement of the temperature of the object is started). Note that the output value prediction apparatus S may be configured such that the time t is not included in the factor X but is included in the factor Z.

The historical performance data ((X, t), y) the predicted target data _(X 0, _{t 0)} the output value from the (predicted value) prediction of _{y 0} is started on the basis of, in the processing S11, first Similar to the embodiment, the distance calculation unit 11 calculates the weight a by the weight calculation unit 111, and the past performance data (X, t) and the prediction target using the weight a in the first to Nth data item spaces. A distance d between the data (X ₀ , t ₀ ) is calculated, and the calculated distance d is stored in the intermediate data storage unit 42 of the storage unit 4.

Next, the process S12, as in the first embodiment, the similarity calculating unit 12, the similarity w between the predicted target data X ₀ and historical performance data X, first through M historical performance data ( X, y) is calculated, and the calculated similarity w is stored in the intermediate data storage unit 42 of the storage unit 4.

Next, in the process S13, instead of the expression 15, the expression 15A; y _j (t) = f (Z, Θ, α _j , t) is used, and the others are processed in the same manner as in the first embodiment. The parameter calculation unit 13 calculates an error parameter α representing an uncertain element for each of the first to M-th past performance data ((X, t), y), and calculates each calculated error parameter α _j . The data is stored in the intermediate data storage unit 42 of the storage unit 4. Here, y _j (t) is a temperature drop amount at the actual measurement time t, and Z, Θ, and α _j are the same as Expression 15 in the first embodiment. In FIG. 17, Formula 15A is indicated by a thick broken line, and in FIG. 18, the model of each past performance data ((X, t), y) is indicated by a thin broken line.

Next, the process S14, using the prediction time t _0, by others treated in the same manner as the first embodiment, the prediction value calculating unit 14, as shown in FIG. 19 (A), obtained by the step S13 the model used, the predicted target data _(X 0, _{t 0)} the predicted value based on the predicted time _{t 0} and the first to each data value _x 01 _{~x 0N} N-th data item in the _y 0 to _{(t 0)} is calculated for each of the error parameters alpha _j obtained in processing S13, the predicted value _y 01 _{(t 0)} which is the calculated _~y 0M _{(t 0)} the storage in association with the degree of similarity _w 1 to w _M Stored in the predicted value storage unit 43 of the unit 4. In FIG. 19A, each predicted value y _0j (t ₀ ) at the predicted time t ₀ is indicated by a _circle .

Next, the process S15, as in the first embodiment, the variation calculation unit 15 uses the respective predicted value _y 01 _{~y 0M} obtained in the processing S14, the variation of the predicted value _{y 0 (t} 0) And the variation of the calculated predicted value y ₀ (t ₀ ) is stored in the variation storage unit 44 of the storage unit 4. More specifically, as shown in FIG. 19 (B), the variation calculation unit 15 takes the predicted value y (t ₀ ) on the vertical axis and the similarity w on the horizontal axis. M prediction values y _0j (t ₀ ) respectively corresponding to M error parameters α _j (j = 1 to M) calculated from each past performance data ((t, x), y). On the other hand, the similarity w _{j is made} to correspond. Next, the variation calculation unit 15 _divides a range including at least each predicted value y _0j (t ₀ ) on the vertical axis y (t ₀ ) in FIG. 19B into a finite number of sections (classes, grades). Then, the weighted frequency F _w is generated by adding all the similarities w _j of the predicted values y _0j (t ₀ ) included in each section, and a histogram representing the variation of the predicted values y ₀ (t ₀ ) is generated. To do. The histogram may be a variation of the predicted value y ₀ (t ₀ ), but in the present embodiment, the variation calculation unit 15 further normalizes the area of the histogram to be 1. This normalized histogram is the probability density of the predicted value y ₀ (t ₀ ) in the prediction target data (t ₀ , x ₀ ), and is the variation of the predicted value y ₀ (t ₀ ). Or the variation calculation part 15 calculates | requires the said curve similarly to the above-mentioned from this histogram, further maintaining the area at 1. This curve is the probability density of the predicted value y ₀ (t ₀ ) in the prediction target data (t ₀ , x ₀ ), and is the variation of the predicted value y ₀ (t ₀ ).

By operating in this way, the output value predicting apparatus S of the second embodiment can obtain the predicted value y ₀ of the output in the process in which the output changes momentarily with time, and the factor x _i in consideration of the degree of influence on the variation of the predicted value y _0, it is possible to determine the variation of the predicted value y _0. Further, in the output value prediction device S of the second embodiment, it is possible to obtain the predicted value y ₀ even in a time region where the past performance data is small (or does not exist), and it is also possible to obtain the variation of the predicted value y _0. It becomes. Further, in the output value prediction apparatus S of the second embodiment, past performance data separated from a predetermined time t ₀ can also be used for estimation of variation in the predicted value y ₀ (t ₀ ), and prediction of prediction target data is performed. Variations in the value y ₀ (t ₀ ) can be obtained with higher accuracy. In addition, as can be seen from FIG. 19, the output value prediction apparatus S of the second embodiment can obtain the predicted values y ₀ (t) at a plurality of different times t, and the variation of the predicted values y ₀ (t). Can also be sought. Therefore, it is possible to determine the processing end timing with the least risk by comparing the variation of the predicted value y ₀ (t) at each time t. For example, in a heating furnace of a steel product manufacturing process, it is possible to terminate the heat treatment at a low risk timing, considering not only whether the steel was heated as intended but also considering the probability of temperature deviation. It becomes. Further, for example, in converter blowing, it is possible to terminate the blowing at a timing with less risk in consideration of the molten steel temperature and the probability that the components in the molten steel deviate from the target. In the case of this converter blowing, the horizontal axis in FIG. 19A is the integrated value of the blown oxygen amount.

Next, another embodiment will be described.
(Third embodiment)
In the steel product manufacturing process, after the converter is blown, the molten steel is transferred from the converter to the ladle, and after the molten steel is processed, the molten steel is transported to the continuous casting equipment. Therefore, it is preferable that the molten steel temperature is slightly higher than the solidification temperature when the ladle arrives at the continuous casting facility. If the molten steel temperature is too low, the molten steel solidifies, which is not preferable. If the molten steel temperature remains high, the casting speed must be reduced. Depending on the charge, the molten steel composition, the amount of molten steel, the type of ladle, the initial state of the ladle (the refractory melting condition, the temperature distribution inside the ladle (cooling condition)), and the ladle when receiving steel from the converter The amount of temperature drop varies depending on the amount of alloy and type of alloy placed in advance. For this reason, it is difficult to predict the molten steel temperature that changes from moment to moment at a single point. Therefore, it is important to accurately estimate the variation in molten steel temperature in the ladle for the charge.

  In the third embodiment, the predetermined output is the molten steel temperature in the ladle in the process from the converter steelmaking process through the molten steel treatment process to the continuous casting process, and the output value prediction device S of the first embodiment is applied. The output value predicting apparatus S in the third embodiment estimates the probability distribution of the temperature drop of the molten steel until the molten steel transferred from the converter to the ladle is transported to the molten steel processing facility. It is. Therefore, the output value prediction apparatus S in the third embodiment is the same as the output value prediction apparatus S in the first embodiment, in the distance calculation process (S11) for calculating the distance, the parameter calculation process (S13) for calculating the parameter, and the predicted value. The predicted value calculation process (S14) for calculating is the same as the output value prediction apparatus S in the first embodiment except that the process is executed as follows, and the description of the same points is omitted.

20 and 21 are diagrams showing the relationship between the weighted distance and the similarity. 20 and 21, the horizontal axis represents the weighted distance _dj , and the vertical axis represents the similarity _wj . FIG. 22 is a diagram illustrating a model of past performance data. FIG. 23 is a diagram showing the probability density at each predicted value. The horizontal axis of FIG. 22 and FIG. 23 is the elapsed time t expressed in minutes (min), and the vertical axis thereof is the temperature drop y (t) expressed in degrees (° C.).

  In the output value prediction apparatus S of the third embodiment, the past record data and the prediction target data are stored in advance in a table format (table format) in the actual measurement data storage unit 41 of the storage unit 4 as in the first embodiment. ing. In the second embodiment, the past performance data and the prediction target data include the temperature drop amount y, the actual measurement time t when the temperature drop amount y is measured, and the factor data x related to the temperature drop amount y. Is done. The temperature drop amount y corresponds to the output value y, and the actual measurement time t can be regarded as one of the factors in the factor related to the output value y. The origin of the actual measurement time t is the time when the temperature drop amount y = 0, that is, the time of the initial temperature of the object (the time when the measurement of the temperature of the object is started). Each factor (data item) in the factors related to the temperature drop y is the number of times the ladle is received, the type of deoxidizer, the concentration of molten steel, the state of the ladle emptying, the temperature at which the steel is tapped, the solidification temperature, and the operation group Etc. for each item. Here, in the present embodiment, the number of times the steel is received by the ladle is converted by a non-linear function so as to be the square root of the number of times the steel is received, for example. The type of deoxidizer is quantified according to the strength of deoxidation. When the ladle is in an empty pan state (a state in which no molten steel is contained), the time of standing, the time of keeping warm, the standing time after keeping warm, etc. are quantified by a non-linear function. The operation group is given an identifier for each group.

When the prediction of the output value (predicted value) y ₀ is started from the prediction target data (X ₀ , t ₀ ) based on the past performance data ((X, t), y), in the process S11, the expression 22 The distance calculation unit 11 calculates the Ath weight a by the weight calculation unit 111, and performs the same processing as in the first embodiment except for the distance d defined by the first to Nth data item spaces. 2, the distance d between the past performance data (X, t) and the prediction target data (X ₀ , t ₀ ) is calculated using the A-th weight a, and the calculated distance d is stored in the middle of the storage unit 4. The data is stored in the data storage unit 42.

Here, f _d (X _ji , x _0i ) is a function that takes 0 when x _ji and x _0i are the same, and takes 1 when x _ji and x _0i are different. In this embodiment, a _i (i = 1 to N) = 1. N is the number of data items. Further, k <N.

  When comparing the operating conditions of the relevant charge with the operating conditions of each past charge, it may be subtracted, such as the operating team or equipment number (number of equipment used for processing among multiple equipment). Some data items cannot be used, or some data items have no meaning in subtraction itself, and the distance d defined by Equation 23 is valid for such data items that have a meaning as to whether they are the same. is there.

In addition, date and time or date may be added as a data item when calculating the similarity. Some processes have characteristics that change over time, such as aging and seasonal variation. In such a case, even if the operating conditions are the same, the result may be different if the date is different. By adding the month and day as a data item, it is possible to reduce the similarity w _j of the old data and make a prediction considering the secular change. Note that the date may be expressed as the number of days elapsed from a reference date (for example, January 1, 1900).

Next, the process S12, as in the first embodiment, the similarity calculating unit 12, the similarity w between the predicted target data X ₀ and historical performance data X, first through M historical performance data ( X, y) is calculated, and the calculated similarity w is stored in the intermediate data storage unit 42 of the storage unit 4.

  Here, as the similarity w, the similarity defined by Expression 14-3, that is, Expression 23, is used.

Here, μ is an average value of the distance d _j (j = 1 to M), and σ is a standard deviation of the distance d _j (j = 1 to M). In the present embodiment, g = 1 is set. FIG. 20 shows the relationship between the weighted distance d _j according to the equation 22 and the similarity w _j according to the equation 23.

In the present embodiment, as shown in FIG. 21, the similarity w _j smaller than a predetermined threshold set in advance is set to zero. By previously removing past performance data ((t, x), y) having a low degree of similarity w, for example, the following arithmetic processing amount such as the arithmetic processing amount for calculating the weighted frequency F _w is reduced (arithmetic processing) Time reduction).

Next, in the process S13, the parameter calculation unit 13 uses the expression 24-1 and the expression 24-2 instead of the expression 15 and the other processes in the same manner as in the first embodiment, so that the parameter calculation unit 13 determines the uncertain element An error parameter α to be expressed is calculated for each of the first to M-th past performance data ((X, t), y), and the calculated error parameters α _j are stored in the intermediate data storage unit 42 of the storage unit 4. .

Here, y _j (t) is the amount of temperature drop at the actual measurement time t, T _0j is the converter steel temperature that is the molten steel temperature at the time when the molten steel is transferred from the converter to the ladle, and T _∽j is the solidification temperature of the molten steel, q _j is a value obtained by converting the amount of heat to a temperature previously entered by that alloy (put every alloy) rob the ladle, when placed every alloy no q _j = 0.

In the present embodiment, the predicted predicted melting data items include the converter steel temperature T _0j , the solidification temperature T _{∽j of the} molten steel, and the _{input of the detained} alloy necessary for estimating the temperature drop q _j due to the _detained alloy. Amount.

When the predicted value y ₀ is predicted by a multiple regression equation, the error parameter α _j in the j-th past performance data ((t _j , x _j ), y _j ) based on the function equation 24 is 25.

Note that the predicted value y _0j (t ₀ ) corresponding to the error parameter α _j obtained from the j-th past performance data is given by Expression 26-1 and Expression 26-2.

Here, T ₀₀ is the converter steel temperature in the prediction target charge, T _∽0 is the solidification temperature of the molten steel in the prediction target charge, and q ₀ is the _detained alloy in the prediction target charge. It is a value obtained by converting the amount of heat taken into temperature.

  FIG. 22 shows a model of each past performance data ((X, t), y).

Next, the process S14, using the prediction time t _0, the other is by treating in the same manner as the first embodiment, the prediction value calculating unit 14 uses the model obtained in the processing S13, the prediction target data ( X _{0, t 0} the predicted time _{t 0} and the predicted value _y 0 on the basis of each data value _x 01 _{~x 0N} the first to N data items _{(t 0)} in), each error parameters obtained in the processing S13 alpha calculated for each _j, the predicted value _y 01 that the calculated _{(t 0) ~y 0M (t} 0) the similarity _w 1 to w _M and association with each other in the predicted value storage unit 43 of the storage unit 4 To do.

Next, the process S15, as in the first embodiment, the variation calculation unit 15 uses the respective predicted value _y 01 _{~y 0M} obtained in the processing S14, the variation of the predicted value _{y 0 (t} 0) And the variation of the calculated predicted value y ₀ (t ₀ ) is stored in the variation storage unit 44 of the storage unit 4. More specifically, the variation calculation unit 15 takes the predicted value y (t ₀ ) on the vertical axis and the similarity w on the horizontal axis. First, each of M pieces of past performance data ((t, x)) , Y) and M prediction values y _0j (t ₀ ) respectively corresponding to the M error parameters α _j (j = 1 to M) respectively calculated from y), the similarity w _{j is made} to correspond. . Next, the variation calculation unit 15 _divides a range including at least each predicted value y _0j (t ₀ ) of the vertical axis y (t ₀ ) into a finite number of sections (classes, grades), and A weighted frequency F _w is generated by adding all the similarities w _j of the included predicted values y _0j (t ₀ ), and a histogram representing the variation of the predicted values y ₀ (t ₀ ) is generated. The histogram may be a variation of the predicted value y ₀ (t ₀ ), but in the present embodiment, the variation calculation unit 15 further normalizes the area of the histogram to be 1. This normalized histogram is the probability density of the predicted value y ₀ (t ₀ ) in the prediction target data (t ₀ , x ₀ ), and is the variation of the predicted value y ₀ (t ₀ ). Or the variation calculation part 15 calculates | requires the said curve similarly to the above-mentioned from this histogram, further maintaining the area at 1. This curve is the probability density of the predicted value y ₀ (t ₀ ) in the prediction target data (t ₀ , x ₀ ), and is the variation of the predicted value y ₀ (t ₀ ). FIG. 23 shows the probability density of the predicted value y ₀ (t ₀ ) of the temperature drop amount every 10 minutes. In other words, the predicted time point is every 10 minutes. In FIG. 23, ○ indicates the actually measured temperature, ○ at time 0 is the temperature at which the steel is discharged from the converter, and ○ after about 50 minutes is the start of the molten steel treatment in the prediction target data (t ₀ , x ₀ ). This is the previous measured temperature. The measured temperature indicated by ○ after about 50 minutes is used to calculate the predicted value y ₀ (t ₀ ) of the temperature drop amount or the probability density of the predicted value y ₀ (t ₀ ) of the temperature drop amount shown in FIG. Is not used, but is displayed in FIG. 23 for reference. Further, the broken line connecting both circles in FIG. 23 is obtained by calculating the molten steel temperature of the charge to be predicted by Equation 24, fitting the error parameter α ₀ of Equation 25, and similarly, the temperature drop amount the calculation of the probability density of the predicted value y _{0 (t 0)} the amount of temperature drop of the predicted value y ₀ shown in calculating and FIG 23 _{(t 0)} is not used, for reference, is shown in Figure 23 ing. In FIG. 23, the scale of the horizontal axis of probability density (corresponding to the horizontal axis of FIG. 15C) is enlarged for easy viewing.

  By operating in this way, in the output value prediction device S of the third embodiment, in the process from the converter steelmaking process through the molten steel treatment process to the continuous casting process, it is given to the variation in the molten steel temperature in the ladle. In consideration of the degree of influence, it is possible to predict the molten steel temperature in the ladle of the charge, and to obtain the predicted variation in molten steel temperature in the ladle with higher accuracy.

  In the third embodiment described above, the predetermined output is the molten steel temperature in the ladle in the process from the converter steelmaking process through the molten steel treatment process to the continuous casting process, but the predetermined output is It may be the molten steel temperature in the tundish in the process from the converter steelmaking process through the molten steel treatment process to the continuous casting process. By configuring in this way, it is possible to predict the molten steel temperature in the tundish in the process from the converter steelmaking process to the continuous casting process through the molten steel treatment process, and obtain the variation of this predicted value It becomes.

  Further, in the above-described third embodiment, the predetermined output may be a molten steel component or a molten steel temperature according to the integrated amount of blown-suction oxygen in the converter process. By being configured in this way, it is possible to predict the molten steel component or molten steel temperature corresponding to the integrated amount of blown-in oxygen in the converter process, and to obtain variations in the predicted value.

  In the third embodiment described above, the predetermined output may be the steel material temperature of the steel material in accordance with the heating time or the integrated amount of heat in the steel material heating furnace step. By comprising in this way, it becomes possible to estimate the steel material temperature of the steel material according to the heating time or the integrated amount of the heating heat amount in the heating furnace process of the steel material, and to obtain the variation of the predicted value.

  As described in the first to third embodiments, the output value predicting apparatus S can predict the output value in each process of the operation process and the manufacturing process and the output value of the final process directly connected to the product together with variations. Here, a configuration example of an output value prediction system in which the output value prediction device S is applied to an operation process or a manufacturing process will be described.

  FIG. 24 is a block diagram illustrating a configuration of the output value prediction system. In FIG. 24, the output value prediction system includes a performance data collection device 101 that collects performance data from the operation process / manufacturing process 201 and a past operation data storage that stores performance data collected by the performance data collection device 101 as past performance data. Device 105, parameter fitting arithmetic device 102 for calculating error parameter α based on past actual data collected by actual data collecting device 101, and parameter estimated value storage device for storing error parameter α calculated by parameter fitting arithmetic device 102 106, a prediction target data operation condition collection device 104 that collects prediction target data from the operation process / manufacturing process 201, a similarity calculation device 108 that calculates a similarity between the prediction target data and each past performance data, and a prediction target Data operating condition collection device The prediction target data output prediction calculation device 103 that predicts the output value (prediction value) of the prediction target data from the prediction target data collected in 104 based on the calculated error parameter α, and the prediction target data output prediction calculation device 103 performs prediction. An output predicted value probability distribution estimation device 107 that calculates the probability density of the predicted value of the prediction target data based on the output value (predicted value) of the prediction target data and the similarity calculated by the similarity calculation device 108; And a probability distribution display device 109 for displaying the probability density of the predicted value of the prediction target data calculated by the probability distribution estimation device 107.

  When the output value prediction system shown in FIG. 24 and the output value prediction device S shown in FIG. 1 are compared, the similarity calculation device 108 includes a distance calculation unit 11 (including a weight calculation unit 111), a similarity calculation unit 12, and an intermediate The parameter fitting calculation device 102 has substantially the same function as the data storage unit 42, the parameter fitting calculation device 102 has substantially the same function as the parameter calculation unit 13, and the prediction target data output prediction calculation device 103 includes the prediction value calculation unit 14 and the prediction The output predicted value probability distribution estimation device 107 has substantially the same function as the variation calculation unit 15 and the variation storage unit 44, and the past operation data storage device 105 has a track record. The parameter estimated value storage device 106 has substantially the same function as the data storage unit 41, and the parameter estimated value storage device 106 has substantially the same function as the intermediate data storage unit 42.

  When the prediction target data is collected during the execution of the process, the output value prediction system having such a configuration can obtain a prediction value and its probability density in the prediction target data, and can display them. . For this reason, a user such as an operator can appropriately adjust the process based on the predicted value in the prediction target data and its probability density, and can execute the process.

  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.

S output value prediction device 1 calculation control unit 4 storage unit 11 distance calculation unit 12 similarity calculation unit 13 parameter calculation unit 14 prediction value calculation unit 15 variation calculation unit 41 actual measurement data storage unit 42 intermediate data storage unit 43 prediction value storage unit 44 Variation storage unit 102 Parameter fitting calculation unit 103 Prediction target data output prediction calculation unit 105 Past operation data storage unit 107 Output predicted value probability distribution estimation unit 108 Similarity calculation unit 111 Weight calculation unit

Claims (13)

The degree of similarity between the prediction target data to be predicted and a plurality of past performance data acquired in the past consisting of a predetermined output and a quantifiable factor related to the output, the factor of the prediction target data and the past performance data A first calculation step of calculating each of the plurality of past performance data based on the factors of
A second calculation step of calculating a variation in the output value of the prediction target data based on the plurality of similarities calculated in the first calculation step and the past performance data;
The first calculation step includes
A distance calculating step of calculating a predetermined distance between the prediction target data and the past performance data for each of the plurality of past performance data based on a factor of the prediction target data and a factor of the past performance data;
A similarity calculation step of calculating the similarity between the prediction target data and the past performance data for each of the plurality of past performance data based on the distance calculated in the distance calculation step;
The distance calculating step includes a weight calculating step of calculating the factor with the degree that the factor contributes to the magnitude of variation in the predetermined output as the A weight, and the A weight calculated in the weight calculating step And calculating the predetermined distance using the output value prediction method. The weight calculation step includes
When the magnitude of the variation in the predetermined output is a first output variable and the variable related to the factor is a first input variable, the first input variable and the first variable are based on the plurality of past performance data. The output value prediction method according to claim 1, wherein a first model representing a relationship with an output variable is generated, and the A-th weight is calculated based on the first model. 3. The output according to claim 2, wherein the weight calculating step calculates a B-weight according to the number of the plurality of past performance data, and generates the first model using the B-weight. Value prediction method. The weight calculation step includes
4. The output value prediction method according to claim 2, wherein the first model is generated based on remaining past performance data obtained by removing predetermined past performance data from the plurality of past performance data. 5. In the distance calculating step, a weight calculating step of calculating the factor with the degree that the factor contributes to the magnitude of variation in the predetermined output and the degree that the factor contributes to the magnitude of the predetermined output as an A weight. 5. The output value prediction method according to claim 1, wherein the predetermined distance is calculated using the A-th weight calculated in the weight calculation step. The second calculation step includes
When a second model representing the relationship between the output variable and the input variable is generated using the input variable when the predetermined output is an output variable and part or all of the factors are input variables In addition, an error parameter, which is a difference between a value obtained by giving the input value of the input variable to the second model and an output value of the output variable corresponding to the input value of the input variable, A parameter calculation step for calculating each of the plurality of past performance data based on input variables and output variables;
A third model representing a relationship between the output variable and the input variable is generated using the input variable and the error parameter, and a factor value corresponding to the input variable and the error among the factors of the prediction target data A predicted value calculating step of calculating each of the plurality of error parameters calculated by the parameter calculating step by using the output value of the prediction target data as a predicted value by giving a parameter value to the third model;
A variation calculating step of calculating variations in the output value of the prediction target data based on the plurality of similarities calculated by the similarity calculating step and the plurality of predicted values calculated by the predicted value calculating step. The output value prediction method according to claim 1, wherein: The output factor prediction method according to claim 1, wherein the factor includes a plurality of elements, and includes at least time as the element. The predetermined power is a molten steel temperature in a ladle or in a tundish in a process from a converter steeling process to a continuous casting process through a molten steel treatment process. The output value prediction method according to any one of the preceding claims. The output value according to any one of claims 1 to 7, wherein the predetermined output is a molten steel component or a molten steel temperature corresponding to an integrated amount of blown-suction oxygen in the converter process. Prediction method. The said predetermined output is the steel material temperature of the said steel material according to the integration amount of the heating time or the amount of heating heat in the heating furnace process of steel materials, The any one of Claims 1 thru | or 7 characterized by the above-mentioned. Output value prediction method. The output value prediction method according to any one of claims 1 to 10, further comprising a presentation step of presenting the variation. An actual measurement data storage unit for storing a plurality of past performance data acquired in the past, which includes a predetermined output and a factor that can be numerically related to the output;
A first calculation unit that calculates the similarity between the prediction target data to be predicted and the past performance data for each of the plurality of past performance data based on the factor of the prediction target data and the factor of the past performance data;
A second calculation unit that calculates variations in the output value of the prediction target data based on the plurality of similarities calculated by the first calculation unit and the past performance data;
The first calculation unit includes:
A distance calculation unit that calculates a predetermined distance between the prediction target data and the past performance data for each of the plurality of past performance data based on a factor of the prediction target data and a factor of the past performance data;
A similarity calculation unit that calculates the similarity between the prediction target data and the past performance data for each of the plurality of past performance data based on the distance calculated by the distance calculation unit;
The distance calculation unit includes a weight calculation unit that calculates the factor by using the degree that the factor contributes to the magnitude of variation in the predetermined output as the Ath weight, and the Ath weight calculated by the weight calculation unit The output value predicting apparatus, wherein the predetermined distance is calculated by using. An output value prediction program for causing a computer to execute,
The degree of similarity between the prediction target data to be predicted and a plurality of past performance data acquired in the past consisting of a predetermined output and a quantifiable factor related to the output, the factor of the prediction target data and the past performance data A first calculation step of calculating each of the plurality of past performance data based on the factors of
A second calculation step of calculating a variation in the output value of the prediction target data based on the plurality of similarities calculated in the first calculation step and the past performance data;
The first calculation step includes
A distance calculating step of calculating a predetermined distance between the prediction target data and the past performance data for each of the plurality of past performance data based on a factor of the prediction target data and a factor of the past performance data;
A similarity calculation step of calculating the similarity between the prediction target data and the past performance data for each of the plurality of past performance data based on the distance calculated in the distance calculation step;
The distance calculating step includes a weight calculating step of calculating the factor with the degree that the factor contributes to the magnitude of variation in the predetermined output as the A weight, and the A weight calculated in the weight calculating step An output value prediction program using the predetermined distance to calculate.

Images