Plant control assistance device, plant control assistance method, and plant control assistance program

Abstract

The present application relates to a plant control assistance device for assisting with plant control in which a plurality of process values of the plant are used, said process values having been respectively predicted using a plurality of prediction models. This device has a first prediction model training unit for training a plurality of first prediction models as a plurality of prediction models, using training data that is operation data of the plant. For a first prediction model which is among the plurality of first prediction models and in which the prediction accuracy is less than a reference value, a second prediction model is retrained using, as training data, the operation data to which additional data has been added. For a second prediction model in which the prediction accuracy is less than the reference value, a third prediction model is retrained using, as training data, similar operation data which is among the operation data and is similar to an operation point of the plant. The prediction model with the best prediction accuracy among the first prediction model, the second prediction model, and the third prediction model is selected as a prediction model.

Description

Plant control support device, plant control support method, and plant control support program

[0001] This application claims priority to Japanese Patent Application No. 2024-070327, filed on April 24, 2024, with the Japan Patent Office, the contents of which are incorporated herein by reference.

[0002] Control devices for controlling plants including various devices are known. In these types of control devices, control is performed based on process values ​​related to the characteristics of the plant. However, when the process values ​​are parameters that are difficult to measure, or when the process values ​​are parameters that are measurable but that are preferable to avoid measuring, for example, for reasons such as cost reduction, a prediction model that simulates the behavior of the plant may be used to obtain the process values ​​as calculated predicted values. Such prediction models can be constructed, for example, by machine learning using past operating data of the plant as learning data. For example, Patent Document 1 discloses that by training a prediction model for each operating condition of the plant, it is possible to appropriately predict process values ​​even when the operating conditions are changed.

[0003] Patent No. 7222943

[0004] In the above-mentioned Patent Document 1, a prediction model is constructed by performing learning for each driving condition, but the number of data points that can be used as learning data may differ depending on the driving condition. In this case, it may be impossible to prepare sufficient learning data depending on the driving condition, which may result in a decrease in the prediction accuracy of the prediction model constructed by learning.

[0005] At least one embodiment of the present disclosure has been made in consideration of the above-mentioned circumstances, and aims to provide a plant control support device, a plant control support method, and a plant control support program that can achieve good prediction accuracy of process values ​​using a prediction model regardless of the operating conditions of the plant.

[0006] In order to solve the above problem, a plant control assistance device according to at least one embodiment of the present disclosure is a plant control assistance device for assisting plant control using a plurality of process values ​​of a plant that are respectively predicted using a plurality of prediction models, the plant control assistance device including: a first prediction model training unit configured to train a plurality of first prediction models as the plurality of prediction models by using learning data that are operating data of the plant; a second prediction model training unit configured to re-train a second prediction model using operating data to which additional data has been added as learning data for a first prediction model among the plurality of first prediction models whose prediction accuracy is less than a reference value; a third prediction model training unit configured to re-train a third prediction model using similar operating data that is similar to an operating point of the plant among the operating data for the second prediction model whose prediction accuracy is less than the reference value as the learning data; and a prediction model selection unit configured to select, as the prediction model, one of the first prediction model, the second prediction model, or the third prediction model that has the best prediction accuracy.

[0007] In order to solve the above-described problems, a plant control support method according to at least one embodiment of the present disclosure is a plant control support method for supporting plant control using a plurality of process values ​​of a plant that are respectively predicted using a plurality of prediction models, the method including: a step of training a plurality of first prediction models as the plurality of prediction models by using training data that are operating data of the plant; a step of re-training a second prediction model using, as the training data, operating data to which additional data has been added, for a first prediction model among the plurality of first prediction models whose prediction accuracy is less than a reference value; a step of re-training a third prediction model using, as the training data, similar operating data that is similar to an operating point of the plant among the operating data, for a second prediction model whose prediction accuracy is less than the reference value; and a step of selecting, as the prediction model, one of the first prediction model, the second prediction model, or the third prediction model having the best prediction accuracy.

[0008] In order to solve the above-described problems, a plant control assistance program according to at least one embodiment of the present disclosure is a plant control assistance program for assisting plant control using a plurality of process values ​​of a plant that are respectively predicted using a plurality of prediction models, the program being executable by a computer device: a step of training a plurality of first prediction models as the plurality of prediction models by using training data that are operating data of the plant; a step of re-training a second prediction model using, as the training data, the operating data to which additional data has been added, for a first prediction model among the plurality of first prediction models whose prediction accuracy is less than a reference value; a step of re-training a third prediction model using, as the training data, similar operating data that is similar to an operating point of the plant among the operating data, for the second prediction model whose prediction accuracy is less than the reference value; and a step of selecting, as the prediction model, one of the first prediction model, the second prediction model, or the third prediction model, which has the best prediction accuracy.

[0009] According to at least one embodiment of the present disclosure, it is possible to provide a plant control support device, a plant control support method, and a plant control support program that can achieve good prediction accuracy of process values ​​using a prediction model regardless of the operating conditions of the plant.

[0010] It is an overall configuration diagram of a plant control system according to one embodiment. It is a schematic diagram showing a virtual space that defines an operating point of a plant. It is an explanatory diagram showing a calculation flow of a predicted value using a prediction model stored in a prediction model storage unit of Figure 1. It is an explanatory diagram showing a calculation flow of an index used for searching for an operating condition in an operating condition searching unit of Figure 1. It is a flowchart showing a plant control support method according to one embodiment.

[0011] Hereinafter, several embodiments of the present disclosure will be described with reference to the accompanying drawings. However, the dimensions, materials, shapes, relative arrangements, etc. of the configurations described as embodiments or shown in the drawings are merely illustrative examples and are not intended to limit the scope of the present disclosure.

[0012] First, the overall configuration of a plant control system 1 including a plant control assistance device 100 according to at least one embodiment of the present disclosure will be described. Fig. 1 is an overall configuration diagram of a plant control system 1 according to one embodiment.

[0013] The plant control assistance device 100 is a device for assisting control performed by a control device 200, which is a control unit for controlling a plant 10. The plant 10 is, for example, a boiler device for generating steam to be supplied to a steam turbine (not shown) connected to a generator in a thermal power plant, but this is merely an example and is not limiting.

[0014] The control device 200 controls the operating state of the plant 10 by transmitting and receiving various signals to and from the plant 10. The control by the control device 200 is performed in accordance with preset operating conditions, and the control device 200 acquires operating parameters from the plant 10 and outputs control signals corresponding to the operating parameters to the plant 10. In the plant 10, each component device is controlled based on the control signals input from the control device 200. Note that the detailed configuration of the control device 200 will be similar to known examples and will not be described here.

[0015] The plant control assistance device 100 includes an external input interface 102, a preprocessing unit 104, an operating data accumulation unit 106, a prediction model learning unit 108, a prediction model storage unit 110, a prediction accuracy determination unit 112, an operating condition search unit 114, a search range setting unit 116, and an external output interface 118.

[0016] The external input interface 102 is an interface configuration for inputting data from a control device 200, which is an external device to the plant control assistance device 100. Various data handled by the control device 200 is input to the external input interface 102. This data includes data that becomes operating data that is processed by the preprocessing unit 104 and handled as learning data in the machine learning of the prediction model M in the prediction model learning unit 108. Furthermore, data input from the control device 200 is performed repeatedly at predetermined time intervals.

[0017] The preprocessing unit 104 is configured to perform preprocessing on the data input from the external input interface 102. The preprocessing includes at least a process of extracting driving data from the data input to the external input interface 102, the driving data being used as learning data in the machine learning of the prediction model M in the prediction model learning unit 108, but the detailed content of the process is not particularly limited.

[0018] The operating data storage unit 106 is a component (e.g., a database) for storing the operating data created by the preprocessing unit 104. In the plant control assistance device 100, every time data is input from the control device 200 to the external input interface 102, preprocessing is performed in the preprocessing unit 104 to process the operating data, and the processed operating data is sequentially stored in the operating data storage unit 106.

[0019] The prediction model learning unit 108 is configured to construct a prediction model M by performing machine learning using the driving data accumulated in the driving data accumulation unit 106 as learning data. This machine learning specifies explanatory variables and an objective function from each parameter included in the driving data, and learns the correlation between the explanatory variables and the objective variable according to a machine learning algorithm, thereby constructing the prediction model M. As the machine learning algorithm, linear regression, neural network, random forest, or the like can be adopted, but is not limited to these.

[0020] The prediction model training unit 108 includes a first prediction model training unit 120 , a second prediction model training unit 122 , and a third prediction model training unit 124 .

[0021] The first prediction model learning unit 120 is configured to construct a prediction model (hereinafter referred to as the "first prediction model M1" as appropriate) by machine learning using the driving data accumulated in the driving data accumulation unit 106 as learning data. The machine learning by the first prediction model learning unit 120 is performed before the machine learning by the second prediction model learning unit 122 and the third prediction model learning unit 124, and is machine learning for constructing a so-called original prediction model M.

[0022] The second prediction model learning unit 122 is configured to build a prediction model (hereinafter referred to as the "second prediction model M2") by additional driving data relearning. In the additional driving data relearning, the second prediction model M2 can be built by relearning using driving data that was used as learning data when the first prediction model M1 was built, with additional data added as learning data. As described above, driving data is sequentially accumulated in the driving data accumulation unit 106, and the additional data is driving data that has been accumulated in the driving data accumulation unit 106 after the first prediction model M1 was learned. In the additional driving data relearning, by performing relearning using driving data that has a larger amount of data due to the addition of the additional data, it is possible to build a second prediction model M2 with improved prediction accuracy compared to the first prediction model M1.

[0023] The third prediction model learning unit 124 is configured to construct a prediction model (hereinafter referred to as the "third prediction model M3") by similar operating data relearning. In the similar operating data relearning, the third prediction model M3 can be constructed by performing similar operating data relearning, using, as learning data, similar operating data that belongs to a predetermined operating region including the operating point of the plant 10, among the operating data stored in the operating data accumulation unit 106.

[0024] Here, the similar operating data treated as learning data by the third prediction model learning unit 124 will be specifically described with reference to Fig. 2. Fig. 2 is a schematic diagram showing a virtual space V that defines the operating point of the plant 10. Fig. 2 shows the current operating point of the plant in the virtual space V, which is a multidimensional space defined with multiple parameters representing the operating point as spatial axes (for ease of illustration, Fig. 2 illustrates a case where the number of dimensions of the virtual space V is "3").

[0025] In FIG. 2, if the coordinate A corresponding to the current operating point of the plant in a virtual space V defined by three spatial axes x, y, and z is (x1, y1, z1), and the coordinate of an arbitrary point B in the virtual space is (x2, y2, z2), the degree of deviation L between the two can be expressed by the following equation: L = {(x1 - x2) ２ +(y1-y2) ２ +(z1-z2)２} ０．５ The similar driving data is defined as, for example, data included in a similar driving range in which the deviation L expressed in this way is within a predetermined value L0 (in FIG. 2 , the similar driving range is schematically shown as being within a sphere indicated by a dashed line). Thus, in the similar driving data relearning, driving data at operating points with similar driving conditions is used as learning data for relearning, thereby suitably improving the prediction accuracy of the prediction model M.

[0026] Although the example here illustrates a case in which the deviation L is defined using Euclidean distance, the deviation L may also be defined using Mahalanobis distance. Furthermore, the similar driving data may be defined as data extracted from the driving data in ascending order of deviation L until a predetermined percentage or a predetermined number is reached. As described above, if the similar driving data is defined as data whose deviation L is within a predetermined value L0, if there is not a sufficient amount of data within that range, sufficient learning data cannot be extracted, and learning of the prediction model M itself may become impossible. In contrast, in this aspect, by defining the similar driving data as data extracted until a predetermined percentage or a predetermined number is reached, even if there is little driving data with a small deviation L, a shortage of data necessary for learning can be prevented, and re-learning of the similar driving data can be performed (this is equivalent to setting the predetermined value L0 to a variable value and determining that a predetermined number or a predetermined percentage of driving data with a small deviation L is extracted from the driving data).

[0027] Returning to Fig. 1 , the prediction model storage unit 110 is configured to store the prediction model M constructed by the prediction model learning unit 108. The prediction model learning unit 108 constructs at least one of a first prediction model M1, a second prediction model M2, and a third prediction model M3 as the prediction model M, and these prediction models M are stored in the prediction model storage unit 110, making them available for use as needed.

[0028] The prediction accuracy determination unit 112 is configured to determine the prediction accuracy of each prediction model M (first prediction model M1, second prediction model M2, or third prediction model M3) stored in the prediction model storage unit 110. The prediction accuracy of the prediction model M is calculated, for example, by aggregating the mean absolute error of predicted values ​​for verification data such as actual measured values, but any known method can be used and is not limited thereto. The prediction accuracy determination unit 112 compares the calculated prediction accuracy with a reference value stored in advance in the acceptable accuracy database 126, thereby making it possible to determine whether the prediction accuracy of the prediction model M is below the reference value.

[0029] The operation condition search unit 114 is configured to search for operation conditions of the control device 200 using the prediction model M stored in the prediction model storage unit 110. The operation condition search is performed by calculating an index corresponding to the performance of the plant 10 based on a predicted value calculated using the prediction model M, and within a search range set by the search range setting unit 116 so that the index becomes optimal.

[0030] Here, FIG. 3 is an explanatory diagram showing the calculation flow of a predicted value P using the prediction model M stored in the prediction model storage unit 110 of FIG. 1, and FIG. 4 is an explanatory diagram showing the calculation flow of an index used to search for an operation condition in the operation condition search unit 114 of FIG. 1.

[0031] As shown in Fig. 3, the prediction model M outputs a predicted value P corresponding to the objective variable when parameters corresponding to the explanatory variables are input. For example, if the plant 10 is a boiler system, the predicted value P is a process value related to the operating state of the boiler system, more specifically, the temperature and pressure of steam in each part of the boiler system, the metal temperature, or the NOx concentration emitted from the boiler system. The explanatory variables are various operating parameters of the boiler system, more specifically, the angle of the burner that injects fuel, the opening degrees of various dampers that adjust the air volume, etc. The prediction model learning unit 108 constructs a prediction model M according to each type of objective variable, and the prediction model storage unit 110 stores multiple prediction models M according to each type of objective variable.

[0032] Note that, for a prediction model M corresponding to a certain dependent variable, if a second prediction model M2 or a third prediction model M3 is stored in addition to a first prediction model M1, the prediction model with the best prediction accuracy is used to calculate the predicted value P. In other words, the predicted values ​​P of multiple dependent variables are calculated using multiple corresponding prediction models M, and each prediction model M is selected from the prediction models for the corresponding dependent variable with the best prediction accuracy (at least one of the first prediction model M1, the second prediction model M2, and the third prediction model M3).

[0033] The operation condition search unit 114 inputs input parameters (initial values) corresponding to preset operation conditions into a plurality of prediction models M stored in the prediction model storage unit 110, thereby obtaining respective predicted values ​​P (hereinafter, when distinguishing between them, they will be appropriately referred to as "predicted values ​​P1, P2, ..."). These predicted values ​​P are converted into an index IN for evaluating the predicted values ​​P using a predetermined conversion formula. Specifically, as shown in FIG. 4 , the operation condition search unit 114 converts each predicted value P1, P2, ... into a corresponding score SC1, SC2, ..., and calculates the index IN (= SC1 + SC2 + ...) as the sum of these scores. The operation condition search unit 114 then searches for operation conditions that will optimize the index IN (e.g., minimum or maximum) (i.e., by repeatedly changing the operation conditions and calculating the index IN, the operation condition that optimizes the index IN is identified).

[0034] The search for operation conditions in the operation condition search unit 114 is performed within a search range set by the search range setting unit 116. This search range is set to a first search range in the default state, but when a search range limiting process is executed as described below, it is limited to a second search range that is specified to be narrower than the first search range. By limiting the search range for operation conditions to a range where the prediction accuracy of the prediction model M is good or a range that does not significantly impair the driving state, it is possible to substantially improve the prediction accuracy and the driving reliability.

[0035] The operating conditions searched for by the operating condition search unit 114 are output to the control device 200 via the external output interface 118. The control device 200 controls the plant 10 in accordance with the operating conditions thus output, thereby suitably controlling the plant 10.

[0036] Next, a description will be given of a plant control support method implemented using the plant control support device 100 having the above configuration. Fig. 5 is a flowchart showing a plant control support method according to one embodiment.

[0037] First, operating data is collected in the operating data accumulation unit 106 (step S100). In step S100, operating data is input from the control device 200 that controls the operating plant 10 to the plant control assistance device 100 via the external input interface 102, and the data is processed into operating data by preprocessing performed by the preprocessing unit 104. The processed operating data is collected by being sequentially accumulated in the operating data accumulation unit 106. Such accumulation of operating data is performed sequentially as described above, and in the following description, it is assumed that sufficient operating data is accumulated in the operating data accumulation unit 106 to train the prediction model M (first prediction model M1).

[0038] Next, the prediction model learning unit 108 constructs a first prediction model M1 through machine learning using the driving data accumulated in the driving data accumulation unit 106 as learning data (step S101). In step S101, the first prediction model learning unit 120 of the prediction model learning unit 108 constructs the first prediction model M1 as the prediction model M. The first prediction model learning unit 120 constructs the first prediction model M1 based on at least one type of machine learning algorithm prepared in advance. In particular, in this embodiment, the first prediction model learning unit 120 has multiple types of machine learning algorithms prepared in advance, and generates multiple prediction model candidates corresponding to the first prediction model M1 using the multiple machine learning algorithms for each dependent variable. The first prediction model learning unit 120 then calculates the prediction accuracy of the multiple prediction model candidates corresponding to the first prediction model M1, and adopts the one with the best prediction accuracy as the first prediction model M1 for that dependent variable. The adopted first prediction model M1 is stored in the prediction model storage unit 110. The first prediction model M1 stored in the prediction model storage unit 110 can be accessed by the operation condition search unit 114 as needed.

[0039] The learning of the first prediction model M1 in step S101 may be performed every predetermined period T1. That is, machine learning may be performed every predetermined period T1 using the driving data accumulated in the driving data accumulation unit 106 as learning data, thereby updating the first prediction model M1 based on the latest driving data.

[0040] Next, the prediction accuracy determination unit 112 determines whether the prediction accuracy of the first prediction model M1 constructed in step S101 is less than a predetermined reference value (step S102). In step S102, the prediction accuracy of the first prediction model M1 is calculated by accessing the first prediction model M1 stored in the prediction model storage unit 110 every predetermined period T2, and the calculated prediction accuracy is compared with the reference value. This predetermined period T2 may be different from the aforementioned predetermined period T1, and in particular may be shorter than the predetermined period T1.

[0041] Next, if the prediction accuracy is less than the reference value (step S102: YES), the prediction model learning unit 108 causes the second prediction model learning unit 122 to perform additional driving data relearning, thereby learning the second prediction model M2 (step S103). In the additional driving data relearning, the second prediction model M2 is constructed by relearning using the driving data that was used as the learning data when the first prediction model M1 was constructed in step S101, but to which additional data has been added. Therefore, if the prediction accuracy of the first prediction model M1 is less than the reference value, the additional driving data relearning is performed using driving data with a larger amount of data due to the addition of the additional data, thereby making it possible to construct a second prediction model M2 with improved prediction accuracy compared to the first prediction model M1.

[0042] Note that the second prediction model M2 may be constructed in the second prediction model training unit 122 using at least one type of machine learning algorithm, as in the first prediction model training unit 120 described above. In particular, in this embodiment, the second prediction model training unit 122 is provided with multiple types of machine learning algorithms, and multiple prediction model candidates corresponding to the second prediction model M2 are generated for each dependent variable using the multiple machine learning algorithms. The second prediction model training unit 122 then calculates the prediction accuracy of the multiple prediction model candidates corresponding to the second prediction model M2, and adopts the one with the highest prediction accuracy as the second prediction model M2 for that dependent variable. The constructed second prediction model M2 is stored in the prediction model storage unit 110 instead of or in addition to the first prediction model M1 described above. The second prediction model M2 stored in the prediction model storage unit 110 can be accessed as needed by the operation condition search unit 114.

[0043] Next, the prediction accuracy determination unit 112 determines whether the prediction accuracy of the second prediction model M2 constructed in step S103 is less than a predetermined reference value (step S104). In step S104, similar to step S102, the prediction accuracy of the second prediction model M2 is calculated and compared with the reference value.

[0044] The reference value used as the determination criterion in step S104 may be the same as the reference value used as the determination criterion in step S102, or may be different.

[0045] Subsequently, if the prediction accuracy is less than the reference value (step S104: YES), the prediction model learning unit 108 causes the third prediction model learning unit 124 to re-learn similar operating data (step S105), or causes the search range setting unit 116 to perform search range restriction processing (step S106). As shown in Fig. 1, the plant control assistance device 100 according to one embodiment includes a display unit 130 and a selection unit 132, and thereby allows the operator to selectively perform either step S105 or S106.

[0046] The display unit 130 is a display device such as a display for displaying various information necessary for the operator. The information displayed on the display unit 130 may include a wide range of information used when the operator selects either step S105 or S106, and may include, for example, the prediction accuracy calculated in the process of making the determination in the prediction accuracy determination unit 112.

[0047] The selection unit 132 is configured to allow an operator to select step S105 or S106 while referring to the information displayed on the display unit 130, and may be an input device such as a keyboard, a mouse, or a touch panel. The operator can select whether to perform step S105 or S106 when the prediction accuracy of the second prediction model M2 is below the reference value by referring to the information displayed on the display unit 130 and operating the selection unit 132 based on their own judgment.

[0048] In general, the search range limiting process tends to require a shorter calculation time than the similar driving data relearning process. Therefore, as a selection criterion for the selector 132, for example, step S105 can be selected when a high-quality prediction model M with good prediction accuracy is required, and step S106 can be selected when priority is given to shortening the calculation time.

[0049] When similar driving data relearning is performed in step S105, the third prediction model M3 is constructed using the similar driving data as learning data by the third prediction model learning unit 124. In the similar driving data relearning, by performing relearning using driving data at operating points with similar driving states as learning data, it is possible to obtain the third prediction model M3 with improved prediction accuracy.

[0050] Note that the construction of the third prediction model M3 in the third prediction model training unit 124 may be performed using at least one type of machine learning algorithm, similar to the construction of the first prediction model M1 in the first prediction model training unit 120 and the construction of the second prediction model M2 in the second prediction model training unit 122. In this embodiment, particularly, the third prediction model training unit 124 is provided with multiple types of machine learning algorithms, and multiple prediction model candidates corresponding to the third prediction model M3 are generated using the multiple machine learning algorithms for each dependent variable. The third prediction model training unit 124 then calculates the prediction accuracy of the multiple prediction model candidates corresponding to these multiple third prediction models M3, and adopts the one with the highest prediction accuracy as the third prediction model M3 for that dependent variable. The constructed third prediction model M3 is stored in the prediction model storage unit 110 instead of or in addition to the second prediction model M2. The third prediction model M3 stored in the prediction model storage unit 110 can be accessed as needed by the operation condition search unit 114.

[0051] On the other hand, when the search range limiting process is performed in step S106, the search range when searching for operating conditions in step S110 (described later) is limited to a specified range, thereby improving the feasibility of optimizing driving using a predictive model. For example, when the search range is limited to a range in which the prediction accuracy of the predictive model M is equal to or greater than a threshold, the prediction accuracy can be substantially improved. Furthermore, when the search range is limited to a range that does not significantly impair the driving state, driving can be optimized using the predictive model M while maintaining safe driving.

[0052] Next, for each of the multiple prediction models M, the prediction model with the best prediction accuracy is selected from the first prediction model M1, the second prediction model M2, and the third prediction model M3 (step S107). As described above, multiple prediction models are prepared depending on the type of dependent variable, and for each of them, at least one of the first prediction model M1, the second prediction model M2, and the third prediction model M3 is constructed in the above-mentioned step. In step S107, the prediction model with the best prediction accuracy is adopted from the first prediction model M1, the second prediction model M2, and the third prediction model M3.

[0053] Subsequently, operation conditions are searched for using each prediction model M selected in step S107 (step S108), and the searched operation conditions are output to the control device 200 via the external output interface 118 (step S109). In step S108, as described above with reference to Fig. 4, an index IN is calculated based on each prediction value P using each prediction model M selected in step S107, and operation conditions are searched for so that the index IN is optimized.

[0054] It should be noted that the search range of the operation conditions in step S108 is limited to a specified range if the search range limiting process has been performed in step S106.

[0055] In addition, within the scope of the present disclosure, the components in the above-described embodiments may be replaced with well-known components as appropriate, and the above-described embodiments may be combined as appropriate.

[0056] The contents described in each of the above embodiments can be understood, for example, as follows.

[0057] (1) A plant control assistance device according to one aspect is a plant control assistance device for assisting plant control using a plurality of process values ​​of a plant that are respectively predicted using a plurality of prediction models, the plant control assistance device including: a first prediction model training unit that trains a plurality of first prediction models as the plurality of prediction models by using learning data that are operating data of the plant; a second prediction model training unit that, for a first prediction model among the plurality of first prediction models whose prediction accuracy is less than a reference value, re-trains a second prediction model by using, as the learning data, operating data to which additional data has been added; a third prediction model training unit that, for the second prediction model whose prediction accuracy is less than the reference value, re-trains a third prediction model by using, as the learning data, similar operating data that is similar to an operating point of the plant among the operating data; and a prediction model selection unit that selects, as the prediction model, one of the first prediction model, the second prediction model, or the third prediction model that has the best prediction accuracy.

[0058] According to the above aspect (1), multiple first prediction models corresponding to multiple process values ​​to be predicted are constructed using operating data as learning data. For those first prediction models whose prediction accuracy is below a reference value, additional operating data relearning is performed using the operating data to which additional data has been added as learning data, thereby constructing a second prediction model with improved prediction accuracy. If the prediction accuracy of the second prediction model is still below the reference value, similar operating data relearning is performed using similar operating data similar to an operating point of the plant from the operating data, thereby improving the prediction accuracy of a third prediction model. In the plant control device, the first prediction model, the second prediction model, or the third prediction model with the best prediction accuracy is selected as each prediction model for predicting the multiple process values. This enables accurate prediction of each process value, thereby suitably improving the controllability of the plant control device.

[0059] (2) In another aspect, in the aspect (1) above, at least one of the first prediction model learning unit, the second prediction model learning unit, or the third prediction model learning unit generates multiple prediction model candidates for each of the multiple prediction models using multiple types of machine learning algorithms, and selects the prediction model candidate with the best prediction accuracy as the first prediction model, the second prediction model, or the third prediction model.

[0060] According to the above aspect (2), at least one of the first prediction model training unit, the second prediction model training unit, and the third prediction model training unit generates multiple prediction model candidates using multiple types of machine learning algorithms.Then, by selecting the prediction model with the best prediction accuracy from among the multiple prediction model candidates, it becomes possible to build a prediction model with excellent prediction accuracy.

[0061] (3) In another aspect, in the aspect (1) or (2) above, the similar driving data is constructed by extracting a predetermined proportion or number of data that have a small deviation from the driving point from the driving data and the additional data.

[0062] According to the above aspect (3), the similar driving data used as learning data in the similar driving data relearning is configured by extracting a predetermined proportion or number of data that have a small deviation from the operating point from the driving data and the additional data. As a result, in the similar driving data relearning, by performing relearning using such similar driving data, it is possible to suitably improve the prediction accuracy of the prediction model.

[0063] (4) In another aspect, in any one of the above aspects (1) to (3), the additional data is data collected from the plant since the previous learning of the first prediction model.

[0064] According to the above aspect (4), the learning data used in the additional operating data relearning includes additional data collected from the plant between the previous learning and the relearning, in addition to the operating data used in the previous learning of the prediction model. This makes it possible to suitably improve the prediction accuracy of the prediction model by relearning using learning data that contains a larger amount of data than in the previous learning.

[0065] (5) In another aspect, in any one of the above aspects (1) to (4), the system further includes an index calculation unit for calculating an index related to the performance of the plant based on the plurality of process values ​​predicted using the plurality of prediction models; an operating condition search unit for searching for operating conditions for the plant based on the index; and a control unit for controlling the plant based on the operating conditions.

[0066] According to the above aspect (5), by using the predicted value of the process value obtained using the prediction model with improved prediction accuracy as described above, it is possible to suitably search for the operating conditions of the plant and control the plant.

[0067] (6) In another aspect, in any one of the above aspects (1) to (5), the plant is a boiler system.

[0068] According to the above aspect (6), the process value relating to the characteristics of the boiler apparatus can be suitably predicted using a prediction model with good prediction accuracy.

[0069] (7) A plant control assistance method according to one aspect is a plant control assistance method for assisting plant control using a plurality of process values ​​of a plant that are respectively predicted using a plurality of prediction models, the method comprising: a step of training a plurality of first prediction models as the plurality of prediction models by using training data that are operating data of the plant; a step of re-training a second prediction model using the operating data to which additional data has been added as the training data for a first prediction model among the plurality of first prediction models, the first prediction model having a prediction accuracy less than a reference value; a step of re-training a third prediction model using similar operating data that is similar to an operating point of the plant among the operating data for the second prediction model having the prediction accuracy less than the reference value as the training data; and a step of selecting, as the prediction model, one of the first prediction model, the second prediction model, or the third prediction model having the best prediction accuracy.

[0070] According to the above aspect (7), a plurality of first prediction models corresponding to a plurality of process values ​​to be predicted are constructed using operating data as learning data. For those first prediction models whose prediction accuracy is below a reference value, additional operating data relearning is performed using the operating data to which additional data has been added as learning data, thereby constructing a second prediction model with improved prediction accuracy. If the prediction accuracy of the second prediction model is still below the reference value, similar operating data relearning is performed using similar operating data similar to an operating point of the plant among the operating data, thereby improving the prediction accuracy of a third prediction model. In the plant control device, the first prediction model, the second prediction model, or the third prediction model with the best prediction accuracy is selected as each prediction model for predicting the plurality of process values. This enables each process value to be predicted with high accuracy, thereby suitably improving the controllability of the plant control device.

[0071] (8) A plant control assistance program according to one aspect is a plant control assistance program for assisting plant control using a plurality of process values ​​of a plant that are respectively predicted using a plurality of prediction models, the program being executable by a computer device: a step of training a plurality of first prediction models as the plurality of prediction models by using training data that are operating data of the plant; a step of re-training a second prediction model using, as the training data, the operating data to which additional data has been added, for a first prediction model among the plurality of first prediction models, whose prediction accuracy is less than a reference value; a step of re-training a third prediction model using, as the training data, similar operating data that is similar to an operating point of the plant among the operating data, for the second prediction model, whose prediction accuracy is less than the reference value; and a step of selecting, as the prediction model, one of the first prediction model, the second prediction model, or the third prediction model, which has the highest prediction accuracy.

[0072] According to the above aspect (8), a plurality of first prediction models corresponding to a plurality of process values ​​to be predicted are constructed using operating data as learning data. For those first prediction models whose prediction accuracy is below a reference value, additional operating data relearning is performed using the operating data to which additional data has been added as learning data, thereby constructing a second prediction model with improved prediction accuracy. If the prediction accuracy of the second prediction model is still below the reference value, similar operating data relearning is performed using similar operating data similar to an operating point of the plant among the operating data, thereby improving the prediction accuracy of a third prediction model. In the plant control device, the first prediction model, the second prediction model, or the third prediction model with the best prediction accuracy is selected as each prediction model for predicting the plurality of process values. This enables each process value to be predicted with high accuracy, thereby suitably improving the controllability of the plant control device.

[0073] REFERENCE SIGNS LIST 1 Plant control system 10 Plant 100 Plant control support device 102 External input interface 104 Preprocessing unit 106 Operation data accumulation unit 108 Prediction model learning unit 110 Prediction model storage unit 112 Prediction accuracy determination unit 114 Operation condition search unit 116 Search range setting unit 118 External output interface 120 First prediction model learning unit 122 Second prediction model learning unit 124 Third prediction model learning unit 130 Display unit 132 Selection unit 200 Control device M Prediction model M1 First prediction model M2 Second prediction model M3 Third prediction model V Virtual space

Claims

1. A plant control assistance device for assisting plant control using a plurality of process values ​​of a plant that are each predicted using a plurality of prediction models, the plant control assistance device comprising: a first prediction model training unit for training a plurality of first prediction models as the plurality of prediction models using training data that are operating data of the plant; a second prediction model training unit for re-training a second prediction model using operating data to which additional data has been added as training data for a first prediction model of the plurality of first prediction models whose prediction accuracy is less than a reference value; a third prediction model training unit for re-training a third prediction model using similar operating data that is similar to an operating point of the plant from the operating data for the second prediction model whose prediction accuracy is less than the reference value as training data; and a prediction model selection unit for selecting the first prediction model, the second prediction model, or the third prediction model that has the best prediction accuracy as the prediction model.

2. The plant control assistance device of claim 1, wherein at least one of the first prediction model learning unit, the second prediction model learning unit, or the third prediction model learning unit generates a plurality of prediction model candidates for each of the plurality of prediction models using a plurality of types of machine learning algorithms, and selects the prediction model candidate with the best prediction accuracy as the first prediction model, the second prediction model, or the third prediction model.

3. A plant control support device according to claim 1 or 2, wherein the similar operating data is constructed by extracting a predetermined proportion or number of data from the operating data and the additional data that have a small deviation from the operating point.

4. A plant control assistance device according to claim 1 or 2, wherein the additional data is data collected from the plant since the previous learning of the first prediction model.

5. The plant control assistance device according to claim 1 or 2, further comprising: an index calculation unit for calculating an index relating to the performance of the plant based on the plurality of process values ​​predicted using the plurality of prediction models; an operating condition search unit for searching for operating conditions for the plant based on the index; and a control unit for controlling the plant based on the operating conditions.

6. The plant control support device according to claim 1 or 2, wherein the plant is a boiler system.

7. A plant control support method for supporting plant control using a plurality of process values ​​of a plant that are each predicted using a plurality of prediction models, the plant control support method comprising: a step of training a plurality of first prediction models as the plurality of prediction models using training data that is operating data of the plant; a step of re-training a second prediction model using the operating data to which additional data has been added as training data for a first prediction model of the plurality of first prediction models whose prediction accuracy is less than a reference value as the training data; a step of re-training a third prediction model using similar operating data that is similar to an operating point of the plant from the operating data for a second prediction model whose prediction accuracy is less than the reference value as the training data; and a step of selecting the first prediction model, the second prediction model, or the third prediction model that has the best prediction accuracy as the prediction model.

8. A plant control assistance program for assisting plant control using a plurality of process values ​​of a plant that are each predicted using a plurality of prediction models, the program being capable of executing the following steps on a computer device: training a plurality of first prediction models as the plurality of prediction models using training data that is operating data of the plant; for a first prediction model among the plurality of first prediction models whose prediction accuracy is less than a reference value, re-training a second prediction model using the operating data to which additional data has been added as the training data; for a second prediction model whose prediction accuracy is less than the reference value, re-training a third prediction model using similar operating data that is similar to an operating point of the plant among the operating data as the training data; and selecting the first prediction model, the second prediction model, or the third prediction model that has the best prediction accuracy as the prediction model.

Images