JP2012074007A - Operation condition optimization system of plant, operation condition optimization method of plant and operation condition optimization program of plant - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an operation condition optimization system of a plant or the like which is easily modeled at low cost, is actually implemented in a plant by using a variable controlled and set by a person and gives an optimum operation condition in a range keeping correlation between variables shown by acquired data.SOLUTION: A measurement data recording part 43 associates operational state data S-DATA acquired by an operational state data acquisition part 41 with operational index data I-DATA calculated by an operational index data acquisition part 42 based on a prescribed item, and records it as a series of measurement data in a measurement DB. A regression model creation part 44 performs a prescribed multivariable analysis using an operational state variable showing the operational state data S-DATA side as an explanatory variable and an operational index variable showing the operational index data I-DATA side as an objective variable and creates a regression model. An operational index variable optimization part 45 calculates an operational state variable optimizing the operational index variable based on the regression model.

Description

  The present invention relates to a plant operation condition optimization system, a plant operation condition optimization method, and a plant operation condition optimization program for obtaining an operation state variable that optimizes an operation index variable based on a regression model.

What is desired in the operation of the plant is product quality and yield in the case of a production plant, and operation efficiency and cost in the case of other plants. Furthermore, it is desired to operate the plant in which the operation index such as the CO ₂ emission amount as an environmental factor is optimum, and realizing these is a great concern in the plant operation. Conventionally, methods, devices and the like as described in the following Patent Documents 1, 2, and 3 have been proposed for this concern.

  For example, in the power plant operation optimization method and apparatus disclosed in Patent Literature 1, a physical model for calculating turbine efficiency as a plant operation evaluation index based on entropy calculation from the temperature and pressure of the turbine inlet steam of the power plant (Patent Literature) 1 of FIG. 9 is used. As described in paragraph [0023] and subsequent paragraphs of Patent Document 1, a non-linear plan calculation unit is provided, and an optimum operating condition is obtained by a non-linear optimization method. In Patent Document 1, complicated optimization methods are combined in order to avoid local minimum.

  In the cogeneration plant and its operation method described in Patent Document 2, for a plurality of units of the cogeneration plant, a linear format (individual model) representing the characteristics of the amount of delivered steam and the amount of fuel consumption obtained by a single test or the like in advance for each unit. Based on the above, various plant operation evaluation indices are represented as a linear form (overall model) of operation conditions such as steam temperature and pressure (see FIG. 2 of Patent Document 2). As described in paragraph [0016] and subsequent paragraphs of Patent Document 2, the operation planning unit attempts to optimize one or more operation evaluation indexes by linear programming based on this linear format.

  In the operation variable selection device described in Patent Document 3, a term indicating the influence of the change in the operation variable of the target process on the quality variable of the target process, and the nonlinearity between the operation variable and the quality variable of the target process are expressed. An evaluation function given as an arbitrary nonlinear function of terms, terms representing costs and penalties related to operations, and these three terms is used.

JP 2008-199825 A JP 2006-17340 A JP 2006-323523 A

Toshihide Ibaraki, Masao Fukushima, FORTRAN77 optimization programming, Iwanami Computer Science, Iwanami Shoten, April 10, 1991 Yoshikatsu Miyashita, Shinichi Sasaki, Computer Chemistry Series 3 Chemometrics -Chemical Pattern Recognition and Multivariate Analysis-, Kyoritsu Publishing Co., Ltd., January 25, 1995

  As described above, although a physical model is used in Patent Document 1, it is unclear whether or not it appropriately represents the input / output physical quantity (relationship) of an actual model. Specifically, in order to match the actual physical quantity, it is necessary to adjust the parameters of the physical model, which requires a great deal of labor.

The power plant operation optimizing method described in Patent Document 1 has a problem in that the calculation is complicated because the nonlinear optimization method is used as described above. In Patent Document 1, complicated optimization methods are combined in order to avoid a local minimum, but this requires more calculation time. As a result, for example, it is impossible to determine an optimum driving condition by evaluating a plurality of driving evaluation indexes, and comparison evaluation between the obtained optimum driving condition and other conditions is performed. There was a problem that it was difficult.
The decision variables cannot be set completely independently. In general, the decision variables may be correlated with each other. In this case, it is necessary to change the correlation between the decision variables, but there is a problem that the general nonlinear optimization method cannot take it into account.

  In addition, since the parameters of the physical model corresponding to the operating conditions are not necessarily variables that can be controlled and set by humans, the optimal operating conditions obtained from the physical model can actually be implemented in the plant as it is. There was also a problem that it was not always.

  On the other hand, the “model” in Patent Document 2 is linear, and the optimization method is linear programming, which is simple, and a plurality of driving evaluation indexes can be considered. However, in order to create the “model”, it is necessary to evaluate the characteristics by a plant test or the like. In general, it is difficult to perform optimization based on a “model” as described in Patent Document 2. In reality, optimization depends on the intuition and experience of an expert, or by trial and error. In many cases, it is determined.

Also in Patent Document 2, as in the problem of Patent Document 1, the parameter of the “model” corresponding to the driving condition is not always a variable that can be controlled and set by a person. There is also a problem that it is not always possible to implement various operating conditions in the plant as it is.
On the other hand, if a non-linear model that is stricter than the linear model is obtained, the prediction accuracy is generally better than that of the linear model. Needless to say, it is desirable to use this nonlinear model with excellent prediction accuracy to determine an optimum decision variable while satisfying various constraints within a range in which the correlation between the decision variables is maintained.

  Further, as described above, since a term representing nonlinearity is used in Patent Document 3, there is a problem common to nonlinear optimization methods such as complicated calculations as in the problem in Patent Document 1.

Therefore, an object of the present invention is to solve the above problems, and to use a physical model that requires a great deal of effort to match an actual physical quantity and a nonlinear optimization method that complicates calculations. It can be modeled inexpensively and easily, can be implemented in the plant by using variables that can be controlled and set by humans, and is optimal within the range that maintains the correlation between the variables indicated by the acquired data An object of the present invention is to provide a plant operating condition optimization system and the like that can obtain various operating conditions.
Further, the object of the present invention is to model the relationship between the operating state and the plant operation evaluation index at that time from the already measured normal operation data of the plant by multivariate analysis, and based on this model, a linear optimization method Therefore, it is to obtain an operation condition that can realize an optimum operation evaluation index efficiently or with a small amount of calculation.
Furthermore, the object of the present invention is to make it possible to use a high-precision model, not limited to linear, and to calculate Q statistics and T ² statistics for original data or variables obtained by performing nonlinear transformation or the like from the original data. Thus, by setting the constraint conditions for these, an optimum solution can be obtained within a range in which the correlation between the variables of the original data is maintained.

The plant operating condition optimization system according to claim 1 is measured by operating state data acquisition means for acquiring operating state data indicating the operating state of the plant measured by a plurality of sensors, and by a sensor provided in the plant. Or, obtained by the operation state data acquisition means, the operation indicator data acquisition means for acquiring the operation indicator data for evaluating the operation of the plant, obtained based on the operation state data acquired by the operation state data acquisition means, and the operation state data acquisition means. A set of measurement data in which the driving condition data and the driving index data obtained by the driving index data acquisition means are associated with each other based on a predetermined item, and the measurement data recording means for recording the measurement data in a data recording unit; , Based on a plurality of sets of measurement data recorded in the data recording unit, the driving state representing the driving state data side Regression model creation means for creating a regression model by performing a predetermined multivariate analysis using the number as an explanatory variable and the driving index variable representing the driving index data side as an objective variable. A component conversion means for converting the number of components into a smaller number of components than the original explanatory variable, a prediction means for predicting an objective variable from the components converted by the component conversion means, and a method corresponding to the component conversion means from the components An inverse transformation means for estimating explanatory variables, and an operation index variable optimization means for obtaining an operation state variable for optimizing the operation index variable based on the regression model created by the regression model creation means, Explanation when optimizing the evaluation function related to the objective variable predicted by the prediction means while satisfying the constraint condition related to the explanatory variable estimated by the inverse transformation means Seeking a number, characterized in that the optimum operating conditions the explaining variable value.
In this specification, the input to the prediction means or the prediction model is referred to as an explanatory variable, and the output from the prediction means or the prediction model is referred to as an objective variable.

  According to a second aspect of the present invention, in the driving condition optimizing system according to the first aspect, when there are a plurality of driving index data obtained by the driving index data acquiring means, the measurement data recording means is the driving state data acquiring means. And a plurality of driving index data obtained by the driving index data acquiring means are set as a set of measurement data related to each other based on a predetermined item, and the measurement data is recorded in the data recording unit. The regression model creating means is an operation that represents a driving state data side based on a plurality of sets of measurement data recorded in the data recording unit, and the set of measurement data may include a plurality of driving index data. Each regression model is analyzed by performing a predetermined multivariate analysis using the state variables as explanatory variables and the multiple driving index variables representing the driving index data as multiple objective variables. The regression model includes component conversion means for converting the explanatory variables into a number of components that are uncorrelated with each other and fewer than the original explanatory variables, and a plurality of target variables from the components converted by the component conversion means. Predicting means for predicting and inverse transform means for estimating explanatory variables from components by a method corresponding to the component transforming means, and the driving index variable optimizing means is created by the regression model creating means A plurality of objective variables predicted by the predicting means while satisfying the constraints on the explanatory variables estimated by the inverse converting means. An explanatory variable value at the time of optimizing one evaluation function based on each evaluation function is obtained, and the explanatory variable value is set as an optimum operating condition.

  According to a third aspect of the present invention, in the plant operating condition optimization system according to the second aspect, the one evaluation function is obtained by a predetermined weighted sum of the respective evaluation functions.

  According to a fourth aspect of the present invention, in the plant operating condition optimization system according to any one of the first to third aspects, the predetermined multivariate analysis is principal component regression.

  According to a fifth aspect of the present invention, in the plant operating condition optimization system according to any one of the first to third aspects, the predetermined multivariate analysis is performed by a partial least square method (PLS). It is characterized by being.

  The invention according to claim 6 is the plant operating condition optimization system according to any one of claims 1 to 5, wherein the evaluation function is a linear expression related to an objective function and the constraint condition is related to an explanatory variable. In the case of a linear expression, the optimization uses linear programming.

  The invention according to claim 7 is the plant operating condition optimization system according to any one of claims 1 to 5, wherein the evaluation function is a nonlinear expression related to an objective variable and the constraint condition is related to an explanatory variable. In the case of a non-linear expression, the optimization uses non-linear programming.

  The invention according to claim 8 is the plant operating condition optimization system according to any one of claims 1 to 7, wherein the optimum operating condition and the operating state data obtained by the operating index variable optimizing means are provided. Based on the driving state data acquired by the acquiring means, it further comprises a validity evaluating means for outputting the validity of the optimum driving condition so that it can be evaluated.

  The invention according to claim 9 is the operation condition optimization system for a plant according to any one of claims 1 to 8, wherein the plant is a customer's plant and is obtained by the operation index variable optimization means. Based on the optimal operating condition value mounting means for mounting optimal operating conditions on the customer's plant side, and the operation index data for evaluating the operation of the customer's plant in a predetermined period before and after mounting by the optimal operating condition value mounting means, the customer A profit calculating means for calculating the profit increase and / or cost saving effect of the customer, and a charge collection means for collecting the consideration based on the customer profit increase and / or the cost saving effect calculated by the profit calculation means. It is characterized by having.

  The invention according to claim 10 is the plant operating condition optimization system according to any one of claims 1 to 9, wherein the plant is a power plant, and the operating state data includes flow rate, pressure, It includes any one or more of temperatures, and the operation index data includes efficiency.

  The plant operating condition optimizing method according to claim 11 is a plant operating condition optimizing method for causing a computer to execute optimization of the plant operating condition, and shows the plant operating state measured by a plurality of sensors. An operation state data acquisition step for acquiring operation state data and an evaluation of the operation of the plant measured by a sensor provided in the plant or obtained based on the operation state data acquired in the operation state data acquisition step The operation index data acquisition step for acquiring the operation index data, the operation state data acquired in the operation state data acquisition step, and the operation index data obtained in the operation index data acquisition step are related based on predetermined items. A set of measurement data and recording the measurement data in the data recording unit Based on the step and a plurality of sets of measurement data recorded in the data recording unit, a predetermined multivariate with an operation state variable representing the operation state data side as an explanatory variable and an operation index variable representing the operation indicator data side as an objective variable A regression model creating step for performing analysis and creating a regression model, the regression model comprising: a component conversion unit that converts explanatory variables into components that are uncorrelated with each other and less in number than the original explanatory variables; and the component conversion A prediction unit that predicts an objective variable from the component converted by the unit, and an inverse conversion unit that estimates an explanatory variable from the component by a method corresponding to the conversion of the component conversion unit. In the regression model creation step, A driving index variable optimization step for obtaining an operating state variable for optimizing the driving index variable based on the created regression model, the explanation estimated by the inverse conversion means Requested explanation variable values in optimizing the evaluation function for the predicted target variable by the prediction unit while satisfying the constraint condition on the number, characterized in that the optimum operating conditions the explaining variable value.

  According to a twelfth aspect of the present invention, in the plant operating condition optimization method according to the eleventh aspect, when there are a plurality of operation index data obtained by the operation index data acquisition step, the measurement data recording step includes the operation state data. The driving state data acquired in the acquiring step and the plurality of driving index data obtained in the driving index data acquiring step are set as a set of measurement data related based on predetermined items, and the measurement data is stored in the data recording unit. And the step of creating the regression model is based on a plurality of sets of measurement data recorded in the data recording unit, wherein the set of measurement data may include a plurality of driving index data. The driving state variable to be represented is an explanatory variable, and a plurality of driving index variables representing the driving index data side are set to a plurality of target variables. Each regression model is created by performing a quantity analysis, and the regression model is converted by the component conversion unit that converts the explanatory variables into components that are uncorrelated with each other and smaller in number than the original explanatory variables. A prediction unit that predicts a plurality of objective variables from the component, and an inverse conversion unit that estimates the explanatory variable from the component by a method corresponding to the conversion of the component conversion unit, and the operation indicator variable optimization step includes: An operation state variable for optimizing an operation index variable is obtained based on each regression model created in the regression model creation step, and the prediction is performed while satisfying the constraint condition on the explanatory variable estimated by the inverse transform unit. Obtaining an explanatory variable value when optimizing one evaluation function based on each evaluation function relating to a plurality of objective variables predicted by the unit, and making the explanatory variable value an optimum operating condition. To.

  A thirteenth aspect of the present invention is the plant operating condition optimization method according to the twelfth aspect, wherein the one evaluation function is obtained by a predetermined weighted sum of the respective evaluation functions.

  The invention according to claim 14 is the plant operating condition optimization method according to any one of claims 11 to 13, wherein the predetermined multivariate analysis is principal component regression.

  The invention according to claim 15 is the plant operating condition optimization method according to any one of claims 11 to 13, wherein the predetermined multivariate analysis is performed by a partial least square method (PLS). It is characterized by being.

  The invention described in claim 16 is the plant operating condition optimization method according to any one of claims 11 to 15, wherein the evaluation function is a linear expression related to an objective variable and the constraint condition is related to an explanatory variable. In the case of a linear expression, the optimization uses linear programming.

  The invention according to claim 17 is the plant operating condition optimization method according to any one of claims 11 to 15, wherein the evaluation function is a nonlinear expression related to an objective variable and the constraint condition is related to an explanatory variable. In the case of a non-linear expression, the optimization uses non-linear programming.

  The invention according to claim 18 is a program for optimizing a plant operating condition for obtaining optimization of the operating condition of the plant, and obtains operating state data indicating the operating state of the plant measured by a plurality of sensors in a computer. Obtaining operation index data for evaluating the operation of the plant, measured by an operation state data acquisition step, measured by a sensor provided in the plant, or obtained based on the operation state data acquired in the operation state data acquisition step Driving index data acquisition step, a set of measurement data in which the driving condition data acquired in the driving condition data acquisition step and the driving index data obtained in the driving index data acquisition step are related based on a predetermined item, A measurement data recording step for recording the measurement data in a data recording unit; Based on the multiple sets of measurement data recorded in the section, the driving state variable representing the driving state data side is used as the explanatory variable, and the driving model variable representing the driving index data side is used as the objective variable to perform a predetermined multivariate analysis and the regression model is determined. Regression model creation step to create, wherein the regression model includes a component conversion unit that converts explanatory variables into a number of components that are uncorrelated with each other and less than the original explanatory variable, and a component converted by the component conversion unit A prediction unit that predicts the objective variable from the component, and an inverse conversion unit that estimates the explanatory variable from the component by a method corresponding to the conversion of the component conversion unit. The regression model created in the regression model creation step includes A driving index variable optimizing step for obtaining a driving state variable for optimizing the driving index variable based on satisfying a constraint condition relating to the explanatory variable estimated by the inverse transformation unit A plant operating condition optimization program for obtaining an explanatory variable value for optimizing the evaluation function related to the objective variable predicted by the prediction unit, and executing a step of setting the explanatory variable value as an optimal operating condition. It is characterized by being.

  According to a nineteenth aspect of the present invention, in the plant operating condition optimization program according to the eighteenth aspect, when there are a plurality of operation index data obtained by the operation index data acquiring step, the measurement data recording step includes the operation state data. The driving state data acquired in the acquiring step and the plurality of driving index data obtained in the driving index data acquiring step are set as a set of measurement data related based on predetermined items, and the measurement data is stored in the data recording unit. Recording, in the regression model creating step, based on the plurality of sets of measurement data recorded in the data recording unit and including a plurality of driving index data in one set of measurement data, The driving state variables to be represented are explanatory variables, and multiple driving index variables representing the driving index data side are set as multiple objective variables. Each regression model is created by performing a predetermined multivariate analysis. The regression model includes a component conversion unit that converts the explanatory variables into a number of components that are uncorrelated with each other and smaller than the original explanatory variable, and the component conversion unit. A prediction unit that predicts a plurality of objective variables from the converted components, and an inverse conversion unit that estimates explanatory variables from the components with a program corresponding to the conversion of the component conversion unit, and the driving index variable optimization In the step, an operation state variable for optimizing the operation index variable is obtained based on each regression model created in the regression model creation step, and satisfying the constraint condition on the explanatory variable estimated by the inverse transformation unit. , Obtaining an explanatory variable value when optimizing one evaluation function based on each evaluation function related to a plurality of objective variables predicted by the prediction unit, and operating the explanatory variable value for optimal operation Characterized by the matter.

  According to a twentieth aspect of the present invention, in the plant operating condition optimization program according to the nineteenth aspect, the one evaluation function is obtained by a predetermined weighted sum of the evaluation functions.

  The invention according to claim 21 is the plant operating condition optimization program according to any one of claims 18 to 20, wherein the predetermined multivariate analysis is principal component regression.

  The invention described in claim 22 is the plant operating condition optimization program according to any one of claims 18 to 20, wherein the predetermined multivariate analysis is performed by a partial least square method (PLS). It is characterized by being.

  The invention according to claim 23 is the plant operating condition optimization program according to any one of claims 18 to 22, wherein the evaluation function is a linear expression related to an objective variable and the constraint condition is related to an explanatory variable. In the case of a linear expression, the optimization uses linear programming.

  The invention according to claim 24 is the plant operating condition optimization program according to any one of claims 18 to 22, wherein the evaluation function is a non-linear expression related to an objective variable, and the constraint condition is related to an explanatory variable. In the case of a non-linear expression, the optimization uses non-linear programming.

The plant operating condition optimizing system according to claim 25 is a plant operating condition optimizing system, and operating state data acquiring means for acquiring operating state data indicating the operating state of the plant measured by a plurality of sensors; The operation index data acquisition means for acquiring the operation index data for evaluating the operation of the plant, which is measured by the sensor provided in the plant or obtained based on the operation state data acquired by the operation state data acquisition means And a set of measurement data in which the driving condition data acquired by the driving condition data acquisition means and the driving index data obtained by the driving index data acquisition means are related based on predetermined items, and the measurement data is Measurement data recording means for recording in the data recording unit, and a plurality of sets of measurement data recorded in the data recording unit. Based on the data, the driving state variable representing the driving state data side as an explanatory variable, the prediction model creating means for creating a driving model variable representing the driving index data side as an objective variable, and the measurement data, The explanatory variables that are correlated with each other are converted into components that are uncorrelated with each other and have a smaller number of components than the original explanatory variables, and are set by projecting the original explanatory variables onto the subspace. And calculating the square of the distance between an arbitrary point in the original explanatory variable space and the projection point on the subspace as a Q statistic, and calculating the projection point and the center of the subspace. Q statistic and T ² statistic calculation means for calculating the square of the normalized distance as a T ² statistic, and an upper limit value as a constraint condition in optimization from the Q statistic and the T ² statistic, respectively. Q series And means for determining the amount · T ² statistic upper limit setting unit, the explanatory variables to optimize objective variable based on the prediction model, said Q statistic and T ² statistic is less than the respective upper limit Is the at least one of the constraint conditions, and while maintaining the correlation between the original explanatory variables, an explanatory variable value for optimizing the evaluation function related to the objective variable is obtained, and this explanatory variable value is determined as the optimal operating condition. And operating condition optimizing means.

  A plant operating condition optimization system according to a twenty-sixth aspect is the plant operating condition optimization system according to a twenty-fifth aspect, wherein a principal component analysis is used as the component conversion means.

  The plant operating condition optimization system according to claim 27 is the plant operating condition optimization system according to claim 25, wherein a partial least square method is used as the component conversion means.

  The plant operating condition optimization system according to claim 28 is the plant operating condition optimization system according to claim 25, wherein the original operating state variable and the original operating state variable are subjected to nonlinear transformation as the prediction model. A variable regression model is used in which all variables combined with the new variable obtained in the above or a variable obtained by extracting an appropriate variable among them are used as explanatory variables and an operation index variable is used as an objective variable.

The plant operating condition optimization system according to claim 29 is the plant operating condition optimization method according to claim 28, wherein the Q statistic / T ² statistic calculating means includes the original explanatory variable and the prediction model. The Q statistic and the T ² statistic are calculated based on a variable including a part or all of the nonlinear term used for creation.

30. The plant operating condition optimization system according to claim 30 is the plant operating condition optimization system according to claim 28 or 29, wherein the predictive model is a new variable obtained by subjecting the original operating state variable to nonlinear transformation. Is used as an explanatory variable, and a principal component regression model having an operation indicator variable as an objective variable is used. The principal component calculated by the Q statistic / T ² statistic calculation means in the principal component regression model The Q statistic and the T ² statistic are calculated based on the analysis.

A plant operating condition optimization system according to claim 31 is the plant operating condition optimization system according to claim 28 or 29, wherein the prediction model is a new variable obtained by subjecting the original operating state variable to nonlinear transformation. Is used as an explanatory variable, and a partial least squares model with an operation indicator variable as a target variable is used, and the Q statistic / T ² statistic calculation means uses the partial least squares model as the prediction model. The Q statistic and the T ² statistic are calculated based on

  A plant operating condition optimizing system according to a thirty-second aspect is the plant operating condition optimizing system according to the twenty-fifth aspect, wherein a neural network is used as the prediction model.

  According to the plant operating condition optimization system and the like of the present invention, sensor data collected in a recording device from a plant connected to a computer via a network is transmitted remotely. The operation state data acquisition unit of the computer acquires an operation state indicating the operation state of the plant measured by a plurality of sensors in the plant. The driving index data acquiring unit acquires driving index data by separately measuring and evaluating with a sensor or an analyzer at a timing corresponding to the driving state data acquired by the driving state data acquiring unit. Or you may obtain | require the driving | operation parameter | index data which evaluate the driving | operation of a plant based on the driving | running state data acquired by the driving | running state data acquisition part. The measurement data recording unit is a set of measurements in which the driving state data acquired by the driving state data acquiring unit and the driving index data obtained by the driving index data acquiring unit are related based on a predetermined item (for example, acquisition timing). The measurement data is recorded in the measurement DB.

  Based on a plurality of sets of measurement data recorded in the measurement DB, the regression model creation unit uses driving state variables representing the driving state data side as explanatory variables, and driving indicator variables representing the driving index data side (assuming there are multiple). A regression model is created by performing a predetermined multivariate analysis using as a target variable. In the regression model, an explanatory variable such as temperature is uncorrelated with each other and converts to a smaller number of components than the original explanatory variable, and an objective variable such as concentration is predicted from the component converted by the component converting unit. A prediction unit, and an inverse conversion unit that estimates an explanatory variable such as temperature from the component by a method corresponding to the conversion method of the component conversion unit. The operation index variable optimization unit obtains an operation state variable that optimizes the operation index variable based on the regression model created by the regression model creation unit. More specifically, an explanatory variable for optimizing the evaluation function related to the objective variable predicted by the prediction unit while satisfying the constraint condition related to the explanatory variable estimated by the inverse transformation unit is obtained, and the explanatory variable value is determined as the optimal value of the plant. The operating condition.

  By using the regression model, it is possible to evaluate how the operating state variable that is an explanatory variable changes and how the driving index that is an objective variable changes. Here, instead of setting the explanatory variable directly, by setting the component (which is an intermediate variable), the estimated value of the explanatory variable of the regression model is calculated from the component by the inverse transformation unit, and the driving indicator is used using the prediction unit The estimated value of is calculated. Here, the explanatory variables (operating state variables) calculated as described above maintain the correlation that the operating state variables had in the original data (used to create the regression model). As described above, the explanatory variable, the objective variable are evaluated by simultaneously evaluating both the estimated value of the explanatory variable calculated from the components in consideration of the correlation between the variables (inverse transformation) and the estimated value of the objective variable. The value of the component can be determined so that the objective variable is optimized (maximized, minimized, or brought close to a desired predetermined value) while satisfying the constraints on As a result, variables that can be modeled cheaply and easily, and can be controlled and set by humans, without using physical models that require a great deal of effort to match actual physical quantities and non-linear optimization methods that complicate calculations. Providing a system for optimizing the operating conditions of a plant that can be implemented in the plant by using, and that can obtain the optimal operating conditions within the range that maintains the correlation between the variables indicated by the acquired data. There is an effect that can be done.

In addition, modeling is performed using actual operation data, and further, an optimum solution is searched for within the upper limit values of the Q statistic and the T ² statistic, so that the correlation between the operation state variables indicated by the data is maintained. In addition, the optimal solution can be obtained in the region where the original data existed in the multivariable space.

1 is a diagram illustrating a plant operating condition optimization system 10 according to a first embodiment of the present invention. It is a figure which shows measurement DB50 provided in the recording device 31. FIG. It is a figure which shows measurement DB50a in the case of taking a power plant as an example. 4 is a conceptual diagram for explaining a regression model 60. FIG. It is a figure explaining the flow of processing of a plant operating condition optimization method and a program. It is a figure which shows measurement DB50b which is an example of measurement DB50. It is a figure which shows measurement DB50c which is another example of measurement DB50. It is a figure which illustrates the data each used for a principal component analysis process and a principal component regression model creation process. It is a figure explaining the flow of a principal component analysis process and a principal component regression model creation process. It is a figure explaining the flow of a process of a partial least square method. It is a figure which shows the example 70 of a display which plots the optimal driving | running condition value 65 obtained in Examples 1-3 in the frequency distribution and scatter diagram of each driving | running state variable S-DATA, for example, and displays it on a screen. It is a figure which illustrates seven driving | running state variables S-DATA. It is a figure which illustrates three driving | operation parameter | index variables I-DATA. It is a figure which shows the time series graph 80 of seven driving | running state variables S-DATA. It is a figure which shows the time series graph 90 of three driving | operation parameter | index variables I-DATA. It is a figure which shows the obtained PLS model 60b. It is a figure which shows the result (optimization result 100) when weighting three driving parameter | index variables I-DATA based on the obtained PLS model 60b, and optimizing with a linear programming method. It is a figure which shows the example of a screen display for evaluation of the validity of the obtained optimal driving | running condition value 65. FIG. It is a figure which shows the operating condition optimization system 10 'in case the plant 1 grade | etc. Is a customer's plant. It is a figure which shows the invoice example 110 issued when the value collection part 38 collects value. It is a block diagram which shows the internal circuit 120 of the computer 30 which executes the computer program of this invention. It is a figure which shows the operating condition optimization system of the plant in Example 8 of this invention. It is a conceptual diagram illustrating a Q statistic and T ² statistics. It is explanatory drawing of the factor in Example 9. FIG. In Example 9, it is explanatory drawing of RMSE (square root of a root mean square error) with respect to each lot. It is explanatory drawing of the optimal conditions with respect to each factor of Example 9.

  Hereinafter, each embodiment will be described in detail with reference to the drawings.

FIG. 1 shows a plant operating condition optimization system 10 according to the first embodiment of the present invention.
In FIG. 1, symbols 1, 2, 3,..., N are plant 1, plant 2, plant 3,..., Plant N to be optimized (hereinafter abbreviated as “plant 1 etc.”). . As shown in FIG. 1, on the plant 1 etc. side, there are a production execution system (MES) 11 and a distributed control system (DCS) 12 that measure and control the plant 1 etc. In the plant 1 or the like, various sensor data measured by a sensor 13 (a flow meter, a pressure gauge, a thermometer, or the like) that measures an object is collected in the recording device 14.
As shown in FIG. 1, routers 21, 22, 23, and 24 (hereinafter abbreviated as “router 21 etc.”) are provided on the plant 1 etc. side, and each router 21 etc. is connected to the network 20. Has been. In FIG. 1, reference numeral 30 is a computer that optimizes operating conditions of the plant 1, 25 is a router provided on the computer 30 side and connected to the network 20, 31 is a recording device connected to the computer 30, and 40 is a computer 30. It is a functional block which shows the function of. As shown in FIG. 1, the sensor data collected in the recording device 14 is remotely transmitted to the computer 30 side via the router 21, the network 20, and the router 25 via the MES 11 and the DCS 12. Alternatively, the sensor data collected by the sensor 13 may be transmitted remotely directly to the computer 30 via the router 21, the network 20, and the router 25.

Next, each function of the computer 30 shown in the function block 40 of FIG. 1 will be described in order.
The operation state data acquisition unit (operation state data acquisition means) 41 in FIG. 1 acquires operation state data indicating the operation state of the plant 1 and the like measured by the plurality of sensors 13 in the plant 1 and the like. In general, a plurality of sensors 13 are installed in the plant 1 and the like, and the operation state data, which is data representing the operation state of the plant 1 and the like, is measured collectively or irregularly by these sensors 13. To the computer 30 side.
FIG. 2 shows a measurement data database 50 (data recording unit; hereinafter referred to as “measurement DB”) provided in the recording device 31. As shown in FIG. 2, the transmitted operation state data is stored in one row (one record) of the measurement DB 50 as one set for each acquired timing. For example, as shown in FIG. 2, in the operation state data S-DATA in which the acquisition timing column 51 of the measurement DB 50 is “3”, the sensor 1 column 52 is “10.6” and the sensor 2 column 52 is “9.3”. ", The sensor N column 54 is" 0.12. " The operation state data S-DATA acquired in a lump need not be acquired at the same time, and may be acquired with a certain delay for each sensor (variable), for example.

  The driving index data acquisition unit (driving index data acquisition unit) 42 in FIG. 1 is separately provided with a sensor 13 or an analyzer (not shown) at a timing corresponding to the driving state data S-DATA acquired by the driving state data acquisition unit 41. The driving index data I-DATA is acquired by performing measurement and evaluation in (1). Or you may obtain | require the driving | operation parameter | index data which evaluate the driving | running state of the plant 1 etc. based on the driving | running state data S-DATA acquired by the driving | running state data acquisition part 41. FIG. As shown in FIG. 2, driving index data I-DAT is also stored in each row of the driving state data S-DATA. For example, in the operation index data I-DATA in which the acquisition timing column 51 of the measurement DB 50 is “2”, the operation index 1 column 55 is “0.67” and the operation index K column 56 is “19.8”. When the operation index data I-DATA is not acquired on a one-to-one basis for all the operation state data S-DATA, and there is no operation index data I-DATA corresponding to some of the operation state data S-DATA There is also a possibility. For example, in the driving index data I-DATA whose acquisition timing column 51 of the measurement DB 50 is “3”, the driving index 1 column 55 is “0.67”, but the driving index K column 56 is blank. In that case, the data (for example, the line “3” in the acquisition timing column 51 in which the driving index K column 56 is blank) is not used except for the optimization target.

  In the following, a power plant is taken as an example of the plant 1 and the like, the operation state data S-DATA shall include any one or more of the flow rate, pressure and temperature in the power plant, and the operation index data I-DATA shall include efficiency. . However, the plant to which the operating condition optimization system 10 of the present invention is applied is not limited to a power plant, but may be a manufacturing plant or the like. In this case, the operating state data S-DATA is used. The driving index data I-DATA can be set as appropriate.

The measurement data recording unit (measurement data recording unit) 43 in FIG. 1 includes the operation state data S-DATA acquired by the operation state data acquisition unit 4 and the operation index data I-DATA obtained by the operation index data acquisition unit 42. Is a set of measurement data related to each other based on a predetermined item (for example, the acquisition timing shown in the acquisition timing column 51), and the measurement data is recorded in the measurement DB 50.
FIG. 3 shows the measurement DB 50a when a power plant is taken as an example. In FIG. 3, the same reference numerals and symbols as those in FIG. As shown in FIG. 3, a set of driving state data S-DATA and driving index data I-DAT is used as one record (row) of the measurement DB 50a, and a plurality of samples are stored using this one record as one sample. . For example, there is a power generation output field = “38.6” of the generator as the operation state data S-DATA of the sample whose measurement time field (acquisition timing field) 51 is “2009/5/12 1:00”, and the operation index data As I-DATA, there is a temperature internal efficiency 1 column (%) = “91.1” obtained based on the operation state data S-DATA.

The regression model creation unit (regression model creation) 44 of FIG. 1 uses the driving state variable representing the driving state data side as an explanatory variable based on a plurality of sets of measurement data recorded in the measurement DB 50, and the driving representing the driving index data side. A regression model is created by performing a predetermined multivariate analysis using an index variable (assuming there are multiple) as an objective variable.
FIG. 4 is a conceptual diagram for explaining the regression model 60. As exemplified in FIG. 4, temperature, pressure, flow rate, and the like are explanatory variables, z1, z2,..., Zn are components, and concentration is an objective variable. The regression model 60 is converted by a component conversion unit (component conversion unit) 46 that converts explanatory variables such as temperature into a number of components z1 that are uncorrelated with each other and smaller than the original explanatory variable, and the component conversion unit 46. A prediction unit (prediction unit) 47 that predicts an objective variable such as a concentration from the component z1 and the like, and an inverse conversion unit (inverse) that estimates an explanatory variable such as a temperature from the component z1 and the like by a method corresponding to the conversion method of the component conversion unit 46. Conversion means) 48.

In general, there are many explanatory variables and they are correlated with each other. Therefore, the component conversion unit 46 converts (aggregates / compresses) these explanatory variables into components z1 and the like that are uncorrelated with each other and smaller than the number of the original explanatory variables (intermediate variable) by multivariate analysis. This conversion generally combines the correlations of the explanatory variables with each other into uncorrelated components. The prediction unit 47 predicts the objective variable from the component 1 and the like converted as described above. As described above, the objective variable can be predicted from the explanatory variable by combining the component conversion unit 46 and the prediction unit 47.
The regression model 60 can represent (predict) an objective variable from the component z1 or the like by the prediction unit 47. However, when the component z1 or the like is given at the same time, the component z1 or the like can be represented by a method corresponding to the component conversion unit 46. The original explanatory variable corresponding to can be calculated (estimated) (inverted). This represents the value of the original explanatory variable that satisfies this correlation from the component z1 or the like since the component z1 or the like aggregates the correlation between the original explanatory variables.

  The driving index variable optimizing unit (driving index variable optimizing means) 45 in FIG. 1 obtains an operating state variable that optimizes the driving index variable based on the regression model 60 created by the regression model creating unit 44. More specifically, an explanatory variable value for optimizing the evaluation function related to the objective variable predicted by the prediction unit 47 while satisfying the constraint condition related to the explanatory variable estimated by the inverse transform unit 48 is obtained, and the explanatory variable value is determined as the plant variable. The optimum operating condition is 1 or the like.

By using the regression model 60, it is possible to evaluate how the driving state variable that is the explanatory variable changes and how the driving index that is the objective function changes. However, here, instead of setting the explanatory variable value directly, the estimated value of the explanatory variable of the regression model 60 is calculated from the component z1 etc. by the inverse transformation unit 48 by setting the component z1 etc. (intermediate variable). Then, the estimated value of the driving index is calculated using the prediction unit 47. Here, the explanatory variables (operating state variables) calculated as described above maintain the correlation that the operating state variables had in the original data (used to create the regression model 60).
As described above, by simultaneously evaluating both the estimated value of the explanatory variable calculated from the component z1 and the like in consideration of the correlation between the variables and the estimated value of the objective variable, the explanatory variable, Component values can be defined that optimize the objective variable (maximize, minimize, or approach a desired predetermined value) while satisfying constraints on the objective variable.

  In particular, when the regression model 60 is generally linear, the above-described evaluation function is a linear expression related to an objective variable (the objective function is a linear combination of driving indices), and the constraint condition is a linear expression related to an explanatory variable, Linear programming can be used as an optimization method.

For example, when the predetermined multivariate analysis described above is principal component analysis and principal component regression, P is a loading matrix for calculating the principal component t from the original explanatory variable x, and the regression coefficient is for calculating the objective variable y from the principal component t. Let the vector be Q. Then, the function of the component conversion unit 46 is principal component analysis, and is expressed as t = P ^T x from the explanatory variable x to the component (principal component score) t. Here, T indicates transposition. The function of the prediction unit 47 is principal component regression, which is expressed as y = Qt = QP ^T x from the principal component t to the objective variable y. These values are variables after being standardized to an average value of 0, a variance of 1 and the like.
The inverse transformation unit 48 represents the estimated value (inverse transformation value) of the explanatory variable as x = Pt, and the prediction unit 47 represents the estimated value of the objective variable y as Qt. Therefore, for example, when the objective variable (y =) Qt is maximized within the range of the upper and lower limit constraints for the explanatory variable x, the following Equation 1 is obtained.

Here, x ^L and x ^U are a lower limit and an upper limit (vectors) of the explanatory variables, respectively. When Formula 1 is applied to a general formulation of linear programming, Formula 2 below is obtained.

  Formula 2 is a general expression of linear programming, and can be solved by a known method (for example, see the two-stage simplex method described in Non-Patent Document 1).

  On the other hand, the objective function is not linear and is expressed by a non-linear function f related to the driving index (y = Qt), or the constraint condition is a non-linear function g related to the driving condition (x = Pt) and the driving index (y = Qt). In other words, in other words, when the evaluation function is a non-linear expression (eg, quadratic expression) related to the objective variable and the constraint condition is a non-linear expression b related to the explanatory variable, And can be solved by using a nonlinear optimization method (eg, quadratic programming) as an optimization method.

FIG. 5 illustrates a process flow of the plant operating condition optimization method and program. In FIG. 5, the portions denoted by the same reference numerals as those in FIGS. 1 and 2 indicate the same elements or functions, and thus description thereof is omitted.
As shown in FIG. 5, the data acquisition device 49 acquires operation state data S-DATA indicating the operation state of the plant 1 and the like measured by the plurality of sensors 13 (operation state data acquisition step). The data acquisition device 49 separately performs measurement and evaluation with the sensor 13 or an analyzer (not shown) at a timing corresponding to the driving state data S-DATA acquired in the driving state data acquisition step, thereby driving index data I. -Acquire DATA (operation index data acquisition step). Or you may obtain | require the driving | operation parameter | index data I-DATA which evaluates driving | operation of the plant 1 etc. based on the driving | operation data S-DATA acquired at the driving | running state data acquisition step. The data acquisition device 49 uses the operation state data S-DATA acquired in the operation state data acquisition step and the operation index data I-DATA obtained in the operation index data acquisition step as predetermined items (for example, the acquisition timing in FIG. 2). Based on the acquisition timing shown in the column 51, a set of related measurement data is recorded, and the measurement data is recorded in the measurement DB 50 in the recording device 31 (measurement data recording step). The functions of the data acquisition device 49 can be realized as a part of the functions of the computer 30 (plant operating condition optimization program), but the data acquisition device 49 can be realized as a computer different from the computer 30. Is preferred.

FIG. 6 shows a measurement DB 50 b that is an example of the measurement DB 50.
As shown in FIG. 6, the sensor data (S-DATA indicated by the pressure 1 column 50b-2, the flow rate 1 column 50b-3, the temperature 1 column 50b-4, etc.) measured at each measurement point is measured at the measurement time. It is acquired at the same time (every predetermined item) shown in the column 50b-1 and stored together as one row, which is generally acquired and accumulated at every measurement time of a fixed sampling period. As the driving index data I-DATA at each time, for example, the efficiency indicated in the efficiency column 50b-5 is calculated by the driving index data acquisition unit 42. When the driving index data I-DATA is not efficient, the driving index data I-DATA is separately measured by the sensor 13 or the like.

  In FIG. 6, since each row is linked by the measurement time (predetermined item), an example of data at the same time is shown. However, for example, linking can be performed not at the same time, such as operating condition data linked to one lot (predetermined item). FIG. 7 shows a measurement DB 50 c that is another example of the measurement DB 50. As shown in FIG. 7, sensor data (S-DATA indicated by measurement time 1 column 50c-2, measurement time 2 column 50c-3, etc.) measured at each measurement point is stored in lot number column 50c-1. The same lot number shown (for each predetermined item) is collectively stored as one line and accumulated. The batching process may be performed as a batch process during or at the end of sensor data acquisition. For example, the efficiency indicated in the efficiency column 50b-5 is measured or calculated as the operation index data I-DATA in each lot. In the above, efficiency is shown as an example of the operation index. However, for example, in the case of a power generation plant, the power generation amount may be used, and in the case of a semiconductor manufacturing plant, quality such as a film thickness in a semiconductor generation process may be used.

  Returning to FIG. 5, the regression model creation unit 44 (multivariate analysis engine) uses the driving state variables representing the driving state data side as explanatory variables based on a plurality of sets of measurement data recorded in the measurement DB 50 or the like, and driving index data. Predetermined multivariate analysis is performed using the driving indicator variable representing the target as the objective variable, and a regression model 60 is created and recorded as the regression model 60 in the recording device 31 (regression model creation step). As shown in FIG. 4, the regression model 60 includes a component conversion unit 46 that converts the explanatory variables into a number of components that are uncorrelated with each other and less than the original explanatory variable, and a component converted by the component reaction unit 46. It has the prediction part 47 which estimates an objective variable, and the inverse conversion part 48 which estimates an explanatory variable from a component by the method corresponding to the conversion of the component conversion part 46. FIG.

  The driving index variable optimizing unit 45 (optimization engine) optimizes the driving index variable based on the regression model 60 created in the regression model creating step (optimum driving condition value 65, which is the optimum driving condition). (Operation indicator variable optimization step). Specifically, an explanatory variable value for optimizing the evaluation function related to the objective variable predicted by the prediction unit 47 while satisfying the constraint condition regarding the explanatory variable estimated (inverse conversion) by the inverse conversion unit 48 is obtained, and the explanatory variable The value is set to the optimum operating condition value 65. Here, the constraint conditions and the like can be given from the input unit 34 such as a keyboard and a mouse of the computer 30, and the optimum operating condition value 65 is output (displayed) to the output unit 33 such as a display of the computer 30, Or expert) 35.

Next, processing when the above-described predetermined multivariate analysis is principal component regression will be described.
FIG. 8 illustrates data matrices used for the principal component analysis process and the principal component regression model creation process, respectively. In FIG. 8, portions denoted by the same reference numerals as those in FIGS. 2 and 6 indicate the same elements, and thus the description thereof is omitted. As shown in FIG. 8, the two-dimensional data X is only net data excluding the measurement time indicated in the measurement time column 50b-1 and the variable name indicated in each variable column 50b-2 in the sensor data S-DATA. The two-dimensional data Y is composed of only net data excluding the measurement time indicated in the measurement time column 50b-1 and the index name indicated in each indicator column 50b-5 in the driving indicator data I-DATA. ing.
FIG. 9 is a diagram for explaining the flow of principal component analysis processing and principal component regression model creation processing. In FIG. 9, the portions denoted by the same reference numerals as those in FIGS. 2 and 6 indicate the same elements, and thus description thereof is omitted. As shown in FIG. 9, from the measurement data stored in the measurement DB 50 or the like, the sample × variable two-dimensional data X regarding the operating conditions is extracted and the principal component analysis process 44a is performed to perform the loading matrix P and the principal component t. Get. Next, two-dimensional data Y about the driving index is extracted from the measurement data stored in the measurement DB 50 and the like, and a principal component regression model creation process 45a is performed using Y and the principal component t obtained by the principal component analysis process 44a. To obtain a regression matrix Q. Each row of the data matrix X for creating the regression model 60a (principal component analysis) shown in Equation 4 below corresponds to a sample, and each column corresponds to a variable. Here, the number of samples (matrix) is M, and the number of variables (number of columns) is N.

The loading matrix (model coefficient matrix) P obtained by the principal component analysis process 44a and the regression matrix Q obtained by the principal component regression process 44b are both stored in the recording device 31 as the regression model 60a. In the following, the NIPALS (Nonlinear Iterative Partial Least Squares) algorithm is used as the principal component analysis (for example, see Non-Patent Document 2 described above).
The principal component analysis model coefficient P is a matrix of N rows and A columns where the number of principal components is A. As the number of principal components A, for example, the cumulative factor contribution ratio described in Non-Patent Document 2 is set to a value that is 80%. The principal component analysis model coefficient P is obtained as follows from 1) to 7).

1) An appropriate initial value is given to p. Here, the transpose of the row (horizontal vector) having the maximum variance among the rows of the data matrix X is set as the initial value of p.
2)

３）前回のスコアｔ^{（ｋ−１）}と現在のスコアｔ^（ｋ）とを比較し、

3) Compare the previous score t ^(k-1) with the current score t ^(k) ,

なら収束であり６）へ進む。
４）

If so, it is converged and the process proceeds to 6).
4)

５）

5)

として２）へ戻る。
６）次の主成分の計算が必要ならＸ＝Ｘ−ｔｐ^Ｔとして１）へ戻って繰り返す。不要なら終了する。
７）以上のようにして１）〜６）の処理をＡ個の縦ベクトルｐが得られるまで行い、得られた縦ベクトルｐを順に左から並べて主成分分析モデル係数Ｐを得る。

Return to 2).
6) If calculation of the next principal component is necessary, return to 1) and repeat as X = X−tp ^T. Exit if not needed.
7) The processes of 1) to 6) are performed as described above until A vertical vectors p are obtained, and the obtained vertical vectors p are sequentially arranged from the left to obtain the principal component analysis model coefficient P.

Before explaining how to determine the regression matrix Q, the principal component analysis will be explained in a little more detail. Principal component axis is obtained as the sum of squares S of the scores t _i becomes maximum. Since the square sum S is a function of the loading coefficient p _i , it is necessary to determine the loading coefficient p _i that maximizes the square sum S. However, since it is necessary to satisfy the constraint condition that the sum of squares of the loading coefficients p _i = 1, Lagrange's undetermined multiplier method is used. In this case, λ is an undetermined multiplier, G (p ₁ , p ₂ , λ) (for convenience of explanation, the data is two-dimensional) is defined, and this G (p ₁ , p ₂ , λ) is maximized. p ₁ , p ₂ and λ are determined. An equation obtained by partial differentiation of G (p ₁ , p ₂ , λ) with respect to p ₁ and p ₂ is expressed as Equation 9 below.

Here, p is a vertical vector in which p ₁ and p ₂ are arranged. Equation 9 indicates that the eigenvector of X ^T X is p and its eigenvalue is λ.
Next, how to determine the regression matrix Q (principal component regression model creation processing 45a in FIG. 9) will be described. Now, assuming that Y is one variable y, let T be a matrix in which the above-mentioned principal component score t (vertical vector) is arranged from the first principal component to the A principal component, and b be a regression vector of A rows and 1 column. The principal component regression model is expressed by Equation 10 below.

  Here, y is a vertical vector of A rows and 1 column. B in Expression 10 is obtained as Expression 11 below.

In Equation 11, Λ is a diagonal matrix having the above eigenvalue λ as an element. Since the regression vector Q is 1 row and A column, it is obtained as Q = b ^T. When y is a multi-variable (= R), b is a regression matrix B with A rows and R columns, y is a matrix Y with A rows and R columns, and Y = TB as in Equation 10. In this case as well, B is obtained by using B = Λ ⁻¹ T ^T Y in the same manner as Equation 11. Regression matrix Q since a R rows A column is determined as Q = ^{B T.}

Using the coefficient P of the principal component regression model obtained as described above and the regression matrix Q, for example, in the case of maximization of the driving index estimated value y = Qt, optimization is performed as shown in Equation 12 below. It is formulated as a problem (linear programming) (processing of the operating condition optimization 46a in FIG. 9).
As shown in FIG. 9, constraints and the like can be given from the input unit 34 such as a keyboard and a mouse of the computer 30, and the optimum operating condition value X ^opt 65 a is output (displayed) to the output unit 33 such as a display of the computer 30. ) To the watcher (or expert) 35.

Note that Formula 12 is a formulation for the variable after being standardized so as to have an average value of 0 and a variance of 1 with respect to the original variable. The optimal solution obtained by the formulation after standardization is defined as t ^opt, and the estimated value of the operation state variable (optimum operation condition value 65a after standardization) and the operation index (optimum operation index after standardization) obtained from this are obtained. ^Are expressed as x ^opt = Pt ^opt and y ^opt = Qt ^opt , respectively, to represent the values in the original unit system before standardization, as shown in the following formulas 13 and 14, and when mounted in an actual plant: These X ^opt and Y ^opt are used.

x ^opt : Optimal condition with standardized variable (unit system) X ^opt : Optimal condition with variable (unit system) before standardization y ^opt : Optimal operation index with standardized variable (unit system) Y ^opt : optimal operation indicators in before being standardized variable (unit system) X _m: X mean X _s: X standard deviation Y _m: Y average Y _s: Y standard deviation

Here, X and Y are vectors because they are composed of a plurality of variables, and their average values and standard deviations are also vectors. Ys. In the product of vectors such as y ^opt , “.” is shown between them to indicate a vector obtained as the product of the elements of the vector.

As described above, according to the first embodiment of the present invention, sensor data collected by the recording device 14 from the plant 1 or the like connected to the computer 30 via the network 20 is transmitted remotely. The operation state data acquisition unit 41 of the computer 30 acquires operation state data S-DATA indicating the operation state of the plant 1 and the like measured by the plurality of sensors 13 in the plant 1 and the like. The operation index data acquisition unit 42 obtains operation index data I-DATA for evaluating the operation of the plant 1 and the like based on the operation state data S-DATA acquired by the operation state data acquisition unit 41. The measurement data recording unit 43 uses the driving condition data S-DATA acquired by the driving condition data acquiring unit 41 and the driving index data I-DATA obtained by the driving index data acquiring unit 42 as predetermined items (for example, acquisition timing). The measurement data is recorded in the measurement DB 50 as a set of measurement data associated with each other based on the acquisition timing shown in the column 51. Based on a plurality of sets of measurement data recorded in the measurement DB 50, the regression model creation unit 44 uses the driving state variable representing the driving state data as an explanatory variable and the driving indicator variable representing the driving index data (assuming there are a plurality of driving index variables). A regression model 60 is created by performing a predetermined multivariate analysis with the objective variable as a target variable. The regression model 60 includes a component conversion unit 46 that converts explanatory variables such as temperature into a number of components z1 and the like that are uncorrelated with each other and less than the original explanatory variable, and a concentration from the component z1 and the like converted by the component conversion unit 46 A prediction unit 47 that predicts an objective variable such as a component, and an inverse conversion unit 48 that estimates an explanatory variable such as a temperature from the component Z1 or the like by a method corresponding to the conversion method of the component conversion unit 46. In general, there are many explanatory variables and they are correlated with each other. Therefore, the component conversion unit 46 converts (aggregates / compresses) these explanatory variables into components z1 and the like that are uncorrelated with each other and smaller than the number of the original explanatory variables (intermediate variable) by multivariate analysis. This conversion generally combines the correlations of the explanatory variables with each other into uncorrelated components. The prediction unit 47 predicts an objective variable from the component z1 and the like converted as described above.
As described above, the objective variable can be predicted from the explanatory variable by combining the component conversion unit 46 and the prediction unit 47. The regression model 60 can represent (predict) an objective variable from the component z1 or the like by the prediction unit 47. However, when the component z1 or the like is given at the same time, the component z1 or the like can be represented by a method corresponding to the component conversion unit 46. The original explanatory variable corresponding to can be calculated (estimated) (inverted). This represents the value of the original explanatory variable that satisfies this correlation from the component z1 or the like since the component z1 or the like aggregates the correlation between the original explanatory variables. The driving index variable optimizing unit 45 obtains an operating state variable that optimizes the driving index variable based on the regression model 60 created by the regression model creating unit 44. More specifically, an explanatory variable value for optimizing the evaluation function related to the objective variable predicted by the prediction unit 47 while satisfying the constraint condition related to the explanatory variable estimated by the inverse transform unit 48 is obtained, and the explanatory variable value is determined as the plant variable. The optimum operating condition is 1 or the like.

By using the regression model 60, it is possible to evaluate how the driving state variable that is the explanatory variable changes and how the driving index that is the objective function changes. However, here, instead of setting the explanatory variable directly, the estimated value of the explanatory variable of the regression model 60 is calculated from the component z1 etc. by the inverse transformation unit 48 by setting the component z1 etc. (intermediate variable). Then, the estimated value of the driving index is calculated using the prediction unit 47. Here, the explanatory variables (operating state variables) calculated as described above maintain the correlation that the operating state variables had in the original data (used to create the regression model 60).
As described above, by simultaneously evaluating both the estimated value of the explanatory variable calculated from the component z1 and the like in consideration of the correlation between the variables and the estimated value of the objective variable, the explanatory variable, Component values can be defined that optimize the objective variable (maximize, minimize, or approach a desired predetermined value) while satisfying constraints on the objective variable. As a result, variables that can be modeled cheaply and easily, and can be controlled and set by humans, without using physical models that require a great deal of effort to match actual physical quantities and non-linear optimization methods that complicate calculations. Providing a system for optimizing the operating conditions of a plant that can be implemented in the plant by using, and that can obtain the optimal operating conditions within the range that maintains the correlation between the variables indicated by the acquired data. can do.

  In the second embodiment, a case where the predetermined multivariate analysis is a partial least square method (Partial Least Squares: PLS) will be described. The partial least square method can be formulated in the same manner as the principal component regression, and can be solved by linear programming. Below, the algorithm of a nonpatent literature 2 is used as a partial least squares algorithm.

FIG. 10 illustrates a processing flow of the partial least square method. In FIG. 10, the portions denoted by the same reference numerals as those in FIGS. 2 and 6 indicate the same elements, and thus description thereof is omitted.
As shown in FIG. 10, from the measurement data stored in the measurement DB 50 of the recording device 31, for the data for creating the PLS model 60 b of the normal part, a sample of explanatory variables shown in Equation 15 × two-dimensional variable A data matrix X and a target variable sample × 1 variable data vector y are extracted and input.

As shown in FIG. 10, in the partial least square method 44b, partial least square coefficients P _pls , W, and Q _pls are calculated from the input data matrix X and data vector y, and these are used as the PLS model 60b. It is stored in the recording device 31. The above relationship is shown in the following mathematical expressions 16-18. Here, when converting N types of data into A types of components, T _pls is a matrix of M rows and A columns of latent variables, which is a linear combination of explanatory variables, and P _pls is N rows A of loading coefficients. A matrix of columns, Q _pls is a 1 _× A horizontal vector, and W is an N × A matrix called PLS weight.

  The PLS model in the case of A = 1 component is composed of the following mathematical formulas 19 and 20.

  As shown in Equations 19 and 20, in the PLS model, a part of the data matrix X is realized through the latent variable t and used for modeling the data vector y. The latent variable t is as shown in Equation 21.

  Here, w is a PLS weight vector (vertical vector of N rows and 1 column), and its norm is 1 as shown in Equation 22.

  The PLS model construction standard is that the correlation between the data vector y and the latent variable t is increased, and at the same time, the variance of the latent variable t is increased, and a large amount of information contained in the data matrix X is used. Therefore, an inner product S that is a covariance between the data vector y and the latent variable t expressed by Equation 23 is considered, and a case where S is maximum under the constraint condition of Equation 22 is considered as an optimal PLS model.

  For this reason, Lagrange's undetermined multiplier method is used. Refer to Non-Patent Document 2 though omitted in the middle. The optimum w that maximizes Equation 23 is Equation 24, and Equations 25 and 26 are obtained. The latent variable t is obtained from Equation 21 (t = Xw).

In the case of A = 2 components, the final prediction model equation can be expressed by Equations 28 and 29 described later. Also in the case of the partial least square method, as in the case of the principal component analysis, the latent variable matrix T _pls is arranged in descending order of variance from the left column. When the variances are accumulated in order from the left column, the ratio (cumulative factor) of the total to the sum of all variances (the sum of all variances in all columns matches in the data matrix X and the latent variable matrix T _pls ) A value that reaches 80% (referred to as a contribution rate) is considered to be appropriate as the number of latent variables, and this is adopted as the value of A, for example. Specifically, when ta is the a-th column of the latent matrix T ₂ and σ _ta is the standard deviation of ta, Equation 27 below is obtained.

  A PLS model expression of A = 2 components can be written as the following Expressions 28 and 29.

  Coefficients related to the first component are calculated by Equations 30 to 33.

  Since the second component is a component that is not related to the first component, the equations 28 and 29 are converted by the following equations 34 and 35.

  Therefore, the coefficients related to the second component can be obtained from the following mathematical formulas 36 to 39.

  When there are more than 3 components, it can be determined in the same manner as above. When generalized, the PLS algorithm becomes the following 1) to 8) (Equations 40 to 49).

２）

３）

４）

５）

６）

７）

８）次のＰＬＳ成分が必要なら３）へ戻って繰り返し、不要なら終了する。 1)

2)

3)

4)

5)

6)

7)

8) If the next PLS component is necessary, return to 3) and repeat, if not, the process ends.

When the number of components is A, P _pls is N rows and A columns, Q _pls is a horizontal vector of 1 rows and A columns, and W is a matrix of N rows and A columns. For example, in the case of maximization of the driving index estimated value y = Q _plst using the coefficients P _pls and Q _{pls of} the PLS model obtained as described above, optimization is performed as in the following formula 50. It can be formulated as a problem (linear programming) (operating condition optimization 45b in FIG. 10).
As shown in FIG. 10, the constraint conditions and the like can be given from the input unit 34 such as a keyboard and a mouse of the computer 30, and the optimum operating condition value X ^opt 65 b is output (displayed) to the output unit 33 such as the display of the computer 30. ) To the watcher (or expert) 35.

As in the case of the principal component regression described in the first embodiment, what is described here is an optimization for variables after standardization to a mean value of zero and a standard deviation of 1, which is the original unit system. When returning to a variable value, the same processing as described above is performed. Specifically, the optimum solution obtained by the formulation after standardization is defined as t ^opt, and the operation state variable obtained at this time (optimum operation condition value 65b after standardization) and the operation index (optimum operation index after standardization) When the estimated values of x ^opt = P _pls t ^opt and y _opt = Q _pls t ^opt are expressed by the values in the original unit system before standardization, the following formulas 13 and 14 (reprinted) are used: Therefore, X ^opt and Y ^opt are used when mounting in an actual plant.

[Equation 13]
Y ^opt = Y _s . y ^opt + Y _m
[Formula 14]
X ^opt = X _s . x ^opt + X _m

x ^opt : Optimal condition with standardized variable (unit system) X ^opt : Optimal condition with variable (unit system) before standardization y ^opt : Optimal operation index with standardized variable (unit system) Y ^opt : optimal operation indicators in before being standardized variable (unit system) X _m: X mean X _s: X standard deviation Y _m: Y average Y _s: Y standard deviation where X, Y are a plurality of variables Therefore, it is a vector, and its average value and standard deviation are also used as vectors. Ys. In the product of vectors such as y ^opt , “.” is shown between them to indicate a vector obtained as the product of the elements of the vector.

The objective function is not linear and is expressed by a non-linear function f related to the driving index (y = Q _pls t), or the constraint condition is non-linear regarding the driving condition (x = P _pls t) and the driving index (y = Q _pls t). In other words, when the evaluation function is a non-linear expression (eg, quadratic expression) relating to the objective variable and the constraint condition is a non-linear expression b relating to the explanatory variable, It can be solved by using a nonlinear optimization method (eg, quadratic programming) as an optimization method.

As described above, according to the second embodiment of the present invention, even when the predetermined multivariate analysis is the partial least square method PLS, it can be formulated in the same manner as the principal component regression, and is solved by linear programming. be able to. In the partial least square method, from the measurement data stored in the measurement DB 50 or the like of the recording device 31, the data for creating a normal portion of the PLS model is expressed by the sample of explanatory variables shown in Equation 15 × the variable two-dimensional data matrix. A target variable sample X and a data vector y of one variable are taken out and input. The partial least squares P _pls , W, and Q _pls are calculated from the input data matrix X and data vector y, and are stored in the recording device 31 as the PLS model 60b. N types of data are converted into A types of components. Since it is considered that a cumulative factor contribution rate of 80% is appropriate as the number of latent variables, for example, this is adopted as the value of A. The PLS algorithm is from 1) to 8) described above. When the number of components is A, P _pls is N rows and A columns, Q _pls is a horizontal vector of 1 rows and A columns, and W is a matrix of N rows and A columns. For example, in the case of maximizing the driving index estimated value y = Q _plst using the coefficients P _pls and Q _{pls of} the PLS model obtained as described above, the optimization problem as shown in the above equation 50 It can be formulated as (linear programming). As a result, even when the predetermined multivariate analysis is a partial least squares PLS, use a physical model that requires a lot of effort to match the actual physical quantity and a nonlinear optimization method that complicates the calculation. It can be modeled inexpensively and easily, and can be implemented in the plant by using variables that can be controlled and set by humans, within the range that maintains the correlation between the variables indicated by the acquired data. It is possible to provide a plant operating condition optimization system or the like that can obtain an optimal operating condition.

  In the third embodiment, the case where there are a plurality of driving index data obtained by the driving index data acquisition unit 42 of FIG. 1 will be described in more detail. In this case, the measurement data recording unit 43 uses the driving condition data S-DATA acquired by the driving condition data acquiring unit 41 and the plurality of driving index data I-DATA obtained by the driving index data acquiring unit 42 as predetermined items. And the measurement data is recorded in the measurement DB 50 or the like.

  The regression model creation unit 44 represents the driving state data side based on a plurality of sets of measurement data recorded in the measurement DB 50 and the like, and a set of measurement data may include a plurality of driving index data I-DATA. Each regression model 60 is created by performing a predetermined multivariate analysis using the driving state variables as explanatory variables and a plurality of driving index variables representing the driving index data as a plurality of objective variables. As described in the first embodiment, the regression model 60 is converted by the component conversion unit 46 that converts the explanatory variables into components z1 and the like that are not correlated with each other and smaller than the original explanatory variable, and the component conversion unit 46. A prediction unit 47 that predicts a plurality of objective variables from the part z1 and the like, and an inverse conversion unit 48 that estimates (reverses) the explanatory variable from the component z1 and the like by a method corresponding to the component conversion unit 46.

  The driving index variable optimizing unit 45 obtains an operating state variable that optimizes the driving index variable based on the regression model 60 created by the regression model creating unit 44. Specifically, one evaluation function based on each evaluation function related to a plurality of objective variables predicted by the prediction unit 47 is considered while satisfying the constraint condition related to the explanatory variable estimated by the inverse conversion unit 48. An explanatory variable value at the time of optimization is obtained, and this explanatory variable value is set as the optimum operating condition 65. Here, the one evaluation function is preferably obtained by a predetermined weighted sum of each evaluation function.

  As described above, there are multiple objective variables (operational indicators) to be optimized, parts that simultaneously optimize the evaluation function for each operation indicator (if you want to optimize each evaluation function), One evaluation function to be evaluated as a whole is calculated. Specifically, the weighted sum for each evaluation function is set as one new evaluation function. For example, when the driving index is efficiency and there are a plurality of types of efficiency and these are maximized at the same time, the evaluation function is the efficiency as the driving index itself and is represented by y = Qt. A new evaluation function considering a plurality of indices is a weighted sum for each element of y. If the weights for each are w1, w2,..., Wn, and the horizontal vector in which these are arranged is w = [w1, w2,..., Wn], the weighted sum of the driving indices is expressed by the following formula 52. .

  Therefore, the optimization problem is expressed by the following formula 53.

  If the evaluation function for each objective variable is represented by some function of the objective variable, the weighted sum is as shown in the following equation 54.

  The weight w = [w1, w2,..., Wn] can be determined as follows, for example. In the PLS model 60b such as the regression model 60 created in the first and second embodiments, when applying the principal component regression and the partial least square method, each variable (operating state data x, driving index data y) is generally averaged. Each of the above models is created after standardization to a value of 0 and a variance (standard deviation) of 1. Here, when there are a plurality of driving indexes and y is a vector, each element is standardized to an average value of 0 and a variance of 1. Therefore, when weighting these ys, for example, if all of them are set to 1, they are equally weighted and optimized. If there is something to be emphasized in y, the weight may be set to 10 times or the like. As described above, by setting the weighted sum of a plurality of evaluation functions as an evaluation function in optimization and solving with an optimization method, it is possible to obtain the optimum operating condition value 65 in which each of the plurality of objective variables is considered by weight. it can.

  In general, when there are a plurality of regression models 60, constraint conditions, and objective variables (evaluation functions) obtained by the regression model creation unit 44 or the like, the driving index variable optimization unit 45 (see FIG. 5) assigns a weight ( When there are a plurality of objective variables, the above w) is given from the input unit 34 such as a keyboard and a mouse of the computer 30 and is optimized by an optimization method such as linear programming. The information can be output (displayed) to the output unit 33 such as a display and shown to the monitoring person (or expert) 35.

As described above, according to the third embodiment of the present invention, there are a plurality of driving index data obtained by the driving index data acquisition unit 42 (a plurality of target variables (driving indices) to be optimized), and each driving When simultaneously optimizing evaluation functions for indices (when it is desired to optimize each evaluation function), the evaluation functions are combined to calculate one evaluation function to be evaluated as a whole.
Specifically, the weighted sum for each evaluation function is set as one new evaluation function. By setting a weighted sum of a plurality of evaluation functions as an evaluation function in optimization and solving with an optimization method, it is possible to obtain an optimum operating condition value in which each of the plurality of objective variables is considered with a weight. As a result, whether the given multivariate analysis is principal component regression or partial least squares PLS, a nonlinear model that requires a great deal of effort to match actual physical quantities and nonlinear optimization Can be modeled cheaply and easily without using a method, and can be implemented in a plant by using variables that can be controlled and set by humans. It is possible to provide a plant operating condition optimization system or the like that can obtain an optimal operating condition within the range maintained.

  In the fourth embodiment, in the operation condition optimization system 10 such as the plant 1 described in the first to third embodiments, the optimum operation condition value 65 obtained by the operation index variable optimization unit 45 and the operation state data obtaining unit 41 are obtained. A validity evaluation unit (validity evaluation means) (not shown) that outputs the validity of the optimum operation condition value 65 so as to be able to be evaluated based on the obtained operation state data S-DATA can be further provided.

FIG. 11 shows a screen (display) by plotting the optimum operation condition value 65 (for example, optimum operation condition value 65b) obtained in Examples 1 to 3 on, for example, the frequency distribution and scatter diagram of each operation state variable S-DATA. A display example 70 displayed on the output unit 33) is shown. 11, the control variable list is an example of each operation state variable S-DATA, and the circle numbers 1, 2, 3,... In the scatter diagram list are numbers assigned from the top of the control variable list. The diagonal elements of the scatter chart list are frequency distributions of the respective control variables, and the elements in the horizontal direction are scatter charts for a certain control variable and other control variables.
For example, the (1, 1) element of the scatter chart list is the frequency distribution of the power generation output, the (1, 2) element is a scatter chart when the vertical axis is the power generation output and the horizontal axis is the condensed water flow rate, (1, 3) The elements are scatter plots where the vertical axis is the power generation output and the horizontal axis is the main steam flow rate (the same applies hereinafter). In FIG. 11, the optimum operating condition value 65b is indicated by a square or circle with black inside, and the actual point close to the optimum operating condition value 65b, which is the calculation result, in the actual data (data used for creating the PLS model 60b). Is indicated by a square or circle with an oblique line inside, and the operating condition indicating the best driving index in the actual data is indicated by a square or circle with the inside being gray. As shown in FIG. 11, the actual validity of the obtained optimum operating condition value 65b can be evaluated based on the actual point close to the optimum operating condition value 65b and the optimum operating condition value 65b.

  As described above, according to the fourth embodiment of the present invention, in the operation condition optimization system 10 such as the plant 1 described in the first to third embodiments, the optimum operation condition value obtained by the operation index variable optimization unit 45 is obtained. Based on the operation state data S-DATA acquired by 65b and the operation state data acquisition unit 41, a validity evaluation unit that outputs the validity of the optimum operation condition value 65b so as to be evaluated can be further provided. As shown in FIG. 11, the actual validity of the obtained optimum operating condition value 65b can be evaluated based on the actual point close to the optimum operating condition value 65b and the optimum operating condition value 65b.

  As described above, the power plant is taken as an example of the plant 1, etc., and the operation state data S-DATA includes any one or more of the flow rate, pressure, and temperature in the power plant, and the operation index data I-DATA indicates the efficiency. Included. However, the plant to which the operating condition optimization system 10 of the present invention is applied is not limited to a power plant, but may be a manufacturing plant or the like. In this case, the operating state data S-DATA is used. The driving index data I-DATA can be set as appropriate. Hereinafter, various variables in the case of a power plant will be exemplified.

In a certain power plant, a PLS model 60b representing three operation index data I-DATA is created for the following operation state variable S-DATA, thereby extracting an operation condition for optimizing the operation index. FIG. 12 illustrates seven driving state variables S-DATA, and FIG. 13 illustrates three driving index variables I-DATA.
FIG. 14 shows a time series graph 80 of seven driving state variables S-DATA, and FIG. 15 shows a time series graph 90 of three driving index variables I-DATA. FIG. 16 shows the obtained PLS model 60b. FIG. 17 shows the result (optimization result 100) when the three driving index variables I-DATA are weighted based on the obtained PLS model 60b and optimized by the linear programming method.
As shown in FIG. 17, the weights are [w1, w2, w3] = [1,1,1], [10,1,1], [1,1,10], [1,10,1]. It carried out about four ways. FIG. 18 shows a screen display example 70 ′ for evaluating the validity of the obtained optimum operating condition value 65b. FIG. 18 is a display example 70 in which one scatter diagram corresponding to the (2,1) element is extracted and displayed in the validity evaluation (scatter graph list) of the optimum operating condition value 65b shown in the display example 70 of FIG. 'Is.
In FIG. 18, the horizontal axis is the power generation output, the vertical axis is the condensed water flow rate, the rhombus is the actual data, the square is the high-efficiency operating condition, the circle is the actual data (similar data), and the triangle is the actual data (maximum efficiency). . As shown in FIG. 18, since both variables have a correlation with each other, it is necessary to set the optimum operating condition value 65b within a range in which this correlation is maintained. In FIG. 18, the optimum operating condition value 65b (square) is also plotted within the range of the existing data maintaining the correlation, and it can be seen that the obtained optimum condition also maintains the correlation. Compared to the data (triangle) that showed the highest efficiency among the existing data (diamonds), the conditions are very close, indicating that it matches the actual situation. That is, the optimum operating condition value 65b can be obtained within a range in which the correlation between the variables indicated by the actual data (diamonds) is maintained, and the actual validity of the obtained optimum operating condition value 65b is evaluated. Can do.

In the sixth embodiment, a business model in the case where the plant 1 or the like is a customer's plant in the above-described first embodiment or the like will be described. FIG. 19 shows an operating condition optimization system 10 ′ when the plant 1 or the like is a customer plant. In FIG. 19, the parts denoted by the same reference numerals as those in FIG. A customer plant 1 ′ in FIG. 19 corresponds to the plant 1 in FIG. 1, and a customer plant 2 ′ corresponds to the plant 2 in FIG. 1 (hereinafter also referred to as “customer plant 1 ′ etc.”). The MES 11 ′ in the customer plant 1 ′ corresponds to the MES 1 in FIG. 1, and the DCS 12 ′ corresponds to the DCS 12 in FIG. 1 (the same applies hereinafter).
As shown in FIG. 19, the following functions are added to the functional block 40 of the computer 30. That is, the optimum operating condition value mounting unit (optimum operating condition value mounting means) for mounting the optimum operating condition value 65 and the like (optimal operating condition) obtained by the operating index variable optimizing unit 45 on the customer's plant 1 'etc. side. 36 and the operation index data I-DATA that evaluates the operation of the customer plant 1 ′ and the like in a predetermined period before and after the mounting by the optimal operating condition value mounting unit 36, the profit increase effect and / or the cost saving effect of the customer is calculated. Profit etc. calculation part (profit etc. calculation means) 37, and a value collection part (value collection means) 38 for collecting the consideration based on the increase in the profit of the customer and / or the cost saving calculated by the profit etc. calculation part 37 I have.

  The optimum driving condition value mounting unit 36 shows the optimum driving condition value 65 and the like obtained by performing regression modeling / PLS modeling from the customer driving state data S-DATA and driving index data I-DATA, as shown in FIG. As described above, it is mounted on the customer control system (MES 11 ′, DCS 12 ′) via the remote network 20 or the like. The operation index (efficiency) data I-DATA for a certain period before and after this condition setting is accumulated, and based on this, the profit increase or cost saving effect of the customer is calculated, and a certain percentage is collected from the customer as a service fee.

  FIG. 20 shows an example bill 110 issued when the consideration collection unit 38 collects the compensation. As shown in the cumulative effect of FIG. 20, for example, in a power plant, it is estimated that fuel costs could be reduced by 148.8 million yen in one month with the same output, so 74,400 yen, which is 5%, was used as expenses. I am charging. In order to make it easier for customers to understand the cost savings, the graph shows a visual improvement in efficiency. The invoice 110 may be sent to the customer by mail, sent to the customer as electronic data (e.g., e-mail), or another computer connected to the computer 30 or network 20 and accessible from the customer. Via the web, the customer may be able to browse at any time. As described above, a business model using the operating condition optimization system 10 'of the present invention can be constructed.

FIG. 21 is a block diagram showing an internal circuit 120 of the computer 30 for executing the computer program (plant operation condition optimization program) of the present invention for realizing the above-described embodiments. As shown in FIG. 21, the CPU 121, ROM 122, RAM 123, image control unit 125, controller 126, input control unit 128, and external I / F (interface) unit 129 are connected to the bus 130. In FIG. 21, the above-described computer program of the present invention is recorded in a recording device 31 such as a ROM 122, an HD (hard disk), or a recording medium (including a removable recording medium) such as a DVD or CD-ROM 127. . The recording device 31 records the PLS model 60b such as the measurement DB 50 and the regression model 60 described above.
The computer program of the present invention is loaded into the RAM 123 from the ROM 122 via the bus 130 or from the recording device 31 or a recording medium such as a DVD or CD-ROM 127 via the controller 126 via the bus 130. The input control unit 128 is connected to the input operation unit 34 such as a mouse and a numeric keypad, and performs input control and the like. The VRAM 124, which is an image memory, has a capacity corresponding to the data capacity of at least one screen of the output unit 33 such as the display of the computer 30, and the image control unit 125 converts the data in the VRAM 124 into image data and outputs it. 33. In this way, the screen display function of the validity evaluation unit is executed. The external I / F unit 129 has an input / output interface function, whereby the computer 30 is connected to the router 25, the network 20, the plant 1, etc., the plant 1 ′, etc. As described above, the CPU 121 executes the computer program of the present invention, whereby the object of the present invention can be achieved.

FIG. 22 is a diagram illustrating a process flow in the plant operating condition optimization system according to the eighth embodiment of the present invention. In FIG. 22, portions denoted by the same reference numerals as those in FIG. 1, FIG. 2, FIG.
Similarly to the above, the data acquisition device 49 acquires operation state data S-DATA indicating the operation state of the plant 1 and the like measured by the plurality of sensors 13. The data acquisition device 49 separately performs measurement and evaluation with the sensor 13 or an analyzer (not shown) at a timing corresponding to the driving state data S-DATA acquired in the driving state data acquisition step, thereby driving index data I. -Get DATA. Or you may obtain | require the driving | operation parameter | index data I-DATA which evaluates driving | operation of the plant 1 etc. based on the driving | operation data S-DATA acquired at the driving | running state data acquisition step. The data acquisition device 49 uses the operation state data S-DATA acquired in the operation state data acquisition step and the operation index data I-DATA obtained in the operation index data acquisition step as predetermined items (for example, the acquisition timing in FIG. 2). Based on the acquisition timing shown in the column 51), the measurement data is recorded in the measurement DB 50 in the recording device 31 as a set of related measurement data. The functions of the data acquisition device 49 can be realized as a part of the functions of the computer 30 (plant operating condition optimization program), but the data acquisition device 49 can be realized as a computer different from the computer 30. Is preferred.
The measurement data recorded in the measurement DB 50 is, for example, as shown in FIG. In FIG. 7, the efficiency is often obtained from the operation state data by a physical formula or the like. For example, in the case of a power plant, the efficiency is obtained from a complicated mathematical expression such as enthalpy calculation and a table.

Measurement data for the past normal operation is input from the measurement DB 50 as model creation data to the prediction model creation means 61 of FIG. The prediction model creation means 61 creates a prediction model 60 c by the multivariate analysis engine using the model creation data and stores it in the recording device 31.
In addition, as described in claim 28, the prediction model 60c extracts the whole of the original explanatory variables and the new variables obtained by performing the non-linear processing, or some of the variables by a variable reduction method or the like. Thus, a multiple regression model can be used in which the explanatory variable of the prediction model is used and the operation index variable is the objective variable. Here, the variable reduction method is described, for example, in “About Variable Selection in Multiple Regression Analysis” by Juichiro Takeuchi (Japan APL Association, Proceedings of the 2004 Symposium).
Further, as described in claims 30 and 31, the prediction model 60 c uses the original explanatory variable and the new variable obtained by performing nonlinear processing as an explanatory variable of the prediction model, Thus, a principal component regression model or a partial least squares (PLS) model using the driving index variable as an objective variable can be used.
Further, as the prediction model 60c, as described in claim 32, a neural network constructed by learning original explanatory variables and driving index variables can be used.

On the other hand, the model creation data is also input to the component conversion means 62.
In general, there are many explanatory variables and they are correlated with each other. Therefore, the component conversion means 62 has fewer explanatory variables in the input model creation data than the original explanatory variables that are uncorrelated with each other by multivariate analysis such as principal component analysis or partial least squares. Component conversion is performed by consolidating and compressing into “components” as intermediate variables. Here, the “component” corresponds to the coordinates in the partial space in the original explanatory variable space, and this partial space constitutes a part of the original explanatory variable space. To convert the original explanatory variable into “component” is to project the original explanatory variable onto the subspace and obtain the coordinates of the subspace of the point, and the conversion coefficient matrix 62a corresponding to this projection is converted into the component conversion Means 62 are provided.

Although not shown in FIG. 22, the Q statistic / T ² statistic calculating means in claim 25 calculates the square of the distance between an arbitrary point in the original explanatory variable space and the projection point in the subspace. Is calculated as a Q statistic, and the square of the normalized distance between the projection point and the center of the subspace is calculated as a T ² statistic. These Q statistics and T ² statistics are zero or more values.
The “component” obtained by the component converting means 62 forms a subspace in the original explanatory variable space, and this partial space maintains the correlation that the explanatory variables of the model creation data have. Projection points on the subspace are correlated with each other in the explanatory variable space.
Here, a point obtained by projecting an arbitrary point (vector corresponding to the driving state) in the space of the explanatory variable onto the subspace is the subspace if the original explanatory variable is not included in the subspace. Since it is the point closest to the original explanatory variable above, it is an approximate point on the subspace of the original explanatory variable. The square of the distance between an arbitrary point in the explanatory variable space and the projected point in the subspace (2 norm in the explanatory variable space) is called the Q statistic. This Q statistic is the distance between an arbitrary point in the explanatory variable space and the partial space, and corresponds to the deviation from the correlation that the explanatory variables of the data for model creation had.
Also, a point projected from any point in the explanatory variable space into the subspace and the center of this subspace (corresponding to the average in the explanatory variable space of the model creation data, the center of the coordinates in the subspace :)) is referred to as the T ² statistic. The normalized distance is a distance weighted for each coordinate (“component”), and this weight is determined as the reciprocal of the variance of the “component” calculated from the original data for each “component”. Is. This indicates the degree of deviation from the average value including a plurality of explanatory variables.
FIG. 23 is a conceptual diagram for explaining the Q statistic and the T ² statistic.

The Q statistic / T ² statistic upper limit setting unit 63 in FIG. 22 calculates, for example, the maximum value excluding outliers of the Q statistic and the T ² statistic for each sample of the operating state variables of the model creation data. The upper limit value is set as the upper limit value, or the upper limit value is set based on a distribution of values excluding outliers of the Q statistic and the T ² statistic.
By setting these upper limit values as constraint conditions and making the Q statistic less than the upper limit value, it is possible to obtain data that maintains the correlation that the original explanatory variables had. Especially, the closer the Q statistic is to zero, the more variable Since the correlation between the variables is high, if the upper limit value is set small, it can be allowed only when the correlation between the variables is higher. In addition, by setting the T ² statistic to be equal to or lower than the upper limit value, the original explanatory variable can be prevented from greatly deviating from the average value, and the range existing in the multivariable space is not deviated.

The operating condition optimizing means 45c of FIG. 22, the prediction model 60c, the constraints defined by the upper limit value or the like of the Q statistic · T ² statistic, and the operating indicator operating condition index or plant states calculated from them, A quantity-related constraint condition and an objective variable (evaluation function) are input, optimization is performed by an optimization method such as mathematical programming, and an optimum operating condition value is output.
Here, since the prediction model is generally non-linear, it is expressed as y ^ = f (x1, x2,..., XN) as a function of explanatory variables (operating state variables).
The operating condition optimizing unit 45c uses the prediction model to formulate as an optimization problem in Expression 55.
In this optimization problem, the purpose is maximization. Therefore, in Equation 55, the value is maximize. However, in the optimization problem that is aimed at minimization, the value is minimize.

Here, y ^ = f (x1, x2, ......, xN) is x1, x2, ......, prediction value by the prediction model based on xN, Q, ^{T 2} is Q statistic and ^{T 2} statistics obtained from decision variables Strictly speaking, it may be expressed as Q (x1, x2,..., XN), T ² (x1, x2,..., XN). Q and T ² with “￣” are the upper limit value of the Q statistic and the upper limit value of the T ² statistic, which are the constraint conditions, and x _i ^L and x _i ^U are the lower limit value and upper limit of each decision variable. Value.
In addition, as an optimization method for solving the above optimization problem, Yanagiura and Ibaraki's "Combinatorial optimization-centering on the meta-line abbreviation-" (New Frontier of Management Science 2, Asakura Shoten, 2001), J Kennedy and R. Eberhart's “A discrete binary version of the particle swarm optimization algorithm” (Proc. Of the IEEE conference on Systems, Man, and Cybernetics (SMC'97), pp.4104-4109, 1997) The described quasi-Newton method, sequential quadratic programming, PSO (Particle Swarm Optimization), or the like can be used.

Next, in this embodiment, a case where principal component analysis processing is performed as multivariate analysis will be described.
For example, square is applied to the original operating state variable as non-linear processing, and the following principal component analysis processing is performed using the original operating state variable and the obtained square term as a whole as an explanatory variable x.
From the measurement data stored in the measurement DB 50 of the storage device 31, the sample × variable two-dimensional data X for the driving condition and the similar two-dimensional data Y for the driving index are extracted and input to the computer 30.
These data are composed of only net data excluding time and variable names as shown in FIG. 8, and each row corresponds to a sample and each column corresponds to a variable. For example, two-dimensional data X corresponds to Formula 56.

ここで、変数の数（列数）はＮ、サンプルの数（行数）はＭである。

Here, the number of variables (number of columns) is N, and the number of samples (number of rows) is M.

In the following, the NIPALS algorithm described in Non-Patent Document 2 is used as in the first embodiment.
In the principal component analysis process, a principal component analysis model coefficient P is calculated from the input two-dimensional data X and stored in the storage device 31 (model database). P is a matrix of N rows and A columns, where A is the number of principal components.
As A, it is set as the value used as the accumulation factor contribution rate 80% described in the nonpatent literature 2, for example.

３）前回のスコアｔ^{（ｋ−１）}と現在のスコアｔ^（ｋ）とを比較し、||ｔ^（ｋ）−ｔ^{（ｋ−１）}||＜１０^−１２なら収束で６）へ移行する。
４）次に、数式５８を計算する。

５）更に、数式５９を計算して２）へ移行する。

６）次の主成分の計算が必要ならＸ＝Ｘ−ｔｐ^Ｔとして１）へ移行し、不要なら終了する。
７）このようにして１）〜６）の処理をＡ個の縦ベクトルｐが得られるまで行い、得られたｐを順に左から並べてＰを得る。 The principal component analysis model coefficient P is obtained as follows.
1) Let the transpose of the row (horizontal vector) with the maximum variance in the matrix X be the initial value of p.
2) Next, Formula 57 is calculated.

3) comparing the previous score ^{t (k-1)} and current score ^{t (k),} proceeds to ^{|| t (k) -t (k} -1) || <6 ^{10 -12} if convergence) To do.
4) Next, Formula 58 is calculated.

5) Further, formula 59 is calculated, and the process proceeds to 2).

6) If calculation of the next principal component is necessary, shift to 1) as X = X-tp ^T , and end if unnecessary.
7) The processes 1) to 6) are performed until A vertical vectors p are obtained in this way, and the obtained ps are arranged in order from the left to obtain P.

Calculation of Q: Q is obtained by the least square method from the matrixes T and Y in which t obtained as described above are arranged.
Q = (T ^T T) ⁻¹ T ^T Y

Next, a Q statistic / T ² statistic calculation method when principal component analysis is used as the component conversion means will be described.
In this case, the principal component analysis process is performed in the same manner as described above using the original operating state variable as the explanatory variable x, and the Q statistic and the T ² statistic are expressed by Equation 60 and Equation 61 using the obtained coefficient matrix P. Respectively.

数式６０において、ｘ＾は、与えられた説明変数のベクトルｘの部分空間上の近似値であって、元の状態変数空間内の元の点から部分空間へ射影した点であり、ｘ＾＝Ｐｔ＝ＰＰ^Ｔｘで表される。

数式６１において、ｔ_ａは与えられた説明変数のベクトルｘから成分変換手段６２により得られた成分ベクトルｔ＝Ｐ^Ｔｘのａ番目の座標値、σ_ｔａはモデル作成用データについて、成分変換手段６２により変換された成分についてのａ番目の座標値に関する標準偏差である。また、Ａは「成分」の数（部分空間の次元）である。

In Equation 60, x ^ is an approximate value in the subspace of the vector x of the given explanatory variable, and is a point projected from the original point in the original state variable space to the subspace, and x ^ = Pt = PP ^T x.

In Equation 61, a second coordinate value of the resulting component vector t = P T ^x by component conversion means 62 from the vector x t _a is given explanatory variables, sigma _ta for modeling data, component conversion means 62 is the standard deviation for the a-th coordinate value for the component transformed by 62. A is the number of “components” (subspace dimension).

Based on these, the optimization problem is formulated as Equation 55 described above. For reference, Equation 55 is shown again.

Note that the above is a formulation for the variable after being standardized so as to have an average value of 0 and a variance of 1 with respect to the original variable. The optimal solution obtained in Formulation after normalization _{^{x 1 opt, x 2 opt,}} ......, and _{^{x N opt, x 1 opt,}} x 2 opt, ......, an ^{x N opt} with the ^{x opt} as a vector If the estimated values of the driving state variable (optimum driving condition after standardization) and the driving index variable (optimal driving index after standardization) obtained at this time are expressed as x ^opt and y ^opt respectively, the original values before standardization are expressed. In order to express the value in the unit system, the following formulas 62 and 63 are obtained. When mounted in an actual plant, these X ^opt and Y ^opt are used.

x ^opt : Optimal condition with standardized variable (unit system) X ^opt : Optimal condition with variable (unit system) before standardization y ^opt : Optimal operation index with standardized variable (unit system) Y ^opt : optimal operation indicators in before being standardized variable (unit system) X _m: X mean X _s: X standard deviation Y _m: Y average Y _s: Y standard deviation

Here, X and Y are vectors because they are composed of a plurality of variables, and their average values and standard deviation values are also vectors. Ys. A symbol “.” between the products of vectors such as y ^opt represents a vector obtained as a product of the elements.

Next, in this embodiment, a case where a partial least square method is used as multivariate analysis will be described. Note that the algorithm described in Non-Patent Document 2 is used as an algorithm of the partial least square method.
For example, the following partial least squares model creation process is applied to the original driving state variable as a non-linear process, and the whole of the original driving state variable and the obtained square term is used as the explanatory variable x. Do.
From the measurement data stored in the measurement database 50 of the storage device 31, for the data for creating a model of a normal part, the explanatory variable and the sample of the objective variable shown in Formula 64 × two-dimensional data X of the variable × y × y of the sample × 1 variable Are extracted and input.

In the partial least squares process, partial least squares coefficients P ₂ , W, and Q ₂ are calculated from the input two-dimensional data based on Expression 65, and stored in the storage device 31 (model database).

Ａが２である場合の最終的な予測モデル式は、次式の通りである。

Next, N types of data are converted into A types of components. When A is 2, the final prediction model equation is expressed by Equation 67 described later.
Also in the case of partial least squares, the latent variable matrix T ₂ is arranged in the descending order of variance from the left column as in the case of the principal component analysis, and the sum of variances in order from the left column is the total variance. Since the ratio (called cumulative factor contribution ratio) to the sum (the sum of the variances of all columns in X and T ₂ coincides) reaches 80% is considered appropriate as the number of latent variables. A value at which the cumulative factor contribution rate is 80% is adopted as the value of A.
Specifically, when the first row a latent variables matrix _{T 2} a _{t a,} the sigma _ta was the standard deviation of _{t a,} it is shown in Equation 66.

The final prediction model formula when A is 2 is as follows.

Here, the coefficient related to the first component is calculated by Formula 68.

Since the second component is a component that is not related to the first component, Formula 68 is converted by Formula 69.

Therefore, the coefficient related to the second component is expressed by Equation 70.

In addition, when there are more components than 3, it can obtain | require similarly.
When the above is generalized, Equation 71 is obtained.

The partial least squares coefficient P ₂ is N rows and A columns, Q ₂ is a vector of A rows and 1 column, and W is a matrix of N rows and A columns.

Next, a method for calculating the Q statistic and the T ² statistic when the partial least square method is used as the component conversion means will be described.
In this case, the principal component analysis process is performed in the same manner as described above using the original operating state variable as the explanatory variable x, and using the obtained coefficient matrix P, the Q statistic is calculated by Equation 72, and the T ² statistic is calculated by Equation 73. Calculate the quantity.
In addition, as shown in claim 30, based on a principal component analysis in which the whole including the original operating state variable and the variable subjected to nonlinear processing (square) as described above is x, the Q statistic · T ² statistic to be calculated, respectively.

In Equation 72, x ^ is an approximate value (estimated value) of a model of a given vector x of explanatory variables, and is a point projected from an original point in the variable space to the model (latent variable space); ^{x ^ = P (W T P} ) represented by -1 ^{W T} x.

数式７３において、Ｔ_ａは与えられた説明変数のベクトルｘから成分変換手段６２により得られた成分ベクトルｔ＝（Ｗ^ＴＰ）^−１Ｗ^Ｔｘのａ番目の座標値に関する標準偏差であり、Ａは「成分」の数（部分空間の次元）である。

In Equation 73, T _a is a standard deviation with respect to the a-th coordinate value of the component vector t = (W ^T P) ⁻¹ W ^T x obtained from the given explanatory variable vector x by the component converting means 62. A is the number of “components” (subspace dimension).

Using the coefficients P and Q of the partial least squares model obtained in this way, for example, if the objective function is maximization of the driving index estimated value, it is formulated as an optimization problem as in the following Expression 74. It becomes.

  As in the case of the principal component regression described above, what has been described here is optimization for variables after standardization to a mean value of zero and standard deviation of 1, and this is returned to the original unit system variable values. In this case, the same processing as described above is performed.

Next, a ninth embodiment of the present invention will be described. This embodiment corresponds to the embodiment of claim 31.
In the film formation process of a solar cell in the production of a film-like solar cell, the quality of the product (conversion efficiency of the solar cell) varies depending on the film formation conditions. As a quality prediction model, a nonlinear model representing the conversion efficiency is created from the film forming conditions, and the film forming conditions under which the film forming conditions are optimum are obtained. Here, the film is composed of about 10 layers, and each layer has a plurality of film forming conditions such as a film thickness, a doping amount, etc., so the film forming conditions (factors) are several dozens as a whole, and they are correlated with each other. Make a move.
In this embodiment, 20 factors out of several tens of factors are determined so that the conversion efficiency of the solar cell, which is the product quality, is optimal. An example of these 20 factors is shown in FIG.

Here, when these 20 factors are input and modeled by the normal (linear) partial least square method with conversion efficiency as output, the relationship between the factors and product quality can be sufficiently modeled. There wasn't.
Therefore, when modeled by partial least squares as explanatory variables including the quadratic terms of each factor, the estimation accuracy was improved as shown in FIG. 25, and sufficient accuracy was obtained as a model for estimating product quality. It was judged.

Therefore, based on the prediction model obtained in this way, correlation between explanatory variables is maintained, the degree of deviation from the average value including multiple explanatory variables is within a certain range, and each explanatory variable is within the upper and lower limits. The film forming conditions were determined so that the product quality would be optimal within the range of the constraints.
In this case, the Q statistic and the T ² statistic are as shown in Expression 75 and Expression 76, respectively. In Equation 75, in the nonlinear model, each quadratic term is added to the original 20 explanatory variables, so the total number of explanatory variables is 40. Further, in Expression 76, the subspace is set to 13 dimensions from the dimension 40 of the original explanatory variable.

When these are used to formulate the optimization problem, Equation 77 is obtained.

By solving this optimization problem by the quasi-Newton method, optimum conditions as shown in FIG. 26 were obtained.
This condition can be implemented practically, and the estimated efficiency value 0.919 based on this condition is sufficiently high, and it can be expected that improvement in conversion efficiency, which is product quality, can be expected from experience in manufacturing results so far. I confirmed that there was.

  The present invention is applicable to various plants including a power plant and a manufacturing plant.

1, 1 ', 2, 2', 3, 3 ': Plant (customer plant)
10, 10 ′: Plant (customer plant) operating condition optimization system 11, 11 ′: MES
12, 12 ': DCS
13, 13 ': Sensor 14, 14', 31: Recording device 20: Network 21, 22, 23, 24, 25: Router 30: Computer 33: Output unit 34: Input unit 35: Expert (monitor)
36: Optimal driving condition value implementation unit 37: Profit etc. calculation unit 38: Consideration collection unit 40: Function block 41: Driving state data acquisition unit 42: Driving index data acquisition unit 43: Measurement data recording unit 44: Regression model creation unit 44a : Principal component analysis (part)
44b: Partial least square method processing 45: Operation index variable optimization unit 45a, 45b: Operation condition optimization (part)
45c: Operating condition optimization means 46: Component conversion unit 47: Prediction unit 48: Inverse conversion unit 50, 50a, 50b, 50c: Measurement DB
50b-1: Measurement time column 50b-2: Pressure 1 column 50b-3: Temperature 1 column 50b-5, 50c-5: Efficiency column 50c-1: Lot number column 50c-2: Measurement time column 1 50c-3: Measurement time 2 column 51: Measurement timing column 52: Sensor 1 column 53: Sensor 2 column 54: Sensor N column 55: Driving index 1 column 56: Driving index K column 60, 60a: Regression model 60b: PLS model 60c: Prediction model 61: prediction model generating means 62: component converting means 62a: transformation coefficient matrix 63: Q statistic · ^{T 2} statistic upper limit value setting means 65 and 65a, 65b: the optimum operating condition unit 70, 70 ': display example 80, 90 : Time series graph 100: Optimization result 110: RAM
120: Internal block 121: CPU
122: ROM
123: RAM
124: VRAM
125: Image control unit 126: Controller 127: Recording medium 128: Input control unit 129: External I / F unit 130: Bus

Claims (32)

A plant operating condition optimization system,
Operating state data acquisition means for acquiring operating state data indicating the operating state of the plant measured by a plurality of sensors;
An operation index data acquisition means for acquiring operation index data for evaluating the operation of the plant, measured by a sensor provided in the plant or obtained based on the operation state data acquired by the operation state data acquisition means; ,
The driving state data acquired by the driving state data acquiring unit and the driving index data obtained by the driving index data acquiring unit are set as a set of measurement data related based on predetermined items, and the measurement data is recorded as data. Measurement data recording means for recording in the section;
Based on a plurality of sets of measurement data recorded in the data recording unit, a predetermined multivariate analysis is performed with the driving state variable representing the driving state data side as an explanatory variable and the driving index variable representing the driving indicator data side as an objective variable. Regression model creation means for creating a regression model, wherein the regression model is converted by the component conversion means and component conversion means for converting the explanatory variables into components that are uncorrelated with each other and fewer than the original explanatory variables. Predicting means for predicting the objective variable from the component obtained, and inverse transform means for estimating the explanatory variable from the component by a method corresponding to the component converting means,
An operation index variable optimizing means for obtaining an operating state variable for optimizing the operation index variable based on the regression model created by the regression model creating means, and satisfying a constraint condition on the explanatory variable estimated by the inverse transform means An operating condition optimization system for a plant characterized in that an explanatory variable value for optimizing an evaluation function related to an objective variable predicted by the prediction means is obtained, and the explanatory variable value is set as an optimal operating condition. In the plant operation condition optimization system according to claim 1, when there are a plurality of operation index data obtained by the operation index data acquisition means,
The measurement data recording means is a set of measurements in which the driving state data acquired by the driving state data acquisition means and a plurality of driving index data obtained by the driving index data acquisition means are related based on a predetermined item. As data, record the measurement data in the data recording unit,
The regression model creation means is a driving state representing the driving state data side based on a plurality of sets of measurement data recorded in the data recording unit and including a plurality of driving index data in the set of measurement data. A variable model is used as an explanatory variable, and a plurality of driving index variables representing the driving index data side are used as a plurality of objective variables to perform each predetermined multivariate analysis to create each regression model. The regression model is uncorrelated with the explanatory variables. And a component converting means for converting the number of components to a smaller number than the original explanatory variable, a predicting means for predicting a plurality of objective variables from the components converted by the component converting means, and a component in a method corresponding to the component converting means And an inverse transformation means for estimating the explanatory variable from
The driving index variable optimizing means obtains an operating state variable for optimizing the driving index variable based on the regression model created by the regression model creating means, and relates to the explanatory variable estimated by the inverse conversion means An explanatory variable value for optimizing one evaluation function based on each evaluation function related to a plurality of objective variables predicted by the prediction means while satisfying the constraint condition is obtained, and the explanatory variable value is set as an optimal operating condition. A system for optimizing the operating conditions of a plant.   3. The plant operating condition optimization system according to claim 2, wherein the one evaluation function is obtained by a predetermined weighted sum of each evaluation function.   The plant operating condition optimization system according to any one of claims 1 to 3, wherein the predetermined multivariate analysis is principal component regression.   4. The plant operating condition optimization system according to claim 1, wherein the predetermined multivariate analysis is a partial least square method (Partial Least Squares: PLS). 5. Optimization system.   6. The plant operating condition optimization system according to claim 1, wherein when the evaluation function is a linear expression related to an objective function and the constraint condition is a linear expression related to an explanatory variable, the optimization is performed. Is a plant operating condition optimization system characterized by using linear programming.   The plant operating condition optimization system according to any one of claims 1 to 5, wherein when the evaluation function is a nonlinear expression related to an objective variable and the constraint condition is a nonlinear expression related to an explanatory variable, the optimization Is a plant operating condition optimization system characterized by using nonlinear programming.   In the plant operating condition optimization system according to any one of claims 1 to 7, an optimum operating condition obtained by the operating index variable optimizing means, and an operating condition data obtained by the operating condition data obtaining means, The plant operating condition optimizing system further comprising validity evaluation means for outputting the validity of the optimum operating condition based on the evaluation. The plant operating condition optimization system according to any one of claims 1 to 8, wherein the plant is a customer's plant,
Optimal operating condition value mounting means for mounting the optimal operating conditions determined by the operating index variable optimizing means on the customer's plant side;
Based on the operation index data for evaluating the operation of the customer's plant in a predetermined period before and after the mounting by the optimal operating condition value mounting means, the profit etc. calculating means for calculating the customer profit increase and / or the cost saving effect,
A plant operating condition optimizing system, further comprising compensation collection means for collecting compensation based on an increase in customer profit and / or a cost saving effect calculated by the profit calculation means.   The plant operating condition optimization system according to any one of claims 1 to 9, wherein the plant is a power plant, and the operating state data includes any one or more of a flow rate, a pressure, and a temperature in the power plant, The operating condition optimization system for a plant, wherein the operating index data includes efficiency. A plant operating condition optimization method that causes a computer to execute optimization of a plant operating condition,
An operation state data acquisition step for acquiring operation state data indicating the operation state of the plant measured by a plurality of sensors;
An operation index data acquisition step for acquiring operation index data for evaluating the operation of the plant, measured by a sensor provided in the plant or obtained based on the operation state data acquired in the operation state data acquisition step; ,
The driving condition data acquired in the driving condition data acquisition step and the driving index data obtained in the driving index data acquisition step are set as a set of measurement data related based on predetermined items, and the measurement data is recorded as data. A measurement data recording step for recording in a section;
Based on a plurality of sets of measurement data recorded in the data recording unit, a predetermined multivariate analysis is performed with the driving state variable representing the driving state data side as an explanatory variable and the driving index variable representing the driving indicator data side as an objective variable. A regression model creation step for creating a regression model, wherein the regression model is converted by the component conversion unit that converts the explanatory variables into components that are uncorrelated with each other and smaller in number than the original explanatory variables. A prediction unit that predicts an objective variable from the component that has been performed, and an inverse conversion unit that estimates an explanatory variable from the component by a method corresponding to the conversion of the component conversion unit,
An operation index variable optimization step for obtaining an operation state variable for optimizing the operation index variable based on the regression model created in the regression model creation step, wherein the constraint condition regarding the explanatory variable estimated by the inverse transformation means is satisfied An operating variable optimization method for a plant characterized in that an explanatory variable value for optimizing an evaluation function related to an objective variable predicted by the predicting means is obtained, and the explanatory variable value is set as an optimal operating condition. The operation condition optimization method for a plant according to claim 11, wherein there are a plurality of operation index data obtained by the operation index data acquisition step.
The measurement data recording step includes a set of measurements in which the driving state data acquired in the driving state data acquisition step and a plurality of driving index data obtained in the driving indicator data acquisition step are related based on predetermined items. As data, record the measurement data in the data recording unit,
The regression model creation step includes a plurality of sets of measurement data recorded in the data recording unit, and a set of measurement data that can include a plurality of driving index data. A variable model is used as an explanatory variable, and a plurality of driving index variables representing the driving index data side are used as a plurality of objective variables, and each regression model is created by performing a predetermined multivariate analysis. The regression model is uncorrelated with the explanatory variables. A component conversion unit that converts the number of components into a smaller number of components than the original explanatory variable, a prediction unit that predicts a plurality of objective variables from the components converted by the component conversion unit, and a method corresponding to the conversion of the component conversion unit And having an inverse transform unit for estimating the explanatory variable from the component,
The operation index variable optimization step is to obtain an operation state variable that optimizes the operation index variable based on each regression model created in the regression model creation step, and is an explanatory variable estimated by the inverse conversion unit An explanatory variable value when optimizing one evaluation function based on each evaluation function related to a plurality of objective variables predicted by the prediction unit while satisfying the constraint condition is obtained, and the explanatory variable value is determined as an optimum operating condition. A method for optimizing plant operating conditions.   13. The plant operating condition optimization method according to claim 12, wherein the one evaluation function is obtained by a predetermined weighted sum of each evaluation function.   14. The plant operating condition optimization method according to claim 11, wherein the predetermined multivariate analysis is principal component regression.   The plant operating condition optimization method according to any one of claims 11 to 13, wherein the predetermined multivariate analysis is a partial least square method (Partial Least Squares: PLS). Optimization method.   16. The plant operating condition optimization method according to claim 11, wherein when the evaluation function is a linear expression related to an objective variable and the constraint condition is a linear expression related to an explanatory variable, the optimization is performed. Is a method for optimizing plant operating conditions, characterized by using linear programming.   The plant operating condition optimization method according to any one of claims 11 to 15, wherein when the evaluation function is a nonlinear expression related to an objective variable and the constraint condition is a nonlinear expression related to an explanatory variable, the optimization Is a method for optimizing plant operating conditions, characterized by using nonlinear programming. A plant operating condition optimization program that seeks to optimize plant operating conditions.
An operation state data acquisition step for acquiring operation state data indicating the operation state of the plant measured by a plurality of sensors,
An operation index data acquisition step for acquiring operation index data for evaluating the operation of the plant, measured by a sensor provided in the plant or obtained based on the operation state data acquired in the operation state data acquisition step,
The driving condition data acquired in the driving condition data acquisition step and the driving index data obtained in the driving index data acquisition step are set as a set of measurement data related based on predetermined items, and the measurement data is recorded as data. Measurement data recording step to record in the section,
Based on a plurality of sets of measurement data recorded in the data recording unit, a predetermined multivariate analysis is performed with the driving state variable representing the driving state data side as an explanatory variable and the driving index variable representing the driving indicator data side as an objective variable. A regression model creation step for creating a regression model, wherein the regression model is converted by the component conversion unit that converts the explanatory variables into components that are uncorrelated with each other and smaller in number than the original explanatory variables. A prediction unit that predicts an objective variable from the component that has been performed, and an inverse conversion unit that estimates an explanatory variable from the component by a method corresponding to the conversion of the component conversion unit,
An operation index variable optimization step for obtaining an operation state variable for optimizing the operation index variable based on the regression model created in the regression model creation step, wherein the constraint condition on the explanatory variable estimated by the inverse transformation unit is satisfied An operating condition optimization program for a plant for obtaining an explanatory variable value when optimizing an evaluation function related to an objective variable predicted by the prediction unit and executing a step of setting the explanatory variable value as an optimal operating condition. The plant operation condition optimization program according to claim 18, wherein there are a plurality of operation index data obtained by the operation index data acquisition step.
In the measurement data recording step, a set of measurements in which the driving state data acquired in the driving state data acquisition step and a plurality of driving index data obtained in the driving indicator data acquisition step are related based on predetermined items. As data, record the measurement data in the data recording unit,
In the regression model creation step, the driving state representing the driving state data side based on a plurality of sets of measurement data recorded in the data recording unit and including a plurality of driving index data in the set of measurement data A variable model is used as an explanatory variable, and a plurality of driving index variables representing the driving index data side are used as a plurality of objective variables, and each regression model is created by performing a predetermined multivariate analysis. The regression model is uncorrelated with the explanatory variables. And a component conversion unit that converts the number of components to a smaller number of components than the original explanatory variable, a prediction unit that predicts a plurality of objective variables from the components converted by the component conversion unit, and a program corresponding to the conversion of the component conversion unit And having an inverse transform unit for estimating the explanatory variable from the component,
In the operation index variable optimization step, an operation state variable for optimizing the operation index variable is obtained based on each regression model created in the regression model creation step, and the explanatory variable estimated by the inverse conversion unit An explanatory variable value when optimizing one evaluation function based on each evaluation function related to a plurality of objective variables predicted by the prediction unit while satisfying the constraint condition is obtained, and the explanatory variable value is determined as an optimum operating condition. A program for optimizing a plant operating condition.   20. The plant operating condition optimization program according to claim 19, wherein the one evaluation function is obtained by a predetermined weighted sum of each evaluation function. In the plant operating condition optimization program according to any one of claims 18 to 20,
The plant operating condition optimization program, wherein the predetermined multivariate analysis is principal component regression. In the plant operating condition optimization program according to any one of claims 18 to 20,
The operation condition optimization program for a plant, wherein the predetermined multivariate analysis is a partial least squares (PLS). In the plant operating condition optimization program according to any one of claims 18 to 22,
When the evaluation function is a linear expression related to an objective variable and the constraint condition is a linear expression related to an explanatory variable, the optimization uses a linear programming method. In the plant operating condition optimization program according to any one of claims 18 to 22,
When the evaluation function is a nonlinear expression related to an objective variable and the constraint condition is a nonlinear expression related to an explanatory variable, the optimization uses a nonlinear programming method. A plant operating condition optimization system,
Operating state data acquisition means for acquiring operating state data indicating the operating state of the plant measured by a plurality of sensors;
An operation index data acquisition means for acquiring operation index data for evaluating the operation of the plant, measured by a sensor provided in the plant or obtained based on the operation state data acquired by the operation state data acquisition means; ,
The driving state data acquired by the driving state data acquiring unit and the driving index data obtained by the driving index data acquiring unit are set as a set of measurement data related based on predetermined items, and the measurement data is recorded as data. Measurement data recording means for recording in the section;
Prediction that creates a prediction model based on a plurality of sets of measurement data recorded in the data recording unit, with driving state variables representing the driving state data side as explanatory variables and driving index variables representing the driving index data side as objective variables Model creation means;
For the measurement data, explanatory variables having correlation with each other are set by projecting the original explanatory variables to the subspace, which are uncorrelated with each other and have a smaller number of components than the original explanatory variables. Component conversion means for converting into a component as a coordinate,
The square of the distance between an arbitrary point in the original explanatory variable space and the projection point on the subspace is calculated as a Q statistic, and the normalized distance between the projection point and the center of the subspace is calculated. Q statistic and T ² statistic calculation means for calculating the square as T ² statistic,
And Q statistic · T ² statistic upper limit value setting means for the upper limit value respectively set as a constraint in the optimization of the Q statistic and T ² statistics,
A means for obtaining an explanatory variable for optimizing an objective variable based on the prediction model, wherein the Q statistic and the T ² statistic are each equal to or less than the upper limit value as at least one of the constraint conditions. An operating condition optimizing means for obtaining an explanatory variable value at the time of optimizing the evaluation function related to the objective variable while maintaining the correlation between the variables, and using the explanatory variable value as an optimal operating condition;
A plant operating condition optimization system characterized by comprising:   26. The plant operating condition optimization system according to claim 25, wherein a principal component analysis is used as the component conversion means.   26. The plant operating condition optimization system according to claim 25, wherein a partial least square method is used as the component conversion means. In the plant operating condition optimization system according to claim 25,
As the prediction model, all the variables that combine the original operating state variable and the new variable obtained by subjecting the original operating state variable to nonlinear transformation, or a variable obtained by extracting an appropriate one of these variables as explanatory variables, A plant operating condition optimization system using a multiple regression model with an indicator variable as an objective variable. The operation condition optimization method for a plant according to claim 28,
Wherein Q statistic · T ² statistic calculation means, and variables, including some or all of the non-linear terms used in the creation of the original explanatory variables and the predicted model, the Q statistic and T ² statistics based on A plant operating condition optimization system characterized by calculating quantity. The plant operating condition optimization system according to claim 28 or 29,
As the prediction model, using a principal component regression model in which a variable obtained by adding a new variable obtained by performing nonlinear transformation on the original driving state variable is an explanatory variable, and a driving index variable is an objective variable,
The Q statistic / T ² statistic calculating means calculates the Q statistic and the T ² statistic based on a principal component analysis calculated in the principal component regression model. Condition optimization system. The plant operating condition optimization system according to claim 28 or 29,
As the prediction model, a variable obtained by adding a new variable obtained by performing nonlinear transformation on the original driving state variable is used as an explanatory variable, and a partial least square method model using a driving index variable as an objective variable is used.
The Q statistic / T ² statistic calculating means calculates the Q statistic and the T ² statistic based on a partial least squares model as the prediction model. system.   26. The plant operating condition optimization system according to claim 25, wherein a neural network is used as the prediction model.

Images