State variable prediction device, state variable prediction method, state variable prediction system, and control method for the state variable prediction system

Abstract

To predict a state quantity accurately while keeping a cost down using a prediction model obtained by adjusting a learned prediction model corresponding to another plant.SOLUTION: A state quantity prediction device predicts a state quantity of a plant corresponding to an input parameter using a prediction model including a machine learning model and a physical model. The state quantity prediction device stores a first prediction model corresponding to a first plant, creates a second prediction model corresponding to a second plant by adjusting the first prediction model using measurement data of the second plant, and predicts a state quantity of the second plant using the second prediction model. The creation of the second prediction model is performed by regularizing and learning a difference between machine learning parameters of the first prediction model and the second prediction model.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a state quantity prediction device, a state quantity prediction method, a state quantity prediction system, and a control method for a state quantity prediction system.

Background Art

[0002] In equipment such as a plant, for the purpose of monitoring, control, or abnormality determination, etc., the state quantity may be predicted by calculation using a numerical model. Examples of this type of numerical model include a physical model based on a static equilibrium equation, a state equation, etc. derived from physical knowledge, and a statistical model such as a machine learning model using a neural network, multiple regression analysis, etc.

[0003] When using a physical model as a numerical model, the physical model makes an approximation such as linearization assuming that the true prediction formula is linear, or models assuming operating conditions such as static conditions, etc. However, in such a physical model, there is a possibility that the prediction error may become large when the assumed prerequisite conditions are not satisfied. Also, in a statistical model such as a machine learning model, there is a possibility that the prediction error may become large in the extrapolation region of the learning data used when constructing the model. Therefore, in Patent Document 1, a prediction method has been proposed in which the deviation between the state quantity derived from the physical model and the measured quantity is learned by machine learning, and the error of the state quantity predicted by the physical model is corrected by the machine learning model. This method is, so to speak, a prediction method that combines a physical model and a machine learning model, but the dynamic characteristics of the state quantity that is the prediction target in the machine learning model are not considered. Therefore, there is a possibility that sufficient prediction accuracy cannot be obtained for an object whose state quantity has dynamic characteristics.

[0004] As one solution to the problems related to such dynamic characteristics, it is conceivable to introduce a Recurrent Neural Network (RNN) as a machine learning method. On the other hand, Neural Ordinary Differential Equation (Neural ODE), a continuous representation neural network with a differential structure, has the advantage of high memory efficiency compared to RNN and can handle a temporally continuous model. In Non-Patent Document 1, a prediction method capable of improving prediction accuracy is disclosed by combining Neural ODE with known differential equations.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Non-Patent Documents

[0006]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0007] When predicting a state quantity using a prediction model that combines a machine learning model and a physical model as in Patent Document 1 and Non-Patent Document 1 above, the prediction model learns using learning data obtained from a plant that is the prediction target, thereby determining the machine learning parameters included in the machine learning model and the physical parameters included in the physical model. The learning data used for learning such a prediction model utilizes measurement data obtained by a measurement device such as a sensor installed in the plant that is the prediction target.

[0008] In the case of a plant where sensors for obtaining measurement data usable as training data are not installed, and the goal is to train a predictive model to predict state variables, installing new sensors to acquire training data is costly. In this case, if a predictive model has been previously trained using training data from another similar plant, it is conceivable to use that predictive model for the plant without sensors. However, since each plant has its own unique characteristics, using a predictive model from another plant as is will reduce the prediction accuracy, requiring fine-tuning of the predictive model.

[0009] At least one embodiment of this disclosure has been made in view of the above circumstances, and aims to provide a state quantity prediction device, a state quantity prediction method, a state quantity prediction system, and a control method for a state quantity prediction system that can predict state quantities accurately while keeping costs down, using a prediction model obtained by adjusting a trained prediction model corresponding to other plants. [Means for solving the problem]

[0010] Some embodiment of the state quantity prediction device described herein solves the above problem, A state quantity prediction device for predicting plant state quantities corresponding to input parameters using a prediction model that includes a machine learning model and a physical model, A first prediction model storage unit for storing a first prediction model, which is a prediction model in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, corresponding to the first plant, A measurement data acquisition unit for acquiring measurement data from the second plant, A second prediction model creation unit for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A state quantity prediction unit for predicting the state quantities of the second plant using the second prediction model, Equipped with, The second prediction model creation unit creates the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and training the model.

[0011] Some embodiment of the state quantity prediction system described herein addresses the above-mentioned problems. A state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A first prediction model storage unit for storing a first prediction model, which is a prediction model in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, corresponding to the first plant, A measurement data acquisition unit, which acquires measurement data from the second plant after a request from the aforementioned information processing device, A second prediction model creation unit for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A state quantity prediction unit for predicting the state quantities of the second plant using the second prediction model, Equipped with, The second prediction model creation unit creates the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and training the model.

[0012] Some state quantity prediction methods according to the embodiments of this disclosure solve the above problems. A method for predicting plant state variables corresponding to input parameters, using a prediction model that includes a machine learning model and a physical model, A step of storing a first prediction model, which corresponds to the first plant and in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, Steps include acquiring measurement data from the second plant, The steps include creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A step of predicting the state variables of the second plant using the second prediction model, Equipped with, In the step of creating the second prediction model, the second prediction model is created by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and training the model.

[0013] A control method for a state quantity prediction system according to some embodiments of this disclosure solves the above problem. A control method for a state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, After a request is received from the aforementioned information processing device, a measurement data acquisition step is performed to acquire measurement data from the second plant, A second prediction model creation step for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model stored in a first prediction model storage unit for storing the first prediction model, which is the prediction model corresponding to the first plant, in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, using the aforementioned measurement data. A state quantity prediction step for predicting the state quantities of the second plant using the second prediction model, Execute, The second prediction model creation step involves creating the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and then training the model.

[0014] Some embodiment of the state quantity prediction device described herein solves the above problem, A state quantity prediction device for predicting plant state quantities corresponding to input parameters using a prediction model that includes a machine learning model and a physical model, A predictive model creation unit for creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The prediction model creation unit creates the prediction model by applying regularization to the machine learning parameters of the prediction model and training it.

[0015] Some embodiment of the state quantity prediction system described herein addresses the above-mentioned problems. A state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A predictive model creation unit for creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The prediction model creation unit creates the prediction model by applying regularization to the machine learning parameters of the prediction model and training it.

[0016] Some state quantity prediction methods according to the embodiments of this disclosure solve the above problems. A method for predicting plant state variables corresponding to input parameters, using a prediction model that includes a machine learning model and a physical model, The steps include creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, The steps include: predicting the state variables of the plant using the prediction model; Equipped with, In the step of creating the prediction model, the prediction model is created by applying regularization to the machine learning parameters of the prediction model and training it.

[0017] A control method for a state quantity prediction system according to some embodiments of this disclosure solves the above problem. A control method for a state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A predictive model creation unit creates the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using the learning data acquired at the plant, after receiving a request from the information processing device. A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The prediction model creation unit creates the prediction model by applying regularization to the machine learning parameters of the prediction model and training it. [Effects of the Invention]

[0018] According to at least one embodiment of this disclosure, a state quantity prediction device, a state quantity prediction method, a state quantity prediction system, and a control method for a state quantity prediction system can be provided that predict state quantities accurately while keeping costs down, using a prediction model obtained by adjusting a trained prediction model corresponding to another plant. [Brief explanation of the drawing]

[0019] [Figure 1] This is a block diagram showing the configuration of a state quantity prediction device according to one embodiment. [Figure 2] Figure 1 is a schematic diagram illustrating an example of a neural network, which is a machine learning model. [Figure 3] This is a block diagram showing the configuration of a state quantity prediction device according to another embodiment. [Figure 4] This figure shows the trend of predicted values ​​calculated based on a second prediction model adjusted using measured values ​​of state variables. [Figure 5] This diagram shows a schematic configuration of a desulfurization plant. [Figure 6] This block diagram shows the configuration of a state quantity prediction system according to one embodiment. [Modes for carrying out the invention]

[0020] Hereinafter, several embodiments of the present invention will be described with reference to the attached drawings. However, the configurations described as embodiments or shown in the drawings are not intended to limit the scope of the present invention, but are merely illustrative examples.

[0021] The state quantity prediction device 1 according to at least one embodiment of the present disclosure is a device for predicting the state quantity x of a plant corresponding to an input parameter u using a prediction model M. The hardware configuration for realizing the state quantity prediction device 1 is not limited. For example, it is configured as an information processing device including a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), and a computer-readable storage medium, etc. And a series of processes for realizing various functions are stored in a storage medium, etc. in the form of a program as an example. The CPU reads this program into the RAM, etc. and executes information processing and arithmetic processing, whereby various functions are realized. Incidentally, the program may be applied in a form pre-installed in a ROM or other storage medium, a form provided in a state stored in a computer-readable storage medium, a form distributed via wired or wireless communication means, etc. A computer-readable storage medium is a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, etc.

[0022] The prediction model M used by the state quantity prediction device 1 is a machine learning model M N and a physical model M P including. In the following embodiments, a physical model M P for predicting the state quantity x of a plant in a static state is combined with a machine learning model M N for predicting the dynamic state of the plant, whereby the prediction model M is configured. The functions of the machine learning model M N and the physical model M P included in such a prediction model M are only examples and are not limited.

[0023] FIG. 1 is a block diagram showing the configuration of a state quantity prediction device 1A according to an embodiment. The state quantity prediction device 1A has a prediction model M in which a machine learning model M N and a physical model M P are combined. Specifically, the machine learning model M NA first prediction value calculation unit 4 for calculating the first prediction value X1 using the physical model M P The system includes a second prediction value calculation unit 6 for calculating a second prediction value X2 using the first prediction value X1 and the second prediction value X2, and a state variable prediction value calculation unit 8 for calculating a predicted value X of the state variable x based on the first prediction value X1 and the second prediction value X2.

[0024] Physical model M P This is a physical model corresponding to a plant in a static state, and is configured to output a second predicted value X2, which is the static component of the plant's state variable x corresponding to the input parameter u in the static state. Such a physical model M P The relationship between the static input parameter u and the second predicted value X2 is generally expressed by the following equation. X2 = f(u,θ) P ) (1) Furthermore, θ P is the physical model M P At least one physical parameter included in, and in this embodiment, k physical parameters θ P1 θ P2 , , , θ Pk (Hereafter, "physical parameter θ" as appropriate) P It is called a 'f'. k is a natural number greater than or equal to 1. f is the input parameter u and the physical parameter θ. P It is any function whose variable is [variable name].

[0025] Machine learning model M N This is a machine learning model that corresponds to a plant in a dynamic state, and is configured to output a first predicted value X1, which is the dynamic component of the plant's state variable x corresponding to the input parameter u when it is in a dynamic state. Such a machine learning model M N This is configured as a neural network that shows the relationship between, for example, the input parameter u during dynamic operation and the first predicted value X1, which is the dynamic component of the plant's state variable x.

[0026] Figure 2 shows the machine learning model M from Figure 1. NThis is a schematic diagram showing an example of a neural network. This neural network comprises an input layer 12 having multiple nodes to which multiple input parameters u1, u2, ... are input, an output layer 14 having a node to output a first predicted value X1 as a prediction result, and an intermediate layer 13 (hidden layer) containing multiple nodes between the input layer 12 and the output layer 14. Each node in the intermediate layer 13 has a weight coefficient, which is the machine learning parameter θ. N1 θ N2 , , , θ Nn (Hereafter, "machine learning parameter θ" as appropriate) N It is called "[...]. A value of n (a natural number greater than or equal to 1) is set.

[0027] The first prediction value calculation unit 4 uses the machine learning model M N The first predicted value X1 is calculated by providing an input parameter u to the physical model M. P The second predicted value X2 is calculated by providing an input parameter u to the state variable x. The state variable prediction value calculation unit 8 calculates a predicted value X of the state variable x based on the first predicted value X1 calculated by the first predicted value calculation unit 4, which corresponds to the dynamic component of the state variable x, and the second predicted value X2 calculated by the second predicted value calculation unit 6, which corresponds to the static component of the state variable x. In this way, the state variable prediction device 1A can obtain a predicted value X of the state variable x, which includes both dynamic and static components.

[0028] Here, the machine learning parameters θ included in the predictive model M are shown. N , and the physical parameter θ P This is determined by learning using training data (hereinafter referred to as the machine learning parameter θ). N , and the physical parameter θ P When referring to them collectively, they will be appropriately called "learning parameters θ". Specifically, during learning, the machine learning parameters θ N , and the physical parameter θ P For each, the initial value θ N0 θ P0 Given a machine learning parameter θ, for example, the evaluation function F1 defined by the following equation is minimized. N , and the physical parameter θP It is searched for. F1 = Σ(XX´) 2 (1) Note that X' is the true value (e.g., the actual value) corresponding to the predicted value X' included in the training data.

[0029] In the state quantity prediction device 1A described above, a prediction model M is used to predict the state quantity x. However, obtaining a prediction model M with good prediction accuracy requires that a sufficient amount of training data used to train the prediction model M be available. Therefore, in plants where sufficient training data cannot be obtained, for example, the prediction accuracy of the prediction model M obtained through training may decrease. To solve this, it is conceivable to add sensors or other devices to the plant to obtain sufficient training data, but this is disadvantageous in terms of cost. To solve this problem, in the following embodiment, by adjusting the trained first prediction model M1 corresponding to the first plant P1 where sufficient training data can be obtained, a second prediction model M2 corresponding to the second plant P2, which is different from the first plant P1, is obtained, thereby enabling the prediction of the state quantity x in the second plant P2.

[0030] Figure 3 is a block diagram showing the configuration of a state quantity prediction device 1B according to another embodiment. The state quantity prediction device 1B is a device for predicting the state quantity x of the second plant P2, and comprises a first prediction model storage unit 15, a measurement data acquisition unit 16, a second prediction model creation unit 17, and a state quantity prediction unit 18.

[0031] The first prediction model storage unit 15 is configured to store the first prediction model M1 corresponding to the first plant P1, which is different from the second plant P2 that is the target of prediction by the state quantity prediction device 1B. The first prediction model M1 is the aforementioned prediction model M corresponding to the first plant P1, and as illustrated in Figure 1, it is a machine learning model M N and physical model M PThis is a predictive model that includes [the specified data]. At the first plant P1, it is possible to acquire sufficient operating data to be used as training data for training the first predictive model M1, and the first predictive model M1 is constructed to have sufficient predictive accuracy by training using such sufficient operating data as training data. The first predictive model M1 constructed in this way is stored in the first predictive model storage unit 15 in a readable format. Furthermore, the first plant P1 and the second plant P2 are similar to each other. Whether or not the first plant P1 and the second plant P2 are similar can be determined by comprehensively considering factors such as the similarity of the plant's equipment configuration and the specifications of each piece of equipment, and the similarity of state variables acquired other than training data (such as the configuration of sensors installed in the plant).

[0032] The measurement data acquisition unit 16 is configured to acquire operational data as measurement data Dm from the second plant P2, which is the target of prediction by the state quantity prediction device 1B. The second plant P2 is equipped with measuring instruments for measuring operational data, but these measuring instruments have limitations, such as differences in the type of measuring instruments and control methods installed in each plant, which prevent the acquisition of a sufficient number of data for the measured value X' of the state quantity x (however, for operational data other than the measured value X', a sufficient number of data can be acquired, similar to the first plant P).

[0033] The second prediction model creation unit 17 is configured to create a second prediction model M2 corresponding to the second plant P2 by adjusting the first prediction model M1 read from the first prediction model storage unit 15 using the measurement data Dm of the second plant P2 acquired by the measurement data acquisition unit 16. The creation of the second prediction model M2 involves adjusting the learning parameters θ (the aforementioned machine learning parameters θ) included in the first prediction model M1 corresponding to the first plant P1 which is similar to the second plant P2 that is the target of prediction. N , and the physical parameter θ PThis is done by fine-tuning the predictive model M. In each plant, the relationship between the input parameter u and the predicted value X depends on characteristics such as individual differences in the plant, so generally the predictive model M is different for each plant. However, for the first plant P1 and the second plant P2, which are similar to each other, the predictive models M will be relatively close. Therefore, the second predictive model creation unit 17 fine-tunes the first predictive model M1, which corresponds to the first plant P1 and has already been trained using sufficient training data, to create a second predictive model M2 that corresponds to the second plant P2, for which it is difficult to obtain sufficient training data for the measured value X' of the state variable x.

[0034] The creation of the second prediction model M2 by the second prediction model creation unit 17 involves the machine learning parameters θ of the first prediction model M1. N1 , and the machine learning parameter θ of the second prediction model M2 N2 This is done by applying regularization to the magnitude of the difference and learning (transfer learning). Specifically, this learning is performed to minimize the evaluation function F2 given by the following equation. F2 = Σ(XX´) 2 +βΣ(θ N2 -θ N1 ) 2 (2)

[0035] This evaluation function F2 comprises a first term similar to the evaluation function F1 mentioned above (see equation (1) above) and a second term which is a regular term. The second term is the machine learning parameter θ of the first prediction model M1 and the second prediction model M2. N It is expressed as the 2-norm of the difference and added to the first term using the weight coefficient β. In learning that minimizes such an evaluation function F2, the machine learning model M N and physical model M P Among the predictive model M that includes the machine learning parameters θ N By restricting the learning degrees of freedom of the physical parameter θ, even when there is little training data for the measured value X' of the state variable x, the physical parameter θ PThis facilitates updates and avoids overfitting. As a result, even when only a small amount of training data (measured values ​​X' of state variables x included in measurement data Dm) is obtained from the second plant P2, the second prediction model creation unit 17 can create a reliable second prediction model M2 by fine-tuning the first prediction model M1 using this small amount of training data.

[0036] Furthermore, the second prediction model creation unit 17, in learning using the evaluation function F2, uses the physical parameter θ P A tolerance range may be set for this. This tolerance range is determined by the physical parameter θ. P The range of possible values ​​can be predetermined. This allows for the machine learning parameter θ obtained through the regularization process described above. N In addition to limiting the learning degrees of freedom of the physical parameter θ, P By restricting the learning degrees of freedom, a more reliable second learning model M2 can be created.

[0037] The state variable prediction unit 18 is configured to calculate the predicted value X of the state variable x for the second plant P2 using the second prediction model M2 created by the second prediction model creation unit 17. Specifically, as described above with reference to Figure 1, the state variable prediction unit 18 obtains the predicted value X of the corresponding state variable x by inputting the input parameter u, which is the operating data acquired at the second plant P2, to the second prediction model M2.

[0038] Furthermore, each block constituting the state quantity prediction device 1B shown in Figure 3 may be configured as a single device or as part of a multi-device configuration. For example, the configuration for creating the second prediction model M2 (first prediction model storage unit 15, measurement data acquisition unit 16, second prediction model creation unit 17, etc.) and the configuration for predicting the state quantity x using the second prediction model M2 (state quantity prediction unit 18, etc.) may be configured as different devices. In this case, the former may create the second prediction model M2, and the latter may use the second prediction model M2 created in the former to predict the state quantity x.

[0039] The second prediction model creation unit 17 may also adjust the second prediction model M2 so that the predicted value X of the state quantity x predicted by the state quantity prediction unit 18 becomes, for example, the measured value X' of the state quantity x measured by equipment installed in the second plant P2. Figure 4 shows the change in the predicted value X calculated based on the second prediction model M2 adjusted by the measured value X' of the state quantity x.

[0040] In Figure 4, at times t1, t2, ... defined for each predetermined or undefined sampling period Ts, the second prediction model M2 is adjusted so that the predicted value X of the state quantity x predicted by the state quantity prediction unit 18 becomes the measured value X' of the state quantity x. The second prediction model M2 is basically created using the evaluation function F2 described above, but for example, if the operating conditions of the second plant P2 are changed midway through (for example, in the embodiment described later with reference to Figure 5, the type of fuel (coal type) used in the boiler attached to the desulfurization unit 20, etc.), the prediction accuracy may decrease. In this embodiment, by adjusting the second prediction model M2 so that the predicted value X of the state quantity x becomes the measured value X' for each predetermined sampling period Ts, the reliability of the second prediction model M2 can be suitably maintained even when the operating conditions of the second plant P2 are changed.

[0041] Furthermore, when training the predictive model M using the aforementioned evaluation function F1 (see equation (1) above), if the amount of data included in the training data is small, it may cause overfitting and reduce the prediction accuracy. Therefore, overfitting can be avoided by performing regularization during the training of the predictive model M. Specifically, instead of evaluation function F1, the evaluation function F3 of the following equation may be used as the evaluation function used when training the predictive model M. F3 = Σ(XX´) 2 +αΣθ i 2 (3)

[0042] In this evaluation function F3, the first term represents the sum of the squared prediction errors, similar to evaluation function F1, and the second term represents the regular term. Equation (3) above is merely an example; for example, the first term could be the average of the first term in equation (3) divided by the number of evaluation data points, and the second term could be the average of the second term in equation (3) divided by the number of parameters. By training using such an evaluation function F3, overfitting can be avoided, and even when the amount of data in the training data is small, an accurate prediction model can be obtained.

[0043] In this learning process using the evaluation function F3, among the learning parameters θ of the predictive model M, the machine learning parameter θ N The degrees of freedom are physical parameters θ P It becomes relatively higher in relation to θ. Therefore, in learning using the evaluation function F3, the machine learning parameter θ N Learning becomes dominant, and the physical parameter θ P There is a risk that learning may not be performed properly, or that learning may not be reproducible.

[0044] To solve these problems, the machine learning parameter θ is one of the learning parameters θ, as shown in the following equation. N The evaluation function F4 may be used to train the system so that regularization is applied only to the specified values. F4 = Σ(XX´) 2 +αΣ(θ Ni ) 2 (4) By training in such a way that the evaluation function F4 is minimized, the machine learning parameter θ N The learning degrees of freedom are restricted, and the physical parameter θ P By facilitating updates, overfitting can be avoided, and as a result, even when the amount of training data (actual measured values ​​X' of state variable x included in measurement data Dm) is small, the machine learning parameter θ N and physical parameter θ P These can be appropriately determined through learning.

[0045] In equation (4), the second term of the evaluation function F4 is the machine learning parameter θ. NThe two norms are expressed by adding them together with a weight coefficient α, but this is just one example, as with the evaluation function F3 mentioned above. For example, the first term could be the average of the first term of equation (4) divided by the number of evaluation data points, and the second term could be the average of the second term of equation (4) divided by the number of parameters.

[0046] Furthermore, in the learning process using the aforementioned evaluation functions F3 and F4, the physical parameters θ of the predictive model M are also considered. P A tolerance range may be set for this. By limiting the learning degrees of freedom of the physical parameters, a learning model M that can predict state variables with greater accuracy can be obtained.

[0047] Furthermore, the prediction model M created by learning using the aforementioned evaluation functions F3 and F4 may also be adjusted, as described above with reference to Figure 4, so that the predicted value X of the state variable x predicted using the prediction model M becomes the actual measured value X' of the state variable x measured by, for example, equipment installed in the plant. By adjusting the predicted value X of the state variable x predicted using the prediction model M to become the actual measured value X' at predetermined timings, the prediction accuracy of the prediction model M can be suitably maintained even when the operating conditions of the plant (for example, in the embodiment described later with reference to Figure 5, the type of fuel (coal type) used in the boiler attached to the desulfurization unit 20) changes.

[0048] Next, specific application examples of the state quantity prediction devices 1A and 1B having the above configuration will be described. Here, as a specific example of predicting the state quantity x of a plant, we will describe the case in which the absorbent concentration of the absorbent liquid used in the absorption tower of a flue gas desulfurization plant is predicted as the state quantity x. However, other examples may include gas turbines, steam turbines, large refrigerators, air conditioners, etc.

[0049] Figure 5 is a schematic diagram showing the configuration of the desulfurization apparatus 20. The desulfurization apparatus 20 is installed in conjunction with a boiler (not shown) of a plant facility such as a thermal power plant and comprises a dust collector 22 that collects particulate matter contained in the exhaust gas G0 flowing through the exhaust passage 23a of the boiler, and an absorption tower 24 installed downstream of the dust collector 22 in the exhaust passage 23b through which the exhaust gas G1 that has passed through the dust collector 22 flows.

[0050] The dust collector 22 is an electrostatic precipitator that collects dust by performing corona discharge on the exhaust gas G0 supplied into the casing, thereby charging the fine particles contained in the exhaust gas G0, and causing them to adhere to the positively and negatively charged attachment points by electrical attraction. The exhaust gas G1 that has been treated by the dust collector 22 is supplied to the absorption tower 24 via the exhaust passage 23b.

[0051] The absorption tower 24 performs desulfurization by absorbing SO2 (sulfur dioxide) from the exhaust gas G1, which has been treated by dust collection in the dust collector 22, by bringing it into contact with an absorbent liquid 26 containing limestone 30. The absorbent liquid 26 is stored at the bottom of the absorption tower 24. The absorbent liquid 26 is produced when limestone 30 supplied from a limestone feeder 28 located outside the absorption tower 24 is mixed with water 32 supplied to the bottom of the absorption tower 24. In this embodiment, the case in which the absorbent liquid 26 is generated by supplying limestone 30 to water 32 from the limestone feeder 28 is illustrated. However, instead, the absorbent liquid 26 may be generated by supplying limestone slurry containing limestone to water 32.

[0052] The absorbent liquid 26 stored at the bottom of the absorption tower 24 is pumped by the absorbent liquid circulation pump 34 and supplied to the upper part of the absorption tower 24 via absorbent liquid circulation piping 36 located outside the absorption tower 24. The absorbent liquid circulation pump 34 consists of multiple pump units connected in parallel, and the operating state of each pump unit is controlled. For example, if the pump units are of a variable capacity type (moving vane type), the flow rate of the absorbent liquid 26 pumped from the absorbent liquid circulation pump 34 can be controlled by adjusting the capacity of each pump unit. If the pump units are of a fixed capacity type (fixed vane type), the flow rate of the absorbent liquid 26 pumped from the absorbent liquid circulation pump 34 can be controlled by adjusting the number of operating pump units. The absorbent liquid 26 supplied to the upper part of the absorption tower 24 comes into contact with the exhaust gas G1 rising inside the absorption tower 24 as it is sprayed from nozzles 38 located at the top of the absorption tower 24 and falls. As a result, the SO2 contained in the exhaust gas G1 reacts with the limestone 30 in the absorbent liquid 26, and desulfurization treatment is performed. Furthermore, the method of spraying and dropping the absorbent liquid 26 from the nozzle 38 may be a grid type, a liquid column type, or a spray type.

[0053] Equation (5) below is the chemical reaction equation for the desulfurization treatment carried out in the absorption tower 24. In the desulfurization reaction, limestone 30 reacts with SO2 contained in the exhaust gas G1 to produce gypsum 34 (CaCO4·2H2O) as a byproduct. The exhaust gas G2 from which SO2 has been removed is discharged to the outside from the top of the absorption tower 24 through the desulfurization exhaust gas pipe 25. SO2+1 / 2O2+CaCO3+2H2O→CaCO4·2H2O+CO2(5)

[0054] Furthermore, a portion of the absorbent liquid 26 stored at the bottom of the absorption tower 24 is pumped by the absorbent liquid circulation pump 34 and sent to the dewaterer 42 via an extraction pipe 40 branched from the absorbent liquid circulation piping 36 outside the absorption tower 24. The dewaterer 42 is composed of, for example, a belt filter, which dewaters the absorbent liquid 26 as it is transported by the belt filter, and the resulting gypsum 34 is discharged outside the system. Furthermore, the filtrate produced during the dewatering process in the dewaterer 42 is reused by being supplied as water 32 to the bottom of the absorption tower 24.

[0055] In addition, oxidizing air 46 is supplied to the bottom of the absorption tower 24. As a result, the absorption liquid 26 contains the oxidizing air 46, which promotes the oxidation of sulfite groups to sulfate groups that are transferred from the SO2 exhaust gas into the absorption liquid 26, and consequently improves the efficiency of SO2 removal from the exhaust gas. Furthermore, if the spraying and dropping method of the absorbent liquid 26 from the nozzle 38 is grid type, the supply of oxidizing air 46 may be omitted because the absorbent liquid 26 is oxidized during the dropping process.

[0056] Such a desulfurization apparatus 20 is equipped with at least one sensor that can be selected as the aforementioned input parameter u. In this embodiment, there is an SO2 concentration sensor 50 for detecting the SO2 concentration u1 (desulfurization outlet SO2 concentration [ppm]) at the outlet side of the absorption tower 24, an SO2 concentration sensor 52 for detecting the SO2 concentration u2 (desulfurization inlet SO2 concentration [ppm]) at the inlet side of the absorption tower 24, and the flow rate u3 (absorption tower limestone slurry flow rate [m³]) of the limestone slurry produced in the absorption tower 24. 3 A limestone slurry flow sensor 54 for detecting the limestone slurry flow rate u5[%] of the absorption tower 24, a limestone slurry concentration sensor 58 for detecting the absorption tower limestone slurry concentration [wt%] of the absorption tower limestone slurry, and the flow rate of the oxidizing air supplied to the absorption tower 24 u7[m 3 The desulfurization apparatus 20 is equipped with an oxidation air flow sensor 60 for detecting [N / h], a pH sensor 62 for detecting the pHu8 (absorption tower pH) of the absorption liquid 26 in the absorption tower 24, and a level sensor 64 for detecting the level u9 (absorption tower level [m]) of the absorption liquid 26 in the absorption tower 24. The detected values ​​from each of these sensors are input to a control device 200, which is the control unit of the desulfurization apparatus 20, and the control device 200 controls each part of the desulfurization apparatus 20. Furthermore, the power generation command signal u4 for a generator (not shown) that generates electricity using steam produced by a boiler (not shown) is also configured to be obtainable as an input parameter u.

[0057] At least one of the detected values ​​from each of these sensors is input as an input parameter u to the aforementioned state quantity prediction devices 1A and 1B, and the state quantity prediction devices 1A and 1B predict the absorbent (calcium carbonate) concentration [mmol / L] in the absorption tower 24 as a state quantity x.

[0058] Furthermore, in the state quantity prediction devices 1A and 1B, the physical parameter θ of the physical model M is used. P Various parameters related to the desulfurization apparatus 20 can be used, for example, including at least one of the activity of limestone in the absorbent liquid 26, the moisture content in the inlet gas of the absorption tower 24, and the humidity increase rate in the absorption tower 24.

[0059] The state quantity prediction devices 1A and 1B, having this configuration, are connected to each sensor located in the desulfurization apparatus 20 via a network, allowing them to acquire the results detected by each sensor as input parameters u. As a result, the state quantity prediction device 1 can predict the absorbent (calcium carbonate) concentration [mmol / L] as a state quantity x corresponding to the input parameter u, and this can be used in the operation of the desulfurization apparatus 20.

[0060] Furthermore, it is possible to configure the system as a state quantity prediction system executed by an information processing device 70 connected to communicate with the state quantity prediction device 1A or the state quantity prediction device 1B in the above embodiment. Here, Figure 6 is a block diagram showing the configuration of a state quantity prediction system according to one embodiment.

[0061] The information processing device 70 includes a display unit 71 that displays the predicted value X calculated by the state quantity prediction device 1A or the state quantity prediction device 1B. Furthermore, the information processing device 70 may be configured to execute various processes in the state quantity prediction device 1A or the state quantity prediction device 1B in accordance with instructions input via the display unit 71, upon request from the information processing device 70.

[0062] Furthermore, it is possible to replace the components in the above-described embodiments with well-known components as appropriate, without departing from the spirit of this disclosure, and the above-described embodiments may also be combined as appropriate.

[0063] The contents described in each of the above embodiments can be understood, for example, as follows:

[0064] (1) A state quantity prediction device according to one embodiment is: A state quantity prediction device for predicting plant state quantities corresponding to input parameters using a prediction model that includes a machine learning model and a physical model, A first prediction model storage unit for storing a first prediction model, which is a prediction model in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, corresponding to the first plant, A measurement data acquisition unit for acquiring measurement data from the second plant, A second prediction model creation unit for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A state quantity prediction unit for predicting the state quantities of the second plant using the second prediction model, Equipped with, The second prediction model creation unit creates the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and training the model.

[0065] According to the embodiment described in (1) above, a second prediction model corresponding to the second plant can be created by adjusting the first prediction model corresponding to the first plant. Such adjustment of the prediction model is performed by applying regularization to the difference in machine learning parameters between the first and second prediction models and then learning. For example, even if only a small amount of second measurement data can be obtained at the second plant through simple measurement, a reliable second prediction model can be created based on the first prediction model by limiting the learning degrees of freedom of the machine learning parameters and promoting the updating of physical parameters. By using the second prediction model created in this way, it becomes possible to predict state variables with good accuracy even at the second plant where it is difficult to obtain sufficient measurement data.

[0066] (2) In other embodiments, in the embodiment of (1) above, The second prediction model creation unit creates the second prediction model by performing transfer learning such that the loss function, which includes the difference in machine learning parameters between the first prediction model and the second prediction model, is minimized.

[0067] According to the embodiment of (2) above, by performing transfer learning in such a way that the loss function, which includes the magnitude of the difference in machine learning parameters between the first prediction model and the second learning model, is minimized, it becomes possible to perform regularized learning so that the learning degrees of freedom of the machine learning parameters do not become excessively large.

[0068] (3) In other embodiments, in the embodiment of (1) or (2) above, The second prediction model creation unit sets tolerance ranges for the physical parameters of the first and second prediction models.

[0069] According to the embodiment described in (3) above, by setting an acceptable range for the physical parameters during learning, a second learning model capable of predicting state variables with greater accuracy can be obtained by limiting the learning degrees of freedom of the physical parameters.

[0070] (4) In other embodiments, in any one embodiment of (1) to (3) above, The second prediction model creation unit adjusts the second prediction model so that the predicted value of the state quantity predicted by the state quantity prediction unit becomes the measured value of the state quantity.

[0071] According to the embodiment described in (4) above, by adjusting the second prediction model so that the predicted values ​​of the state variables predicted by the second prediction model become the measured values, the prediction accuracy of the second prediction model can be suitably maintained even when the operating conditions of the second plant change.

[0072] (5) In other embodiments, in any one embodiment of (1) to (4) above, The aforementioned plant is a flue gas desulfurization plant that desulfurizes the flue gas by bringing an absorbent liquid into contact with the flue gas in an absorption tower. The aforementioned state quantity is the concentration of the absorbent in the absorbent solution in the absorption tower.

[0073] According to the embodiment of (5) above, the concentration of the absorbent in the absorption liquid in the absorption tower of the flue gas desulfurization plant can be suitably predicted as a state variable.

[0074] (6) In other embodiments, in the embodiment of (5) above, The input parameters include at least one of the following: the SO2 concentration at the desulfurization outlet of the absorption tower, the SO2 concentration at the desulfurization inlet of the absorption tower, the flow rate or concentration of the limestone slurry produced in the absorption tower, a power generation command signal to a generator that generates electricity using steam produced in a boiler that discharges the flue gas, the air flow rate in the boiler that discharges the flue gas, the flow rate of oxidizing air supplied to the absorption tower, the pH of the absorbent in the absorption tower, or the level of the absorbent in the absorption tower.

[0075] According to the embodiment of (6) above, by including at least one of these parameters as an input parameter, the absorbent concentration of the absorbent liquid in the absorption tower of the flue gas desulfurization plant can be suitably predicted as a state variable.

[0076] (7) A state quantity prediction system according to one embodiment is: A state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A first prediction model storage unit for storing a first prediction model, which is a prediction model in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, corresponding to the first plant, A measurement data acquisition unit, which acquires measurement data from the second plant after a request from the aforementioned information processing device, A second prediction model creation unit for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A state quantity prediction unit for predicting the state quantities of the second plant using the second prediction model, Equipped with, The second prediction model creation unit creates the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and training the model.

[0077] According to the embodiment of (7) above, a state quantity prediction system is realized that includes an information processing device capable of communicating with the state quantity prediction device according to each of the embodiments described above. In this system, the state quantity prediction device predicts the state quantities of the second plant as described above in response to a request from the information processing device, and the prediction result can be obtained by the information processing device.

[0078] (8) A state quantity prediction method relating to one embodiment is: A method for predicting plant state variables corresponding to input parameters, using a prediction model that includes a machine learning model and a physical model, A step of storing a first prediction model, which corresponds to the first plant and in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, Steps include acquiring measurement data from the second plant, The steps include creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A step of predicting the state variables of the second plant using the second prediction model, Equipped with, In the step of creating the second prediction model, the second prediction model is created by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and training the model.

[0079] According to the embodiment described in (8) above, a second prediction model corresponding to the second plant can be created by adjusting the first prediction model corresponding to the first plant. Such adjustment of the prediction model is performed by applying regularization to the difference in machine learning parameters between the first and second prediction models and then learning. For example, even if only a small amount of second measurement data can be obtained at the second plant through simple measurement, a reliable second prediction model can be created based on the first prediction model by limiting the learning degrees of freedom of the machine learning parameters and promoting the updating of physical parameters. By using the second prediction model created in this way, it becomes possible to predict state variables with good accuracy even at the second plant where it is difficult to obtain sufficient measurement data.

[0080] (9) A control method for a state quantity prediction system according to one embodiment is: A control method for a state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, After a request is received from the aforementioned information processing device, a measurement data acquisition step is performed to acquire measurement data from the second plant, A second prediction model creation step for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model stored in a first prediction model storage unit for storing the first prediction model, which is the prediction model corresponding to the first plant, in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, using the aforementioned measurement data. A state quantity prediction step for predicting the state quantities of the second plant using the second prediction model, Execute, The second prediction model creation step involves creating the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and then training the model.

[0081] According to the embodiment of (9) above, a state quantity prediction system is controlled, which includes an information processing device that can communicate with the state quantity prediction device according to each of the embodiments described above. The state quantity prediction device predicts the state quantities of the second plant as described above by performing a state quantity prediction method in response to a request from the information processing device, and the prediction result can be obtained by the information processing device.

[0082] (10) A state quantity prediction device according to one embodiment is A state quantity prediction device for predicting plant state quantities corresponding to input parameters using a prediction model that includes a machine learning model and a physical model, A predictive model creation unit for creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The prediction model creation unit creates the prediction model by applying regularization to the machine learning parameters of the prediction model and training it.

[0083] According to the embodiment of (10) above, when training a predictive model using training data acquired at the plant, the predictive model is created by applying regularization to the machine learning parameters of the predictive model during training. Generally, when training a predictive model using a small amount of training data, the training of machine learning parameters may become dominant, and the training of physical parameters may not be performed properly. In contrast, in this embodiment, even when the amount of training data is small, training is performed with regularization applied to prevent the machine learning parameters from becoming excessively large. This limits the degrees of freedom of the machine learning parameters, promotes the updating of physical parameters, avoids overfitting, and allows for the acquisition of a highly accurate predictive model.

[0084] (11) In other embodiments, in the embodiment of (10) above, The prediction model creation unit learns the prediction model such that the loss function, which includes the magnitude of the machine learning parameters, is minimized.

[0085] According to the embodiment of (11) above, by learning in such a way that the loss function including the magnitude of the machine learning parameters of the prediction model is minimized, it is possible to perform regularized learning so that the machine learning parameters of the prediction model do not become excessively large.

[0086] (12) In other embodiments, in the embodiment of (10) or (11) above, The predictive model creation unit sets an acceptable range for the physical parameters during training.

[0087] According to the embodiment described in (12) above, by setting an acceptable range for the physical parameters during training, the degrees of freedom for learning the physical parameters can be limited, thereby enabling the creation of a learning model that can predict state variables with greater accuracy.

[0088] (13) In other embodiments, in any one of the embodiments described in (10) to (12) above, The prediction model creation unit adjusts the prediction model so that the predicted value of the state quantity predicted by the state quantity prediction unit becomes the measured value of the state quantity.

[0089] According to the embodiment of (13) above, by adjusting the prediction model so that the predicted values ​​of the state variables predicted by the prediction model become the measured values, the prediction accuracy of the prediction model can be suitably maintained even when the operating conditions of the plant change.

[0090] (14) In other embodiments, in any one embodiment of (10) to (13) above, The aforementioned plant is a flue gas desulfurization plant that desulfurizes the flue gas by bringing an absorbent liquid into contact with the flue gas in an absorption tower. The aforementioned state quantity is the concentration of the absorbent in the absorbent solution in the absorption tower.

[0091] According to the embodiment of (14) above, the concentration of the absorbent in the absorption liquid in the absorption tower of the flue gas desulfurization plant can be suitably predicted as a state variable.

[0092] (15) In other embodiments, in the embodiment of (14) above, The input parameters include at least one of the following: the SO2 concentration at the desulfurization outlet of the absorption tower, the SO2 concentration at the desulfurization inlet of the absorption tower, the flow rate or concentration of the limestone slurry produced in the absorption tower, a power generation command signal to a generator that generates electricity using steam produced in a boiler that discharges the flue gas, the air flow rate in the boiler that discharges the flue gas, the flow rate of oxidizing air supplied to the absorption tower, the pH of the absorbent in the absorption tower, or the level of the absorbent in the absorption tower.

[0093] According to the embodiment of (15) above, by including at least one of these parameters as an input parameter, the absorbent concentration of the absorbent liquid in the absorption tower of the flue gas desulfurization plant can be suitably predicted as a state variable.

[0094] (16) A state quantity prediction system according to one embodiment is: A state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A predictive model creation unit for creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The prediction model creation unit creates the prediction model by applying regularization to the machine learning parameters of the prediction model and training it.

[0095] According to the embodiment of (16) above, a state quantity prediction system is realized that includes an information processing device capable of communicating with the state quantity prediction device according to each of the embodiments described above. In this system, the state quantity prediction device predicts the state quantities of the plant as described above in response to a request from the information processing device, and the prediction results can be obtained by the information processing device.

[0096] (17) A method for predicting state variables according to one embodiment is: A method for predicting plant state variables corresponding to input parameters, using a prediction model that includes a machine learning model and a physical model, The steps include creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, The steps include: predicting the state variables of the plant using the prediction model; Equipped with, In the step of creating the prediction model, the prediction model is created by applying regularization to the machine learning parameters of the prediction model and training it.

[0097] According to the embodiment of (17) above, when training a predictive model using training data acquired at the plant, the predictive model is created by applying regularization to the machine learning parameters of the predictive model during training. Generally, when training a predictive model using a small amount of training data, the training of machine learning parameters may become dominant, and the training of physical parameters may not be performed appropriately. In contrast, in this embodiment, even when the amount of training data is small, training is performed with regularization applied to prevent the machine learning parameters from becoming excessively large. This limits the degrees of freedom of the machine learning parameters, promotes the updating of physical parameters, avoids overfitting, and allows for the acquisition of a highly accurate predictive model.

[0098] (18) A control method for a state quantity prediction system according to one embodiment is: A control method for a state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A predictive model creation unit creates the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using the learning data acquired at the plant, after receiving a request from the information processing device. A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The prediction model creation unit creates the prediction model by applying regularization to the machine learning parameters of the prediction model and training it.

[0099] According to the embodiment of (18) above, a state quantity prediction system is controlled, which includes an information processing device that can communicate with the state quantity prediction device according to each of the embodiments described above. The state quantity prediction device predicts the state quantities of the plant as described above by performing a state quantity prediction method in response to a request from the information processing device, and the prediction results can be obtained by the information processing device. [Explanation of symbols]

[0100] 1(1A,1B) State variable prediction device 4. First Prediction Value Calculation Unit 6. Second Prediction Value Calculation Unit 8. State variable prediction value calculation unit 12 Input Layers 13. Middle Class 14 Output Layer 15. First Prediction Model Memory Unit 16 Measurement data acquisition unit 17. Second Prediction Model Creation Department 18 State Quantity Prediction Unit 20 Desulfurization equipment 22 Dust collector 23a, 23b Exhaust passage 24 Absorption Towers 25 Desulfurization exhaust gas pipe 26 Absorbent solution 28 Limestone feeder 30 Limestone 32 water 34 Absorbent fluid circulation pump 36 Absorbent fluid circulation piping 38 nozzles 40 Extraction tube 42 Dehydrator 46. ​​Oxidizing air 50, 52, 58 Concentration Sensors 54 Flow Sensor 56 Boiler air flow sensor 60. Oxidation air flow sensor 62 sensors 64-level sensor DE Neural O Dm measurement data M Predictive Model M1 First Prediction Model M2 Second Prediction Model M P Physical model M N Machine learning models P1 Plant No. 1 P2 Second Plant u Input parameter x State variable

Claims

1. A state quantity prediction device for predicting plant state quantities corresponding to input parameters using a prediction model that includes a machine learning model and a physical model, A first prediction model storage unit for storing a first prediction model, which is a prediction model in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, corresponding to the first plant, A measurement data acquisition unit for acquiring measurement data from the second plant, A second prediction model creation unit for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A state quantity prediction unit for predicting the state quantities of the second plant using the second prediction model, Equipped with, The second prediction model creation unit is a state quantity prediction device that creates the second prediction model by applying regularization to the difference in the machine learning parameters of the first prediction model and the second prediction model and learning from it.

2. The state quantity prediction device according to claim 1, wherein the second prediction model creation unit creates the second prediction model by performing transfer learning such that the loss function including the difference in machine learning parameters between the first prediction model and the second prediction model is minimized.

3. The state quantity prediction device according to claim 1 or 2, wherein the second prediction model creation unit sets an acceptable range for the physical parameters of the first prediction model and the second prediction model.

4. The state quantity prediction device according to claim 1 or 2, wherein the second prediction model creation unit adjusts the second prediction model such that the predicted value of the state quantity predicted by the state quantity prediction unit becomes the measured value of the state quantity.

5. The aforementioned plant is a flue gas desulfurization plant that desulfurizes the flue gas by bringing an absorbent liquid into contact with the flue gas in an absorption tower. The state quantity is the absorbent concentration in the absorption liquid in the absorption tower, according to claim 1 or 2, a state quantity prediction device.

6. The state quantity prediction device according to claim 5, wherein the input parameters include at least one of the SO2 concentration at the desulfurization outlet of the absorption tower, the SO2 concentration at the desulfurization inlet of the absorption tower, the flow rate or concentration of the limestone slurry produced in the absorption tower, a power generation command signal to a generator that generates power using steam produced in a boiler that discharges the flue gas, the air flow rate in the boiler that discharges the flue gas, the flow rate of oxidizing air supplied to the absorption tower, the pH of the absorbent in the absorption tower, or the level of the absorbent in the absorption tower.

7. A state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A first prediction model storage unit for storing a first prediction model, which is a prediction model in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, corresponding to the first plant, A measurement data acquisition unit, which acquires measurement data from the second plant after a request is received from the aforementioned information processing device, A second prediction model creation unit for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A state quantity prediction unit for predicting the state quantities of the second plant using the second prediction model, Equipped with, The state quantity prediction system includes a second prediction model creation unit which creates the second prediction model by applying regularization to the difference in machine learning parameters between the first prediction model and the second prediction model and learning from that difference.

8. A method for predicting plant state variables corresponding to input parameters, using a prediction model that includes a machine learning model and a physical model, A step of storing a first prediction model, which corresponds to the first plant, and in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, Steps include acquiring measurement data from the second plant, The steps include creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model using the aforementioned measurement data, A step of predicting the state variables of the second plant using the second prediction model, Equipped with, A method for predicting state variables, wherein in the step of creating the second predictive model, the second predictive model is created by applying regularization to the difference in machine learning parameters between the first predictive model and the second predictive model and learning from that difference.

9. A control method for a state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, After a request is received from the information processing device, a measurement data acquisition step is performed to acquire measurement data from the second plant, A second prediction model creation step for creating a second prediction model, which is the prediction model corresponding to the second plant, by adjusting the first prediction model stored in a first prediction model storage unit for storing the first prediction model, which is the prediction model corresponding to the first plant, in which the machine learning parameters included in the machine learning model and the physical parameters included in the physical model have been learned, using the aforementioned measurement data. A state quantity prediction step for predicting the state quantities of the second plant using the second prediction model, Execute, The method for controlling a state quantity prediction system is to create the second prediction model by performing regularization on the difference in machine learning parameters between the first prediction model and the second prediction model and learning from that difference.

10. A state quantity prediction device for predicting plant state quantities corresponding to input parameters using a prediction model that includes a machine learning model and a physical model, A predictive model creation unit for creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The aforementioned prediction model creation unit is a state quantity prediction device that creates the prediction model by applying regularization to the machine learning parameters of the prediction model and learning from them.

11. The state quantity prediction device according to claim 10, wherein the prediction model creation unit learns the prediction model such that the loss function including the magnitude of the machine learning parameters is minimized.

12. The state quantity prediction device according to claim 10 or 11, wherein the prediction model creation unit sets an acceptable range for the physical parameters during learning.

13. The state quantity prediction device according to claim 10 or 11, wherein the prediction model creation unit adjusts the prediction model so that the predicted value of the state quantity predicted by the state quantity prediction unit becomes the measured value of the state quantity.

14. The aforementioned plant is a flue gas desulfurization plant that desulfurizes the flue gas by bringing an absorbent liquid into contact with the flue gas in an absorption tower. The state quantity is the absorbent concentration in the absorption liquid in the absorption tower, according to claim 10 or 11, a state quantity prediction device.

15. The state quantity prediction device according to claim 14, wherein the input parameters include at least one of the SO2 concentration at the desulfurization outlet of the absorption tower, the SO2 concentration at the desulfurization inlet of the absorption tower, the flow rate or concentration of the limestone slurry produced in the absorption tower, a power generation command signal to a generator that generates power using steam produced in a boiler that discharges the flue gas, the air flow rate in the boiler that discharges the flue gas, the flow rate of oxidizing air supplied to the absorption tower, the pH of the absorbent liquid in the absorption tower, or the level of the absorbent liquid in the absorption tower.

16. A state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A predictive model creation unit for creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The aforementioned prediction model creation unit is a state quantity prediction system that creates the prediction model by applying regularization to the machine learning parameters of the prediction model and learning from them.

17. A method for predicting plant state variables corresponding to input parameters, using a prediction model that includes a machine learning model and a physical model, The steps include creating the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using training data acquired at the plant, The steps include: predicting the state variables of the plant using the prediction model; Equipped with, A method for predicting state variables, wherein the step of creating the predictive model involves creating the predictive model by applying regularization to the machine learning parameters of the predictive model and training it.

18. A control method for a state quantity prediction system comprising a state quantity prediction device and a communicationable information processing device for predicting plant state quantities corresponding to input parameters using a prediction model including a machine learning model and a physical model, The state quantity prediction device is, A predictive model creation unit creates the predictive model by learning the machine learning parameters included in the machine learning model and the physical parameters included in the physical model using the learning data acquired at the plant, after receiving a request from the information processing device. A state quantity prediction unit for predicting the state quantities of the plant using the prediction model, Equipped with, The control method for a state quantity prediction system includes a prediction model creation unit which creates the prediction model by applying regularization to the machine learning parameters of the prediction model and learning from them.

Images