JP2007004769A - Parameter prediction apparatus and parameter prediction method for oil refinery plant - Google Patents

Abstract

A prediction model optimized for a prediction time and a first parameter and a second parameter used for model creation is constructed, and a specific parameter of an oil refinery plant is accurately predicted using the prediction model.
A PLS (Partial Least Squares) analysis is performed on at least a part of first and second parameters measured based on a model creation condition, an input unit, a measurement unit, a parameter storage unit, and a predicted time and A prediction apparatus comprising: a model creation unit that creates a prediction model optimized with respect to the first and second parameters used for model creation, a model storage unit, and an output unit.
[Selection] Figure 1

Description

  The present invention relates to a prediction apparatus that creates a prediction model capable of predicting a specific parameter ahead of a predetermined time from a predetermined parameter that characterizes an oil refinery plant, and predicts the specific parameter with high accuracy using the prediction model.

  In an oil refining plant, once an abnormal state occurs, it takes a lot of time and cost to resolve it. Therefore, there is a demand for a technology that can detect an abnormality at a predetermined time ahead of a plant of interest at an early stage.

  There are two methods for predicting parameters of an oil refinery plant: a method based on a physical model of the plant and a method based on a mathematical model represented by a neural network model or a regression model.

  The physical model can actually reflect the phenomena occurring inside the plant, but many complicated phenomena occur inside the oil refinery plant, and it is difficult to express them all as physical models. It was. In addition, the calculation amount is enormous and there is a problem in terms of prediction accuracy.

  On the other hand, a mathematical model creates a model based on data that can be collected and used from a plant. Therefore, unlike a physical model, model construction and tuning are unnecessary, and it can be said to be a practically effective prediction method.

  Patent Document 1 describes a plant abnormality diagnosis apparatus using a neural network. In this method, in order to ensure the accuracy of estimation results used for abnormality diagnosis, a method of collecting data from a plurality of operation regions and learning with a neural network for a plurality of plant data set in advance has been proposed.

  In the estimation / prediction method using the neural network model described in Patent Document 1, it is necessary to model using plant data having a degree of relevance as much as possible to the prediction target in order to ensure and improve the prediction accuracy. However, in this prediction method, which type of plant data is used in which range is arbitrary, and is set mainly based on the knowledge and experience of the designer. For this reason, depending on the selected plant data, the degree of relevance with respect to the prediction target is shown in one operation region, but the degree of relevance is weak in another operation region, which may cause a significant decrease in prediction accuracy. It was.

  In the oil refinery plant, the parameters are complexly correlated with each other. For this reason, there is a linear relationship depending on the parameter to be selected (multicollinearity), and when the regression model is used, there is a case where the variance becomes a large value and the prediction accuracy is greatly lowered.

Therefore, PLS (Partial Least Squares) analysis can be cited as a mathematical model that can be accurately estimated even when multi-collinearity is recognized between parameters regardless of which parameter is selected for model creation.
Japanese Patent Laid-Open No. 9-113671

  In order to perform the PLS analysis, first, an index (input data) used for prediction calculation in the prediction model and an index (output data) predicted for a predetermined time from the time of measurement of the input data are measured. Then, the relationship between the input data and the output data is created as a prediction model. In the conventional PLS analysis, it is determined which data is used as the data used for the prediction calculation when creating this model, and how much ahead data is predicted as the predicted data (prediction time). There was a need.

  However, there are cases where specific abnormal values are included in these data, and when a model is created using this data, the prediction accuracy may be significantly reduced.

  Therefore, as a result of intensive studies, the present inventors have been able to predict parameters with high accuracy by optimizing parameters used for prediction time and model creation in PLS analysis.

  In addition, depending on the data to be predicted, the result of the PLS analysis may be greatly affected by disturbances and time series history of past input / output data. Accordingly, the present inventors have found that RNN analysis is excellent as an analysis method that has a characteristic that is strongly influenced by a non-linear element and is not easily affected by disturbance, and that can effectively reflect the influence of time series history of past data. did. Accordingly, it has been discovered that by combining the PLS analysis and the RNN analysis to predict the output data, it is possible to effectively combine the characteristics of both to predict the parameters with higher accuracy.

In order to solve the above problems, the prediction apparatus of the present invention has the following configuration. That is, the present invention uses the first and second parameters measured in the oil refinery plant to create a prediction model that predicts the second parameter ahead by the prediction time from the time of measurement of the first parameter, and the prediction model A prediction device for predicting the second parameter using
Input means for inputting first and second parameter measurement conditions and model creation conditions;
Measurement means for measuring the first and second parameters from the oil refinery plant based on the first and second parameter measurement conditions;
Parameter storage means for storing the first and second parameters;
Based on the model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction optimized with respect to the prediction time and the first and second parameters used for model creation A model creation means for creating a model;
Model storage means for storing the optimized prediction model;
Output means for outputting a second parameter predicted by the optimized prediction model.

The present invention uses the first and second parameters measured in the oil refinery plant to create a prediction model that predicts the second parameter ahead by the prediction time from the time of measurement of the first parameter, and uses the prediction model A prediction device for predicting the second parameter,
Input means for inputting first and second parameter measurement conditions, model creation conditions and model correction conditions;
Measurement means for measuring the first and second parameters from the oil refinery plant based on the first and second parameter measurement conditions;
Parameter storage means for storing the first and second parameters;
Based on the model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction optimized with respect to the prediction time and the first and second parameters used for model creation A model creation means for creating a model;
Based on the model correction condition, a model correction unit that further performs an RNN (Recurrent Natural Network) analysis on the prediction model created by the model creation unit to create a modified prediction model obtained by correcting the prediction model;
Model storage means for storing the modified prediction model;
And an output unit that outputs a second parameter predicted by the modified prediction model.

The present invention uses the first and second parameters measured in the oil refinery plant to create a prediction model that predicts the second parameter ahead by the prediction time from the time of measurement of the first parameter, and uses the prediction model Predicting the second parameter,
A process for measuring the first and second parameters from the oil refinery plant based on the input first and second parameter measurement conditions;
A process for storing the first and second parameters;
Based on the input model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction time and the first and second parameters used for model creation are optimized. A process of creating the prediction model;
Storing the optimized prediction model;
And a process of outputting a second parameter predicted by the optimized prediction model.

The present invention uses the first and second parameters measured in the oil refinery plant to create a prediction model that predicts the second parameter ahead by the prediction time from the time of measurement of the first parameter, and uses the prediction model Predicting the second parameter,
A process for measuring the first and second parameters from the oil refinery plant based on the input first and second parameter measurement conditions;
A process for storing the first and second parameters;
Based on the input model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction time and the first and second parameters used for model creation are optimized. A process of creating the prediction model;
A process of creating a modified prediction model in which the prediction model is further corrected by performing an RNN (Recursive Neural Network) analysis on the prediction model based on the input model modification condition;
Storing the modified prediction model;
And a process of outputting a second parameter predicted by the modified prediction model.

  The first prediction apparatus of the present invention has input means, measurement means, parameter storage means, model creation means, model storage means, and output means, and the second prediction apparatus of the present invention is in addition to these means. Furthermore, it has a model correction means.

  Parameter input conditions (first parameter measurement conditions, second parameter measurement conditions), model creation conditions, and model correction conditions, if necessary, are input to the input means.

  Here, the “first parameter measurement condition” is a condition for designating the measurement means used for measurement of the first parameter, the measurement interval, and the number of measurement points of each measurement means in the oil refinery plant.

  The “second parameter measurement condition” is a condition for designating a measurement means used for measuring the second parameter in the oil refinery plant. In this prediction apparatus, the measurement interval of the second parameter and the number of measurement points of each measurement means are set to be the same as the value input as the first parameter measurement condition. In addition, the number of parameters measured as “first parameters” is set to be larger than the number of parameters measured as “second parameters”.

  The “model creation condition” is the maximum time difference (maximum prediction time) between the two parameters when PLS analysis is performed using the measured first and second parameters to create a prediction model. The model is created, and the number of model constituent variables constituting the prediction model is specified.

The “model correction condition” defines conditions for creating a corrected prediction model in which the prediction model obtained by PLS analysis is further corrected by RNN analysis. Specifically, when the RNN analysis is performed to create the modified prediction model, which residue obtained by the PLS analysis is used and which first parameter among the first parameters used in the PLS analysis is used for the RNN analysis And a condition for designating the maximum number of intermediate layers.
Note that it is not necessary to input all these conditions, and some or all of the conditions may be set in advance.

  The measuring means includes a first measuring means capable of measuring the first parameter and a second measuring means capable of measuring the second parameter. The first measurement means measures the first parameter from the oil refinery plant at the measurement interval and the number of measurement points input as the first parameter measurement condition. The second measurement means measures the second parameter from the oil refinery plant at the same measurement interval and number of measurement points as the first measurement means.

  Here, the “first parameter” and the “second parameter” are parameters representing the characteristics of the oil refinery plant, and are not particularly limited as long as they represent the characteristics of the plant. For example, pressure at a predetermined position in the plant, temperature, flow rate of raw materials / products, product properties (flash point, initial fraction, distillation point) and the like can be mentioned.

  The parameter storage means stores the first and second parameters measured by the first measurement means and the second measurement means, respectively.

  The model creation means performs PLS analysis on at least a part of the measured first parameter and second parameter, and creates a prediction model optimized for the prediction time and the first and second parameters used for model creation.

  The model correction means creates a modified prediction model by further modifying the prediction model created by the model creation means by further performing an RNN analysis on the result of the PLS analysis.

The model storage means stores the prediction model created by the model creation means or the model correction means.
The output means outputs the second parameter ahead by the prediction time using the optimized prediction model.

  In the present invention, various means may be formed so as to realize the function. For example, dedicated hardware that exhibits a predetermined function, a prediction device provided with a predetermined function by a computer program, a computer It can be realized as a predetermined function realized by the prediction device by a program, a combination thereof, or the like.

  Further, in the present invention, the various means do not need to be independent of each other, a plurality of means are formed as one member, a certain means is a part of another means, It is also possible that some of the above and some other means overlap.

  The storage means of the present invention may be hardware in which a computer program for causing the prediction device to execute various processes is stored in advance, for example, a ROM (Read Only Memory) fixed to the prediction device, It is possible to implement with HDD (Hard Disc Drive), CD (Compact Discs) -ROM, FD (Flexible Disc-cartridge), a combination thereof, and the like that are exchangeably mounted on the prediction device.

  The prediction apparatus of the present invention may be hardware that can read a computer program and execute a corresponding processing operation. For example, a CPU (Central Processing Unit) is a main component, and ROM, RAM (Random Access Memory) is included in this. ), Hardware to which various devices such as an I / F (Interface) unit are connected.

  In the present invention, causing the prediction device to execute various operations corresponding to the computer program also means causing the prediction device to control operations of various devices. For example, storing various data in the prediction device means that the CPU stores various data in an information storage medium such as a RAM fixed to the prediction device, and information such as an FD that is exchangeably loaded in the prediction device. The CPU may store various data in the storage medium by FDD (FD Drive).

  The first prediction apparatus of the present invention constructs a prediction model optimized for the prediction time and the first and second parameters used for model creation by using PLS analysis, and uses this prediction model to refine oil with high accuracy. Specific parameters such as the plant can be predicted.

  The second prediction apparatus of the present invention combines the PLS analysis and the RNN analysis to reflect the influence of the past time series history with high accuracy by taking advantage of the characteristics of both analysis methods and being less susceptible to disturbances. Certain parameters, such as an oil refinery plant, can be predicted.

(Embodiment 1)
The first prediction apparatus of the present invention includes input means, measurement means (first measurement means, second measurement means), parameter storage means, model creation means, model storage means, and output means. Each constituent means is connected by a cable so that data can be exchanged.

  FIG. 2 shows a distillation column as an example of an oil refinery plant that performs parameter prediction using the prediction apparatus of the present invention. The distillation column 14 is an apparatus for separating crude oil into fractions such as light gasoline, heavy gasoline, kerosene, light oil, heavy light oil, in addition to the tower top gas and tower bottom oil. In the distillation column 14, raw materials are supplied from the predetermined positions 11 to 13 to the preliminary distillation columns 15 to 17, respectively. The preliminary distillation towers 15 to 17 are separated into tower top gas 23 to 25 and tower bottom oils 26 to 28 from the supplied raw materials and components discharged from the predetermined positions 20 to 22 of the distillation tower, respectively. The column top gases 23 to 25 are returned to the distillation column 14 again, and the column bottom oils 26 to 28 are taken out as products. The raw material is also supplied from a predetermined position 19 near the tower bottom. Further, a tower top gas 29 and a tower bottom oil 30 are taken out from the distillation tower 14. The intermediate raw material in the distillation column is transported from 31 to 33 to 34 to 36 by a compressor.

  In this distillation column 14, measuring means are arranged at 11 to 13 and 15 to 22, and arranged so that the characteristics of 23 to 30 tower top gas or tower bottom oil can be measured. Of these measuring means, the one that measures the first parameter is the first measuring means, and the one that measures the second parameter is the second measuring means.

  As parameters, raw material supply flow rate, product discharge flow rate, temperature, pressure, parameters calculated by combining these parameters, or parameters obtained by performing special arithmetic processing on these parameters, etc. may be used. it can. Note that these combination calculations and special calculation processing can be performed by calculation means built in the measurement means. These parameters can be arbitrarily assigned as the first parameter and the second parameter, but since the first parameter is used to predict the second parameter, the first parameter is ahead of the second parameter in the distillation column. It must be a parameter that represents the state. Also, the first parameter and the second parameter cannot overlap, and the number of first parameters must be greater than the number of second parameters.

  For example, in the distillation column of FIG. 2, the parameters obtained by measuring the characteristics of the bottom oil 30 and the top gas 29 are parameters that represent the state of the product finally obtained in the distillation process, and are therefore only used as the second parameter. Cannot be used. Further, the parameters measured at the predetermined positions 11 to 13 and 19 are parameters representing the state of the raw material supplied first, and therefore can only be used as the first parameter.

  In the PLS analysis of the present invention, some or all of the above parameters may be used. The parameters are not limited to those shown in FIG. 2 and are not particularly limited as long as they represent the characteristics of the distillation column. In the present invention, since PLS analysis is used for model creation, it is possible to predict the value of a specific parameter for a predetermined time with high accuracy regardless of the parameters used for model creation.

  Details of the processing steps by the prediction apparatus of the present invention will be described with reference to FIG. First, parameter measurement conditions (first parameter measurement condition, second parameter measurement condition) and model creation conditions are input to the input means 51. As the first parameter measurement conditions, the measurement means for measuring the first parameter in the oil refinery plant, the measurement interval, and the number of measurement points to be measured by each measurement means are input. As the second parameter measurement condition, measurement means for measuring the second parameter in the oil refinery plant is input (in the present embodiment, the case where there is one kind of second parameter) is input. As model creation conditions, the maximum time difference (maximum prediction time) between the first parameter and the second parameter is set for the PLS analysis, and the number of model configuration variables constituting the prediction model is input.

  Next, based on the first parameter measurement conditions input to the input unit, the predetermined first measurement unit 52 disposed in the oil refinery plant 50 measures the first parameter at a predetermined measurement interval and number of measurement points. . The second measurement means 53 is synchronized with the measurement of the first measurement means 52 by the predetermined second measurement means 53 arranged in the oil refinery plant 50 based on the second parameter measurement conditions inputted to the input means. Then, the second parameter is measured at the same measurement interval and number of measurement points as the first measurement means. The measured data of the first parameter and the second parameter are conceptually shown in FIG. 3 (FIG. 3 shows a case where the number of measurement points is 36).

  The first parameter and the second parameter measured in this way are stored after being corrected by the parameter storage means as necessary. As a method for correcting the data of the first parameter x, the method of equation (1) can be mentioned.

Here, x _np ^* is the n-th measured value of the p-type out of x measured at the M points for each of the Q types, x _p ^′ is the average value of the p-th type x measured for the M points, σ _xp represents the standard deviation of the p-th x.

  Next, PLS analysis is performed on the measured first parameter and second parameter by the model creation means, and a prediction model optimized with respect to the prediction time and the first and second parameters used for model creation is created. The process of creating this prediction model will be described with reference to FIG.

  First, a plurality of temporary prediction models are created by temporary model creation means. In this process, first, which part of the measured first parameter and second parameter is to be used as data for creating a prediction model (data region) is designated.

  In order to specify this data area, it is necessary to determine the prediction time. First, the prediction time is changed stepwise within the range of the maximum prediction time specified by the model creation conditions inputted in advance (S1, S2). ). However, the predicted time must be less than or equal to the time represented by (measurement interval) × {(number of measurement points) −1} input in advance as the first parameter measurement condition. The rate of change in stages must be at least a multiple of (measurement interval) and (measurement interval) × (number of measurement points) / 2 or less.

  For example, when the measurement interval is 1 minute and the number of measurement points is 100, (measurement interval) × {(number of measurement points) −1} = 1 × (100-1) = 99 minutes, and the predicted time is 1 to 99 minutes. Can vary up to. For example, when the prediction time is changed every minute (measurement interval), the generated prediction time is 1, 2, 3,. . . 99 minutes. When the prediction time is changed every 2 minutes, 1, 3, 5. . . 99 minutes.

  Sampling of the first and second parameters is performed according to a predetermined condition for each predicted time thus changed. For this purpose, first, a data area for sampling is designated. Examples of the designated data area are shown in FIGS. In FIG. 4, the data area is represented by a gray area. 4A to 4C show data areas when different prediction times 61 are generated by the above process. At this time, the last part 63 of the data area of the second parameter is set to coincide with the end of measurement, and the measurement interval and the number of measurement points are designated in advance as the parameter measurement conditions, so the prediction time 61 is determined. For example, the data area of the first parameter and the second parameter is automatically determined.

  The data area of the first parameter and the data area of the second parameter are overlapped as shown in FIG. 4 (a) or are contacted as shown in FIG. 4 (b), but are separated as shown in FIG. 4 (c). May be.

  Next, the first parameter used for creating the prediction model is sampled from the data area of the first parameter specified in this way under a predetermined condition (S3). Sampling conditions are not particularly limited. For example, sampling is performed one or more times at a sample number that falls within a predetermined range, a method of sampling in sequence at a predetermined measurement interval, or other sampling according to a predetermined law. Can be mentioned.

  It is preferable to sample at random. In random sampling, sampling is not performed according to a predetermined law, so that it is difficult to capture abnormal values. In addition, the number of samples can be set freely. The sampling condition may be input as a model creation condition, or a predetermined condition may be set in the prediction device in advance. FIG. 5A shows data 62 sampled at random.

  Next, the second parameter 64 ahead is sampled from the first parameter 62 sampled in this way by the estimated time determined by the above process (FIG. 5B: S3). Since the second parameter is specified from the first parameter by the predicted time 61, if the first parameter 62 is determined, the second parameter 64 to be automatically sampled is also determined.

  Next, a first parameter matrix X (formula (2)) is created from the first parameters sampled in this way, and a second parameter matrix Y (formula (3)) is created from the second parameters.

Here, X _np represents the first parameter sampled n-th in the p-th type, and Y _n represents the second parameter sampled n-th. Next, a model configuration variable matrix (equation (4)) having the same number as the sampled first parameters is created (S4).

Here, W ₁ represents the first model configuration variable. Thereafter, a matrix t represented by the product of the first parameter matrix X and the model constituent variable matrix (4) is created, and the inner product φ of the second parameter matrix Y and the matrix t is calculated.

It can be evaluated that the larger the value of the inner product φ, the smaller the error between the second parameter predicted using the model configuration variable W ₁ and the measured second parameter. Therefore, a model configuration variable W ₁ that maximizes φ is obtained. Specifically, the model configuration variable W ₁ is obtained by the undetermined multiplier method under the condition of norm 1 (W ₁ ² −1 = 0) (S5).

Next, if the portion of the second parameter matrix Y that cannot be represented by the model configuration variable W ₁ is Y ⁽¹⁾ , Y ⁽¹⁾ can be represented by W ₁ and its weight coefficient r ₁ ((10) Since ⁽ formula), Y ⁽¹⁾ and r ₁ XW ₁ are orthogonal to each other, r ₁ represented by formula (11) is determined.

When r ₁ represented by the equation (11) is used, the second parameter Y (= r ₁ XW ₁ ) predicted using the model configuration variable W ₁ and the first parameter X can be expressed.

Here, the number of model configuration variables previously entered as a model creation condition when 2 or more (S6, S7), determining the next model configuration variables W _2. The process for obtaining the model configuration variable W ₂ will be described below. First, X created using the first and second parameters by sampling, when the portion which can not be represented in the model configuration variable W ₁ of the Y and X ⁽¹⁾ and ^{Y (1), X (1} ) and Y ⁽¹⁾ is expressed as in equations (12) and (13) (S8).

Next, the model configuration variable W ₂ is obtained by substituting Y ⁽¹⁾ and X ⁽¹⁾ into the equations (2) to (11) as Y → Y ⁽¹⁾ and X → X ⁽¹⁾ . If the number of model configuration variables input as model creation conditions is 3 or more, the process is performed in the same manner. Then, the process ends when the number of model configuration variables input as the model creation condition is reached.

That is, when the number of model configuration variables input as model creation conditions is K, the following processing is performed to sequentially obtain model configuration variables from W ₁ to W _K.

The following X ⁽⁰⁾ and Y ⁽⁰⁾ can be set as appropriate according to the type of data to be used. Thus to compute each r _k XW _k from each model configuration variable W _k obtained, it is possible to create a temporary predictive model using the respective r _k XW _k. Through the above process, the provisional model creation means creates a plurality of provisional prediction models in which the first and second parameters are sampled according to a predetermined condition for each changed prediction time. Next, for each temporary prediction model, the first parameter used to create the temporary prediction model is input, and the second parameter ahead by the prediction time is predicted. Then, an optimum model having the smallest error between the predicted second parameter and the sampled second parameter is determined from the created temporary prediction models.

  A statistical index can be used to evaluate the error. As the statistical index, for example, the following indices can be used.

Here, e _i is (second parameter predicted by the provisional prediction model) − (sampled second parameter), and e ^′ is {(sampled second parameter) − (average value of sampled second parameter) )}, Y _i is the sampled second parameter, and y ^′ is the sampled second parameter average. These statistical indicators may be used alone or in combination for error evaluation.

  Thus, the temporary prediction model evaluated as the model with the least error by the statistical index is set as the optimum prediction model and stored in the model storage unit. Further, when an output command is input, the output means predicts a specific parameter ahead by the prediction time by using the optimum prediction model stored in the model storage means.

(Embodiment 2)
The second prediction apparatus of the present invention further includes an update command means, and is configured to generate a prediction model and update the model again when the model update condition previously input to the update condition input unit is satisfied. Yes. Thus, the second parameter can be accurately predicted over a long period of time by updating the prediction model according to the model update condition.

  This process will be described with reference to FIG. The update command means 59 has an update condition input unit 58 for inputting model update conditions. Then, an update command is issued to update the prediction model when the model update condition is satisfied. When this update command is issued, the first measurement unit and the second measurement unit automatically start measuring the first parameter and the second parameter, respectively, and the optimized prediction model is again determined by the process of the first embodiment. create.

  An example of the update condition is a case where the prediction model is automatically updated every predetermined time. In this case, the predetermined time can be calculated from when the prediction model is created. In addition, the second parameter is regularly predicted, and the second parameter is actually measured at the time when the second parameter is predicted (the time that is advanced by the predicted time from the measurement of the first parameter used for prediction). When an error between the predicted second parameter and the measured second parameter is calculated by a statistical index such as equations (16) to (18) and the error exceeds a predetermined value, It may be a case where an update command is issued after determination.

(Embodiment 3)
According to a third prediction apparatus of the present invention, the first and second prediction apparatuses further include model correction means. In this prediction device, an RNN analysis is further performed on the result of the PLS analysis. The RNN analysis has characteristics different from the PLS analysis, such as being less susceptible to disturbances and capable of reflecting the time series history of past input / output data in the output data effectively. For this reason, it is possible to construct a prediction model with further improved prediction accuracy by performing further RNN analysis on the prediction model subjected to PLS analysis.

  FIG. 8 shows an outline of a prediction apparatus in which model correction means is further added to the second prediction apparatus. The prediction apparatus includes an input unit, a measurement unit (first measurement unit, second measurement unit), a parameter storage unit, a model creation unit, a model correction unit, a model storage unit, an output unit, and an update command unit.

In the prediction apparatus of FIG. 8, a modified prediction obtained by further modifying the prediction model created by the model creation unit 55 from the model correction condition input to the input unit 51 and the measured value of the first parameter stored in the parameter storage unit. A model is created by the model correcting means 71. The corrected prediction model is stored in the model storage unit 56 and can be output by the output unit 57 as necessary.
A method for creating a modified prediction model that combines the PLS analysis and the RNN analysis will be described in detail below.

FIG. 10 is a flowchart showing the process of RNN analysis. First, the residue Y ^(k) (formula (14)) that cannot be expressed by the model configuration variable W _k among the sampled second parameters is calculated by PLS analysis as in the first or second embodiment. Next, it is determined which type of residue to be used for creating a modified prediction model by RNN analysis from among the plurality of types of residues thus obtained. This residue may be all the residues obtained by PLS analysis or a part of the residues. When some residues are used, for example, the determination can be made by sampling at random from all the residues obtained by the PLS analysis (S11). In this case, the number of samples can be freely set within a range that does not interfere with the RNN analysis.

  Next, a first parameter used for creating a prediction model by RNN analysis is determined. The first parameters may be all the first parameters used for model creation by the PLS analysis or some of the first parameters. When some first parameters are used, for example, sampling can be performed randomly (S11). In this case, the number of samples can be freely set within a range that does not interfere with the RNN analysis.

Next, an RNN analysis is performed using the residue Y ^(k) determined as described above and the first parameter X. FIG. 9 shows an outline of this RNN analysis. In the RNN analysis, there are three processing steps (virtual regions) of an input layer, an intermediate layer, and an output layer. The residue and the first parameter determined as described above are first input to the input layer. .

In the input layer, Y ^(k) · d ^obtained by multiplying the residue Y ^(k) by the forgetting rate d is calculated, and this Y ^(k) · d and the first parameter are input to the sigmoid function (Formula A). Is output (formulas B and C). Here, the forgetting rate d is set to a value of 0 <d ≦ 1 depending on conditions.

Next, the number of intermediate layers is sequentially changed one by one from 1 to the maximum number of intermediate layers previously input as model correction conditions (S12 to S17). Next, a weight function W corresponding to the number of intermediate layers is defined, and a load sum z is output. For example, when the number of intermediate layers is 1, the weight function is W _ij ⁽¹⁾ and its load sum z _j (formula D) is output as shown in FIG. When the number of intermediate layers is 2, the weight functions for the first and second intermediate layers are W _ij ⁽¹⁾ and W _jk ⁽²⁾ , the load sums are z _j and z _k , and the number of intermediate layers is 3, respectively. In this case, the weight functions for the first, second, and third intermediate layers are W _ij ⁽¹⁾ , W _jk ⁽²⁾ , W _kl ⁽³⁾ , and the load sums are z _j , z _k , z _l . . . Thus, the weight function W and the load sum z of the intermediate layer are defined according to the number of each intermediate layer.

Next, in the output layer, the load sum u is output from the load sum z and the weight function W output from the intermediate layer. For example, when the number of intermediate layers is 1, as shown in FIG. 9, the output layer weight function is W _jk ⁽²⁾ and its load sum u _k (formula E) is output. When the number of intermediate layers is 2, the output layer weight function is W _kl ⁽³⁾ and the load sum is u _l , and when the number of intermediate layers is 3, the output layer weight function is W _lm ⁽⁴⁾ and the load sum. U _m . . . Thus, the weight function W and the load sum u of the output layer are defined according to the number of intermediate layers.

However, in (Formula B) and (Formula C), (m−1) types of first parameters X and (M−m) types of residues Y ^(k) are input to the input layer, respectively. M, n, and M are natural numbers.

  Thereafter, the weight function W of the intermediate layer and the output layer in the RNN analysis for each number of intermediate layers is determined so that the load sum u output from the output layer is minimized. As a method for obtaining these weighting functions W, various known methods can be used. For example, a BPTT method, an RTRL method, or the like can be used. Next, a temporary correction prediction model is created for each RNN analysis of each number of intermediate layers using the weight function W thus obtained. Next, the error of each temporary correction prediction model is evaluated. For example, statistical indexes such as equations (16) to (18) can be used for error evaluation.

  The temporarily corrected prediction model evaluated as the model with the least error by the statistical index as described above is determined as a corrected prediction model obtained by correcting the prediction model by combining the RLS analysis with the PLS analysis, and is stored in the model storage unit. . Further, when an output command is input, the output means predicts a specific parameter ahead by the prediction time by using the modified prediction model stored in the model storage means.

  In this prediction apparatus, when the model update condition input in advance to the update condition input unit is further connected to the update command unit, the prediction model is again created by the model creation unit and the modified prediction by the model correction unit. It may be configured to create a model. Thus, the second parameter can be accurately predicted over a long period by updating the prediction model in accordance with the model update condition. Note that the model update process can be performed by the same procedure and method as the second prediction apparatus.

(experimental method)
With the prediction device of the present invention, the parameters (characteristics) of the distillation tower in Mizushima Refinery, Nippon Oil Refining Co., Ltd. were predicted. First, sensors were attached to 28 locations of the distillation tower, and characteristic values such as temperature, pressure, and raw material flow rate were measured by each sensor. The types of measured parameters and the positions in the distillation column are shown in FIG. In this measurement, 1200 points of the first parameter (X _{1 to} X ₂₇ ) and the second parameter (Y ₁ ) were measured every 3 minutes.

Analysis is performed using at least a part of the data of the first 1 to 500 points among the 1200 points, the measured first parameter (X _{1 to} X ₂₇ ), and the second parameter (Y ₁ ) sample, and a prediction model is obtained. Created. Next, data of 501 to 1000 points out of 501 to 1200 points was predicted using this prediction model. The prediction procedure is shown below.

The data of 200 points was extracted from the data of 500 points, and the analysis was performed with the estimated time being 15 minutes (delay time of 5 samples).
In Example 1, the second parameter of the 501st to 600th points was predicted by PLS analysis in which the number of latent variables was 5.
In Example 2 and 3, the analysis which combined PLS analysis and RNN analysis was performed. At this time, the number of latent variables for PLS analysis was set to 5, and the number of intermediate layers for RNN analysis was set to 10. In Example 2, the second parameter of the 501st to 600th points was predicted, and in Example 3, the second parameter of the 601st to 1000th points was predicted.

The accuracy of each prediction model was evaluated from the error with respect to the actual measurement value Y ₁ of the second parameter predicted as described above. The error was evaluated by RMSE (formula (16)) and INDEX (formula (18)). These results are shown in Table 1.

  From the result of Example 1, it can be seen that the second parameter can be predicted with high accuracy by the prediction model created by the PLS analysis. Moreover, it can be seen from the results of Examples 1 and 2 that the prediction accuracy is further improved by using the PLS analysis and the RNN analysis as compared with the case of the PLS analysis alone. In addition, the results of Example 3 show that the prediction model maintains the high prediction accuracy even when a certain amount of time has elapsed after the creation of the prediction model, from 601 to 1000 points.

It is a schematic diagram which shows an example of the internal structure of the prediction apparatus of this invention. It is a schematic diagram showing a distillation apparatus. It is a conceptual diagram showing the measurement state of the parameter of this invention. It is a conceptual diagram showing the sampling state of the parameter of this invention. It is a conceptual diagram showing the sampling state of the parameter of this invention. It is a schematic diagram which shows the internal structure of the prediction apparatus of this invention. It is a flowchart which shows the process in which the optimized prediction model is produced. It is a schematic diagram which shows an example of the internal structure of the prediction apparatus of this invention. It is a conceptual diagram showing the method of the RNN analysis of this invention. It is a flowchart which shows the process in which the optimized prediction model is produced. It is a figure showing the measurement position of the parameter in the distillation column in an Example.

Explanation of symbols

50 Oil refinery plant 51 Input means 52 First measurement means 53 Second measurement means 54 Parameter storage means 55 Model creation means 56 Model storage means 57 Output means 58 Input unit 59 Update command means 61 Predetermined time 62 Sampled first parameter 64 Sampled second parameter 71 Model correction means

Claims (14)

Using the first and second parameters measured in the oil refining plant, a prediction model for predicting the second parameter ahead by the prediction time from the time of measurement of the first parameter is created, and the second parameter is created using the prediction model. A prediction device for predicting
Input means for inputting first and second parameter measurement conditions and model creation conditions;
Measurement means for measuring the first and second parameters from the oil refinery plant based on the first and second parameter measurement conditions;
Parameter storage means for storing the first and second parameters;
Based on the model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction optimized with respect to the prediction time and the first and second parameters used for model creation A model creation means for creating a model;
Model storage means for storing the optimized prediction model;
An output unit configured to output a second parameter predicted by the optimized prediction model. The prediction device further includes
An update instruction means for issuing an update instruction for updating the prediction model when the model update condition is satisfied, having an update condition input unit for inputting a model update condition,
The prediction apparatus according to claim 1, wherein when the update command is issued, the prediction apparatus is configured to create a newly optimized prediction model.   The prediction apparatus according to claim 2, wherein the model update condition is a case where a predetermined time has elapsed since the creation of the optimized prediction model. The output means periodically predicts the second parameter,
The measuring means measures the second parameter at a time when the second parameter is predicted;
The prediction apparatus according to claim 2, wherein the model update condition is a case where an error between the predicted second parameter and the measured second parameter is determined to be an abnormal state by a statistical index. Using the first and second parameters measured in the oil refining plant, a prediction model for predicting the second parameter ahead by the prediction time from the time of measurement of the first parameter is created, and the second parameter is created using the prediction model. A prediction device for predicting
Input means for inputting first and second parameter measurement conditions, model creation conditions and model correction conditions;
Measurement means for measuring the first and second parameters from the oil refinery plant based on the first and second parameter measurement conditions;
Parameter storage means for storing the first and second parameters;
Based on the model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction optimized with respect to the prediction time and the first and second parameters used for model creation A model creation means for creating a model;
Based on the model correction condition, a model correction unit that further performs an RNN (Recurrent Natural Network) analysis on the prediction model created by the model creation unit to create a modified prediction model obtained by correcting the prediction model;
Model storage means for storing the modified prediction model;
An output means for outputting a second parameter predicted by the modified prediction model. The model creation means includes
The prediction time is changed within the range of the prediction time specified by the model creation condition, and the first and second parameters are determined according to a predetermined condition from the measured first and second parameters for each changed prediction time. A temporary model creating means for sampling and performing a PLS analysis on the sampled first and second parameters to create a plurality of temporary prediction models;
The prediction apparatus according to claim 1, further comprising: a model determination unit that evaluates accuracy of each temporary prediction model and determines an optimal prediction model. Sampling of the first parameter according to the predetermined condition is a predetermined number of random samplings,
7. The prediction according to claim 6, wherein the sampling of the second parameter according to the predetermined condition is a sampling of the second parameter measured for the prediction time earlier than the sampled first parameter. apparatus.   The prediction apparatus according to claim 6, wherein the accuracy of each temporary prediction model is evaluated using a statistical index. The model creation means further includes
The prediction apparatus according to claim 1, wherein a prediction model in which the number of model configuration variables constituting the prediction model is optimized is created. The oil refinery plant is a distillation column;
The first and second parameters are each independently at least one parameter selected from the group consisting of pressure, temperature, raw material supply flow rate, and product discharge flow rate in a predetermined region of the distillation column. The prediction apparatus according to any one of claims 1 to 9. Using the first and second parameters measured in the oil refining plant, a prediction model for predicting the second parameter ahead by the prediction time from the time of measurement of the first parameter is created, and the second parameter is created using the prediction model. Is a method for predicting
A process for measuring the first and second parameters from the oil refinery plant based on the input first and second parameter measurement conditions;
A process for storing the first and second parameters;
PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters based on the input model creation conditions, and the prediction time and the first and second parameters used for model creation are optimized. A process of creating the prediction model;
Storing the optimized prediction model;
And a process of outputting a second parameter predicted by the optimized prediction model. Using the first and second parameters measured in the oil refining plant, a prediction model for predicting the second parameter ahead by the prediction time from the time of measurement of the first parameter is created, and the second parameter is created using the prediction model. Is a method for predicting
A process for measuring the first and second parameters from the oil refinery plant based on the input first and second parameter measurement conditions;
A process for storing the first and second parameters;
Based on the input model creation conditions, PLS (Partial Least Squares) analysis is performed on at least a part of the first and second parameters, and the prediction time and the first and second parameters used for model creation are optimized. A process of creating the prediction model;
A process of creating a modified prediction model in which the prediction model is further corrected by performing an RNN (Recursive Neural Network) analysis on the prediction model based on the input model modification condition;
Storing the modified prediction model;
And a process of outputting a second parameter predicted by the modified prediction model. The process of creating the prediction model is as follows:
A process of changing the prediction time within the range of the prediction time specified by the model creation condition;
A process of sampling the first and second parameters in accordance with a predetermined condition from the measured first and second parameters at each changed predicted time;
A process of performing a PLS analysis on the sampled first and second parameters to create a plurality of temporary prediction models;
The prediction method according to claim 11, further comprising: a process of evaluating an accuracy of each temporary prediction model and determining an optimal prediction model. Sampling of the first parameter according to the predetermined condition is a predetermined number of random samplings,
14. The prediction according to claim 13, wherein sampling of the second parameter according to the predetermined condition is sampling of the second parameter measured in advance for the prediction time with respect to the sampled first parameter. Method.

Images