WO2022210866A1 - Device, remote monitoring system, method for controlling device, and method for controlling remote monitoring system - Google Patents

Abstract

This device is for executing control of a plant on the basis of a prediction result obtained using a learning model. When a prediction result obtained using the learning model satisfies a prescribed condition, the device selects, from operational data as additional training data, data that is largely divergent from training data that was used to construct the learning model. Thereafter, new training data including the training data and the additional training data is used to reconstruct the learning model.

Description

Apparatus, remote monitoring system, apparatus control method, and remote monitoring system control method

The present disclosure relates to a device, a remote monitoring system, a control method for the device, and a control method for the remote monitoring system.
This application claims priority based on Japanese Patent Application No. 2021-061268 filed with the Japan Patent Office on March 31, 2021, the content of which is incorporated herein.

In a wet flue gas desulfurization system, which is an example of a plant, exhaust gas generated in a combustion device such as a boiler is introduced into an absorption tower of the desulfurization system, and brought into gas-liquid contact with an absorption liquid circulating in the absorption tower. During the gas-liquid contact process, the absorbent (e.g., calcium carbonate) in the absorbent reacts with the sulfur dioxide (SO ₂ ) in the flue gas, so that the SO ₂ in the flue gas is absorbed by the absorbent and removed from the flue gas. SO2 is removed ₍ exhaust gas is desulfurized). On the other hand, the absorbent that has absorbed SO ₂ drops and is stored in the storage tank below the absorption tower. An absorbent is supplied to the storage tank, and the absorbent whose absorption performance has been restored by the supplied absorbent is supplied to the upper part of the absorption tower by a circulating pump, and is subjected to gas-liquid contact with exhaust gas ₍ SO2 absorption). be done.

In such a wet flue gas desulfurization apparatus, it takes a considerable amount of time until changes in the circulation flow rate of the absorbent and the concentration of the absorbent are reflected in the concentration of SO ₂ in the flue gas. Therefore, in the control of wet flue gas desulfurization equipment, the relationship between the control parameters such as the circulation flow rate of the absorbent and the concentration of the absorbent and the SO ₂ concentration in the exhaust gas is constructed as a learning model by machine learning. Based _on the SO2 concentration in the exhaust gas predicted by the model, it is possible to reduce the _SO2 concentration in the exhaust gas to below the standard value by determining the control target values for the circulation flow rate of the absorbent and the concentration of the absorbent. Become. For example, in Patent Document 1, in the control of such a wet flue gas desulfurization system, a control target for the circulation amount of the absorbent is used using a first learning model that represents the relationship between the circulation amount of the absorbent and the SO ₂ concentration in the exhaust gas. In addition to obtaining the value, the control target value for the absorbent concentration of the absorbent is obtained by using the _second learning model representing the relationship between the absorbent concentration of the absorbent and the SO2 concentration in the exhaust gas. and methods of controlling absorbent concentration are disclosed.

JP 2020-11163 A

In plant control as described above, a learning model is prepared in advance for predicting control target values. The construction of such a learning model is performed by machine learning using learning data selected from plant operating data. sometimes you can't get it. Moreover, even if the prediction error of the learning model is sufficiently small at the beginning, the prediction error may increase later due to changes in plant operating conditions from the time the learning model was constructed.

In this way, if the prediction error of the learning model is not sufficient, it will be necessary to reconstruct the learning model. The learning model is reconstructed, for example, using learning data to which new data has been added. Sometimes it's gone. In addition, since the amount of driving data that is the basis of learning data is enormous, it is required to efficiently select data for reducing prediction errors by reconstructing the learning model.

At least one embodiment of the present disclosure has been made in view of the above circumstances, and a device capable of realizing good control accuracy by efficiently selecting learning data used to reconstruct a learning model, remote monitoring It is an object of the present invention to provide a system, a device control method, and a remote monitoring system control method.

In order to solve the above problems, the device according to at least one embodiment of the present disclosure,
A device for executing processing related to plant control based on prediction results using a learning model,
Additional learning data selection for selecting, as additional learning data, data having a large deviation from the learning data used to construct the learning model from the driving data when the prediction result using the learning model satisfies a predetermined condition. Department and
a learning model construction unit for reconstructing the learning model using new learning data including the learning data and the additional learning data;
Prepare.

In order to solve the above problems, the remote monitoring system according to at least one embodiment of the present disclosure,
A remote monitoring system comprising a terminal capable of communicating with a device for executing processing related to plant control based on prediction results using a learning model,
The device comprises:
In response to a request from the terminal, when the prediction result using the learning model satisfies a predetermined condition, data with a large deviation from the learning data used to construct the learning model from the driving data is selected as additional learning data. an additional learning data selection unit for
a learning model construction unit for reconstructing the learning model using learning data including the learning data and the additional learning data;
Prepare.

In order to solve the above problems, a device control method according to at least one embodiment of the present disclosure includes:
A control method for a device for executing processing related to plant control based on prediction results using a learning model,
Additional learning data selection for selecting, as additional learning data, data having a large deviation from the learning data used to construct the learning model from the driving data when the prediction result using the learning model satisfies a predetermined condition. a step;
A learning model building step for reconstructing the learning model using learning data including the learning data and the additional learning data;
Prepare.

In order to solve the above problems, a remote monitoring system control method according to at least one embodiment of the present disclosure includes:
A control method for a remote monitoring system comprising a terminal capable of communicating with a device for executing processing related to plant control based on prediction results using a learning model,
In response to a request from the terminal, when the prediction result using the learning model satisfies a predetermined condition, data with a large deviation from the learning data used to construct the learning model from the driving data is selected as additional learning data. an additional learning data selection step for
a learning model building step of reconstructing the learning model using learning data including the learning data and the additional learning data;
Prepare.

According to at least one embodiment of the present disclosure, by efficiently selecting learning data used to reconstruct a learning model, a device capable of realizing good control accuracy, a remote monitoring system, a device control method, and , can provide a control method for a remote monitoring system.

It is a lineblock diagram of the flue gas desulfurization equipment concerning one embodiment. 1 is a configuration schematic diagram of a remote monitoring system according to one embodiment; FIG. 4 is a flow chart showing basic control of a wet flue gas desulfurization system according to one embodiment. 4 is a graph showing transitions of predicted values of SO ₂ concentration in outflow gas, measured values of SO ₂ concentration by a gas analyzer, and true values of predicted values of SO ₂ concentration. FIG. 4 is a diagram schematically showing an example of a first relationship table created in basic control of a wet flue gas desulfurization system according to one embodiment; FIG. 4 is a diagram schematically showing an example of a second relationship table created in basic control of a wet flue gas desulfurization system according to one embodiment; It is a flow chart which shows a plant control method concerning one embodiment. 8 is a flowchart showing a method of selecting additional learning data in step S104 of FIG. 7; FIG. 9 is a diagram showing a process of selecting additional learning data in step S204 of FIG. 8; FIG. 9 is a diagram showing a process of selecting additional learning data in step S204 of FIG. 8; FIG. 10 is a diagram showing the distribution of learning data used for reconstruction of the first learning model for each number of times reconstruction is performed; FIG. 11 is a diagram showing changes in predicted values of the first learning model reconstructed using the learning data shown in FIG. 10; 1 is a configuration diagram of an information processing system according to an embodiment; FIG. 13 is a diagram showing an internal configuration of the information processing apparatus of FIG. 12 together with a control device; FIG.

Several embodiments of the present invention will be described below with reference to the drawings. However, the scope of the present invention is not limited to the following embodiments. The dimensions, materials, shapes, relative positions, and the like of components described in the following embodiments are not intended to limit the scope of the present invention, but merely illustrative examples.

First, referring to FIG. 1, the configuration of a wet flue gas desulfurization system 10, which is an example of a plant controlled by a plant control device according to at least one embodiment of the present disclosure, will be described. FIG. 1 is a configuration diagram of a wet flue gas desulfurization apparatus 10 according to one embodiment.

In the following description, the wet-type flue gas desulfurization system 10 will be described as an example of a plant, but the control target is not limited to the wet-type flue gas desulfurization system 10. Control parameters based on control target values predicted using a learning model can broadly include plants in which is controlled.

The wet flue gas desulfurization device 10 is plant equipment for desulfurizing the exhaust gas generated in the combustion device 1 . The combustion device 1 is, for example, a boiler for generating steam, and is configured as part of a power plant capable of generating power by supplying the steam generated by the combustion device 1 to the power generator 5 . The wet flue gas desulfurization apparatus 10 includes an absorption tower 11 communicating with a combustion apparatus 1 via a pipe 2, and a plurality of circulation pumps 12a, 12b provided in a circulation pipe 3 for absorbing liquid circulating in the absorption tower 11. 12c (when these are collectively referred to as a “circulation pump 12” as appropriate), and a slurry (absorbent slurry) of calcium carbonate (CaCO ₃ ), which is an absorbent contained in the absorbent, is supplied into the absorption tower 11. and a gypsum recovery unit 14 for recovering gypsum in the absorbent. The absorber 11 is provided with an outflow pipe 16 through which the exhaust gas desulfurized by the operation described later flows out from the absorber ₁₁ as an outflow gas. gas analyzer 17 is provided.

The absorbent slurry supply unit 13 includes an absorbent slurry production facility 21 for producing absorbent slurry, an absorbent slurry supply pipe 22 communicating between the absorbent slurry production facility 21 and the absorption tower 11, and an absorbent slurry. An absorbent slurry supply amount control valve 23 for controlling the flow rate of the absorbent slurry flowing through the supply pipe 22 is provided. The gypsum recovery unit 14 includes a gypsum separator 25, a gypsum slurry withdrawal pipe 26 communicating between the gypsum separator 25 and the absorption tower 11, and a gypsum slurry withdrawal pump 27 provided in the gypsum slurry withdrawal pipe 26. I have.

The wet flue gas desulfurization system 10 is provided with a control device 15 that is a device for controlling the plant of at least one embodiment of the present disclosure. The control device 15 acquires operation data including various detectors for acquiring various operation data (e.g., temperature and pressure at various parts, flow rates of various fluids, etc.) of the combustion device 1 and the wet flue gas desulfurization device 10. A driving data receiving unit 30 electrically connected to the unit 20 is provided. The operating data acquisition unit 20 includes a gas analyzer 17 .

The control device 15 includes a first learning model constructing unit 38 electrically connected to the operation data receiving unit 30, a first relationship table constructing unit 31 electrically connected to the first learning model constructing unit 38, a first A circulation flow determination unit 32 electrically connected to the 1-relation table creation unit 31 and a circulation pump adjustment unit 33 electrically connected to the circulation flow determination unit 32 are provided. The circulation pump adjustment unit 33 is electrically connected to each of the circulation pumps 12a, 12b, 12c.

The control device 15 further includes a second learning model constructing unit 39 electrically connected to the operation data receiving unit 30, a second relationship table generating unit 35 electrically connected to the second learning model constructing unit 39, An absorbent slurry supply amount determination unit 36 electrically connected to the second relationship table creation unit 35, and an absorbent slurry supply control unit 37 electrically connected to the absorbent slurry supply amount determination unit 36. there is The absorbent slurry supply controller 37 is electrically connected to the absorbent slurry supply amount control valve 23 .

The control device 15 further includes a prediction error calculator 40 electrically connected to the first learning model construction unit 38 and the second learning model construction unit 39, and additional learning data electrically connected to the prediction error calculation unit 40. A selection unit 42 is provided.

FIG. 2 shows the configuration of a remote monitoring system 44 for remotely monitoring the control state of the wet flue gas desulfurization equipment 10 (see FIG. 1). The remote monitoring system 44 is electrically connected to a distributed control system (DCS) 46 of each device constituting the combustion device 1 (see FIG. 1) and the wet flue gas desulfurization device 10 (see FIG. 1), and to the DCS 46. Equipped with an edge server 48 equipped with a control device 15 and a remote monitoring device 50 such as a desktop personal computer, tablet computer, etc. electrically connected to the edge server 48 via a cloud or virtual private network (VPN). there is A control state of the wet flue gas desulfurization apparatus 10 can be remotely monitored by a remote monitoring device 50 which is normally located at a location remote from the edge server 48 .

Next, the operation of desulfurizing the exhaust gas generated in the combustion device 1 by the wet flue gas desulfurization device 10 will be described.
As shown in FIG. 1, the exhaust gas generated in the combustion device 1 flows through the pipe 2, flows into the absorption tower 11, and rises in the absorption tower 11. As shown in FIG. By operating at least one of the circulation pumps 12 , the absorbent flows through the circulation pipe 3 and into the absorption tower 11 , where the absorption liquid flows down. The absorbent flowing down in the absorption tower 11 accumulates in the absorption tower 11 , flows out from the absorption tower 11 by the circulation pump 12 , and flows through the circulation pipe 3 . Thus, the absorbent circulates within the absorption tower 11 .

In the absorption tower 11, the rising exhaust gas and the flowing absorbing liquid come into gas-liquid contact. SO ₂ contained in the exhaust gas is expressed by the following reaction formula: SO ₂ +CaCO ₃ +2H ₂ O+1/2O ₂ →CaSO _{4.2H 2} O+CO ₂
, reacts with CaCO ₃ in the absorbent to precipitate gypsum (CaSO ₄ .2H ₂ O) in the absorbent.

In this way, since part of the SO ₂ in the flue gas is removed as gypsum in the absorbent, that is, the flue gas is desulfurized, SO ₂ concentration is lower than the SO ₂ concentration in the flue gas flowing into the absorption tower 11 via the pipe 2 . _The effluent gas that has flowed out of the absorption tower 11 flows through the outflow pipe 16 and is released into the atmosphere. It is transmitted to the receiver 30 .

The SO ₂ concentration in the effluent gas tends to decrease as the circulation flow rate of the absorbent circulating in the absorption tower 11 increases, unless the CaCO ₃ concentration in the absorbent fluctuates greatly. The control device 15 controls the number of operating circulation pumps 12 by a control method described later to control the circulation flow rate, thereby controlling the concentration of SO ₂ in the outflow gas, for example, below a preset value. _So the SO2 concentration in the effluent gas can be controlled.

The gypsum precipitated in the absorbing liquid in the absorption tower 11 is extracted as gypsum slurry from the absorption tower 11 by the gypsum slurry extraction pump 27 , and the gypsum slurry flows through the gypsum slurry extraction pipe 26 to the gypsum separator 25 . flow into Gypsum and water are separated in the gypsum separator 25, the gypsum is recovered, and the water is sent to a drainage facility (not shown).

Since CaCO ₃ in the absorbing liquid reacts with SO ₂ to form gypsum, the concentration of CaCO ₃ in the absorbing liquid decreases as the exhaust gas is desulfurized. The control device 15 controls the opening degree of the absorbent slurry supply amount control valve 23 by a control method to be described later, and absorbs the absorbent slurry produced in the absorbent slurry production facility 21 through the absorbent slurry supply pipe 22. Feed into column 11 . As a result, the CaCO ₃ concentration in the absorbent falls within a preset range, and large fluctuations in the CaCO ₃ concentration during desulfurization of the flue gas are suppressed.

Next, basic control of the wet flue gas desulfurization system 10 by the control device 15 will be described. FIG. 3 is a flow chart showing basic control of the wet flue gas desulfurization system 10 performed by the controller 15 of FIG.

In the basic control, first, in step S1, after collecting various operational data of the combustion device 1 and the wet-type flue gas desulfurization device 10, in step S2, various operational data and future SO ₂ in the effluent gas flowing out from the absorber 11 A first learning model is constructed by machine learning for the relationship with concentration. Next, in step S3, a first relationship table, which will be described later, is created using the constructed first learning model. In the following step S4, based on the first relationship table, the circulation flow rate of the absorbent that makes the SO ₂ concentration in the outflow gas equal to or lower than a preset value is determined, and in step S5, based on the determined circulation flow rate to adjust the operating conditions of the circulation pump 12. Thereby, the SO ₂ concentration in the outflow gas is controlled so as to be equal to or lower than the preset value.

After step S1, aside from steps S2 to S5, in step S12, a second learning model is constructed by machine learning regarding the relationship between various operating data and the future CaCO ₃ concentration in the absorbent. Next, in step S13, a second relationship table, which will be described later, is created using the constructed second learning model. In subsequent step S14, based on the second relationship table, the supply amount of the absorbent slurry that brings the CaCO ₃ concentration within a preset range is determined, and in step S15, the absorbent slurry supply unit 13 is controlled. That is, by controlling the opening degree of the absorbent slurry supply amount control valve 23, the absorbent slurry is supplied into the absorption tower 11 at the determined supply amount. As a result, the CaCO ₃ concentration in the absorbent falls within a preset range, and large fluctuations in the CaCO ₃ concentration during desulfurization of the flue gas are suppressed.

Next, each step of the method of controlling the wet flue gas desulfurization apparatus 10 by the control device 15 will be described in detail.
In step S1, as shown in FIG. 1, after the operation data acquisition unit 20 acquires various operation data of the combustion device 1 and the wet-type flue gas desulfurization device 10, the acquired various operation data are transmitted to the control device 15. The control device 15 collects various types of operation data as the operation data receiving unit 30 receives the operation data. As described above, since the operating data acquisition section ₂₀ includes the gas analyzer 17, various operating data include the SO2 concentration in the outflow gas.

In step S2, the first learning model constructing unit 38 constructs a first model by machine learning regarding the relationship between the various operational data collected by the control device 15 and the future SO ₂ concentration in the outflow gas. In step S3, using the constructed first learning model, the first relationship table creation unit 31 calculates the circulation flow rate of the absorbent at a first time and the outflow at a second time, which is a time later than the first time. A _first relationship table is created that is a correlation with the predicted SO2 concentration in the gas. Since the first relationship table is created using the first learning model constructed by machine learning, the first relationship table can be created quickly.

In the first relationship table, the circulation flow rate of the absorbent and the predicted value of the SO2 concentration in the outflow gas _are different in _time . is the predicted value of the SO2 concentration, for example, a _few minutes from now. For this reason, various operational data include at _least the concentration of SO2 in the outflow gas at an arbitrary time and the circulation flow rate of the absorbent at a time interval obtained by subtracting the first time from the second time. and are included. From actual operating data, including the SO2 concentration in the effluent gas at an arbitrary time and the circulation flow rate of the absorbent at a time earlier than the arbitrary time by the time interval of the _second time minus the first time, the future Since the _SO2 concentration is directly predicted, the prediction performance of future _SO2 concentration can be improved. Note that the shorter the interval between the first time and the second time, the better the performance of predicting the future SO ₂ concentration. Therefore, the interval between the first time and the second time is the time required for the SO ₂ concentration in the outflow gas to change due to changes in the circulation flow rate of the absorbent, and the time required for the gas analyzer 17 to change the SO ₂ concentration. is preferably the sum of the time required to measure .

In FIG. 4, the interval between the first time and the second time is the time required for the SO ₂ concentration in the outflow gas to change due to changes in the circulation flow rate of the absorbent, and the time required for the SO 2 concentration in the outflow gas to change. transition (a) of the predicted value of the SO ₂ concentration, transition ( _b ) of the measured value of the SO ₂ concentration by the gas analyzer 17, and SO ₂ 3 shows the transition (c) of the true value of the density. In each graph, the right side is the past value, and the leftmost side is the latest value. The latest measured value of the SO ₂ concentration by the gas analyzer 17 is the value at the first hour, and the latest value of the predicted SO ₂ concentration is the value at the second hour. _The interval (i) between the latest measured value of the SO2 concentration by the gas analyzer 17 and the latest value of the true value of the SO2 concentration is the _time required for the gas analyzer 17 to measure the _SO2 concentration, i.e. Corresponding to the measurement lag, the interval _{( ii} ) between the latest true value of the SO2 concentration and the latest value of the predicted SO2 concentration is the amount of SO ₂ corresponds to the time required for the concentration to change.

FIG. 5 shows an example of the first relationship table. In this embodiment, the first relationship table is represented as a graph in which the horizontal axis represents the predicted value of the SO ₂ concentration in the outflow gas and the vertical axis represents the circulating flow rate of the absorbent. , and may be in the form of a matrix, formula, or the like. In step S4, the circulation flow rate determination unit 32 determines, based on the first relationship table, the absorption liquid circulation flow rate Q (control target value) at which the future SO ₂ concentration in the outflow gas will be the preset set value SV. to decide.

In step S5, as shown in FIG. 1, the circulation pump adjustment unit 33 determines the number of operating circulation pumps 12a to 12c so that the determined circulation flow rate Q (control target value) or more is determined. Make sure that the number of circulation pumps in operation is in operation. For example, when the three circulating pumps 12a to 12c operate at the same supply rate, it is possible to adjust the circulating flow rate in three steps. By increasing the number of circulation pumps, finer adjustment of the circulation flow rate becomes possible. Further, for example, when the three circulation pumps 12a to 12c operate with different supply amounts, it is possible to adjust the circulation flow rate in a maximum of six stages by combining the operating circulation pumps. Further, for example, if each of the three circulation pumps 12a to 12c can adjust the supply amount, finer adjustment of the circulation flow rate becomes possible.

It should be noted that the adjustment of the circulation flow rate is not limited to being performed by controlling the number of circulation pumps. A single circulation pump whose supply amount is adjustable may be used, and the supply amount of the circulation pump may be adjusted so as to achieve the circulation flow rate determined by the circulation flow rate determination unit 32 .

In this way, by adjusting the circulation flow rate of the absorbent circulating in the absorption tower 11, it is possible to control the future SO ₂ concentration in the outflow gas to a preset value or less. For this, it is necessary that the CaCO ₃ concentration in the absorbing liquid does not fluctuate greatly. Therefore, in this embodiment, as described above, in addition to steps S2 to S5, steps S12 to S15 are performed to control the CaCO ₃ concentration in the absorbent so that it is within a preset range. . Next, each of steps S12 to S15 will be described in detail.

In step S12, the second learning model construction unit 39 uses machine learning to create a second learning model for the relationship between the various operating data collected by the control device 15 and the future CaCO ₃ concentration in the absorbent in the absorption tower 11. To construct. In step S13, using the constructed second learning model, the second relationship table creation unit 35 determines the amount of absorbent slurry supplied to the absorption tower 11 at the third time and A second relationship table is created which is a correlation with the predicted value of CaCO ₃ concentration at a certain fourth hour. Since the second relationship table is created using the second learning model constructed by machine learning, the second relationship table can be created quickly.

In the second relationship table, the amount of absorbent slurry supplied to the absorption tower 11 and the predicted value of CaCO ₃ concentration differ in time, and if the amount of absorbent slurry supplied is the current value, the predicted value of CaCO ₃ concentration , for example, the predicted value of the CaCO ₃ concentration a few minutes from now. For this reason, the various operation data include at least the CaCO ₃ concentration at an arbitrary time and the supply amount of the absorbent slurry at a time interval obtained by subtracting the third hour from the fourth hour before the arbitrary time. include. future CaCO ₃ from actual operating data including the CaCO ₃ concentration at a given time and the supply rate of absorbent slurry at a time prior to the given time by the time interval of the fourth hour minus the third hour; Since the concentration is predicted directly, the prediction performance of future CaCO ₃ concentration can be improved.

In this embodiment, the CaCO ₃ concentration at an arbitrary time uses a value calculated using a simulation model based on mass balance calculation. A sensor for detecting the CaCO ₃ concentration is generally expensive. However, if the CaCO ₃ concentration is calculated using a simulation model based on mass balance calculation, an expensive sensor becomes unnecessary, and an increase in the cost of the wet flue gas desulfurization apparatus 10 can be suppressed.

It should be noted that the shorter the interval between the third time and the fourth time, the better the prediction performance of the future CaCO ₃ concentration. Therefore, the interval between the third time period and the fourth time period is preferably the time required for the CaCO ₃ concentration to change due to the change in the supply amount of the absorbent slurry. Transition of the predicted value and transition of the true value of the supply amount of the absorbent slurry have the same relationship as the transition (a) of the predicted value of the SO ₂ concentration and the transition of the true value (c) of FIG. 4, respectively. In this embodiment, the CaCO ₃ concentration is calculated using a simulation model based _on mass balance calculation. The transition of the measured value and the transition of the true value have the same relationship as the various transitions (a) to (c) of the SO ₂ concentration in FIG.

In general, the number of steps required to change the SO2 concentration in the effluent gas flowing out of the absorber tower ₁₁ is greater _than the number of steps required to change the _CaCO3 concentration. The delay in controlling the SO ₂ concentration is large compared to the conventional method. Therefore, by making the time from the third time to the fourth time shorter than the time from the first time to the second time, the influence of the control delay can be appropriately considered, so future CaCO ₃ Concentration prediction performance can be further improved.

FIG. 6 shows an example of the second relationship table. In this embodiment, the second relationship table is represented as a graph in which the horizontal axis indicates the predicted value of the CaCO ₃ concentration and the vertical axis indicates the supply amount of the absorbent slurry. Instead, it may be in the form of a matrix, formula, or the like. In step S14, the absorbent slurry supply amount determination unit ₃₆ determines, based on the second relationship table, the absorbent slurry supply amount F (control target value ).

In step S15, as shown in FIG. 1, the absorbent slurry supply control unit 37 determines the amount of absorbent slurry to be supplied into the absorption tower 11 via the absorbent slurry supply pipe 22. The opening degree of the absorbent slurry supply amount control valve 23 is controlled so as to approach the absorbent slurry supply amount F (control target value). By adjusting the supply amount of the absorbent slurry to the absorption tower 11 in this way, it is possible to control the CaCO ₃ concentration in the future to be within a preset range.

In this way, from the operation data of the combustion device 1 and the operation data including the circulation flow rate of the absorbent of the wet-type flue gas desulfurization device 10, the circulation flow rate of the absorbent at the first time and the future time from the first time Since the future SO2 concentration is directly predicted from the actual operation data by creating the _first relationship table between the SO2 concentration in the effluent gas flowing out of the absorber 11 at the _second time, the future can obtain a first relationship table that improves the prediction performance of the SO2 concentration of , and based on this _first relationship table, the SO2 concentration in the outflow gas at the _second time is equal to or less than the preset value The circulation flow rate of the absorbent at the first time is determined as described above, and the operating conditions of the circulation pumps 12a to 12c are adjusted based on the determined circulation flow rate at the first time. can be adjusted appropriately.

In this embodiment, steps S12 to S15 _{are performed so that the CaCO 3 concentration} in the absorbent falls within a preset range. If the amount of absorbent slurry supplied to the absorption tower 11 is adjusted as needed based on the value, steps S12 to S15 can be eliminated. In this case, the control device 15 does not have to include the second learning model construction section 39 , the second relationship table creation section 35 , the absorbent slurry supply amount determination section 36 , and the absorbent slurry supply control section 37 .

Next, in addition to the basic control shown in FIG. 3, a plant control method according to one embodiment implemented by the control device 15 will be described. FIG. 7 is a flow chart showing a plant control method according to one embodiment.

In this plant control, the predicted value of the first learning model is calculated in step S100, apart from steps S2 to S5 shown in FIG. Next, in step S101, the analysis result by the gas analyzer 17 is obtained. In subsequent step S102, the predicted value calculated in step S100 and the analysis result obtained in step S101 are compared to calculate the prediction result of the first learning model, that is, the prediction error. In subsequent step S103, when the prediction error calculated in step S102 satisfies a predetermined condition, for example, it is determined whether the prediction error is greater than a threshold. If the prediction error is greater than the threshold (step S103: YES), additional learning data is selected in subsequent step S104, and the first learning model is reconstructed in step S105 using the additional learning data selected in step S104. Then, in step S106, a prediction error is calculated for the first learning model reconstructed in step S105, and in step S107, it is determined whether or not the prediction error is equal to or less than a threshold. If the prediction error calculated in step S106 is still larger than the threshold (step S107: NO), the process returns to step S104 to repeatedly select additional learning data and reconstruct the learning model. Such repeated processing is performed until the predicted value of the reconstructed first learning model becomes equal to or less than the threshold.
If the prediction error is equal to or less than the threshold in step S103 (step S103: NO), the process ends, but the series of processes shown in FIG. 7 may be repeated at a predetermined timing.

Next, each step in FIG. 7 will be described in detail.
In step S100, the first learning model constructed by the first learning model construction unit 38 is used to calculate the predicted value of the SO ₂ concentration in the outflow gas. Calculation of the predicted value by the first learning model in step S100 is the same as the calculation of the predicted value of the SO ₂ concentration in the outflow gas to create the first relationship table in step S3. A predicted value of the SO ₂ concentration in the outflow gas at a second time, which is a time later than the first time, is calculated for the circulating flow rate of the absorbent at the first time that is input to the first learning model. be.

In step S101, based on the analysis result by the gas analyzer 17, a measured value of SO ₂ concentration in the outflow gas is obtained. This measured value is the actual SO2 concentration in the outflow gas at the _{second time} corresponding to the predicted value of the SO2 concentration in the outflow gas calculated in step S100.

_In step S102, the prediction error calculation unit 40 calculates the difference between the predicted value of the SO2 concentration in the outflow gas calculated in step S100 and the measured value of the _SO2 concentration in the outflow gas obtained in step S101. Calculate the prediction error. This prediction error is an error corresponding to the prediction accuracy of the first learning model constructed by the first learning model construction unit 38, and includes various factors. For example, since the driving data received by the driving data receiving unit 30 has some variation, the first learning model constructed by machine learning using the driving data as learning data has a learning error caused by the variation. . In addition, the prediction error may increase later due to changes in plant operating conditions from the time the model was constructed.

At step S103, it is determined whether or not such a prediction error is greater than a preset threshold value ε. The success/failure determination in step S103 may be made so as to be established when the prediction error continues to exceed the threshold value ε for a predetermined period of time or longer. The magnitude of the prediction error may also vary depending on the operating state of the wet-type flue gas desulfurization system 10. If it is determined in step S103 based on a short-term determination, the first learning model will be rebuilt frequently. There is a risk that the burden of model management will increase. Therefore, in step S103, when the state in which the prediction error is larger than the threshold value ε continues for a predetermined time or longer, it is determined in step S103 that the first learning model is reconstructed appropriately and efficiently. model management becomes possible.

In step S104, when the establishment determination is made in step S103, the additional learning data selection unit 42 selects additional learning data included in the learning data used for reconstructing the first learning model. The learning data used for reconstruction includes new additional learning data in addition to the old initial learning data used when building the first learning model last time (that is, learning data used for reconstruction = initial learning data + additional training data). The operation data receiving unit 30 continuously receives operation data, and appropriate additional learning data is selected from the operation data received after the previous first learning model was constructed.

In addition, the selection of additional learning data in step S104 may be performed with respect to the operation data acquired during steady operation of the plant. For example, the operation data acquired during non-steady operation such as when an abnormality occurs in the plant, when the operation is started, when the operation is stopped, etc., are excluded from the selection targets of the additional learning data. If the operating data includes data acquired during unsteady operation, the operating data may be excluded by performing preprocessing.

Here, a method for selecting additional learning data by the additional learning data selection unit 42 will be specifically described with reference to FIG. FIG. 8 is a flow chart showing a method of selecting additional learning data in step S104 of FIG.

In step S200, first, by analyzing the operating data received by the operating data receiving unit 30, at least one explanatory variable of the first learning model is selected from a plurality of parameters included in the operating data. For example, the selection of such explanatory variables contributes to the SO ₂ concentration in the outflow gas, which is the objective variable of the first learning model, for each of the plurality of operating data included in the operating data, by a method such as multiple regression. may be calculated based on the contribution. For example, Z parameters may be selected as explanatory variables in descending order of contribution. By selecting some of the parameters included in the driving data as explanatory variables for the first learning model in this way, the learning accuracy is lower than when all the parameters included in the driving data are subject to learning. can be effectively reduced while suppressing

In step S201, the explanatory variables selected in step S200 from among the learning data (operating data) used in the previous construction of the first learning model are selected as initial learning data. At this time, as the initial learning data, an average value over W hours for V selected from the learning data (operating data) used in the previous construction of the first learning model may be used. In this case, by using the average value of a specific parameter included in the operation data over a predetermined period of time as the learning data, it is possible to effectively reduce the amount of calculation during learning while suppressing a decrease in learning accuracy.

In step S202, additional learning data candidates are selected from the driving data received by the driving data receiving unit 30 for the explanatory variables selected in step S200. The additional learning data candidate is selected from new driving data received by the driving data receiving unit 30 from the previous construction of the first learning model to the present, and corresponds to the initial learning data selected in step S201. Contains parameters to For example, when the average value over W hours is used as the initial learning data as described above, the average value over W hours is also used as the additional learning data candidate.

In step S203, the degree of divergence between the initial learning data selected in step S201 and the additional learning data candidates selected in step S202 is calculated. For calculating the degree of deviation, various methods for evaluating the degree of deviation can be used, such as the k nearest neighbor method and the Mahalanobis distance. In step S204, additional learning data to be added to the learning data is selected based on the deviation calculated in step S203.

Here, FIGS. 9A and 9B are diagrams showing the process of selecting additional learning data in step S204 of FIG.

In the embodiment of FIG. 9A, in a space defined by arbitrary variables 1 and 2 included in the explanatory variables of the first learning model, a plurality of additional learning data candidates Dc1, Dc2, and Dc3 are obtained for certain initial learning data Ds. , . . . are shown, and the distances indicating the degrees of divergence between the initial learning data Ds and the additional learning data candidates Dc1, Dc2, Dc3, . In this example, the additional learning data selection unit 42 selects the additional learning data candidate Dc5 having the largest distance among the plurality of additional learning data candidates as the additional learning data.

9B, in the space defined by arbitrary variables 1 and 2 included in the explanatory variables of the first learning model, the plurality of additional learning data candidates Dc1, Dc2, Dc3, . A plurality of initial learning data Ds1, Ds2, . . . selected in step S201 are shown for . Then, the distance to the closest initial learning data is calculated for each of the additional learning data candidates Dc1, Dc2, Dc3, . The additional learning data selection unit 42 selects, as additional learning data, the one with the largest distance among the plurality of additional learning data candidates. In FIGS. 9A and 9B, the number of variables used for determination of addition of learning data is two, but the scope of the present invention is not limited, and the number may be one or three or more in practice.

The additional learning data selection unit 42 calculates the degree of divergence between the initial learning data and the additional learning data candidates in this way, and based on the degree of divergence, should be added to the learning data for reconstructing the first learning model. Select learning data candidates. The number of additional learning data to be newly added may be arbitrary. ) additional learning data can be selected.

In the present embodiment, on the premise that the first learning model has already been constructed by the first learning model constructing unit 38, the learning data used when constructing the first learning model last time is treated as the initial learning data. However, if there is no construction history of the first learning model (for example, when the first learning model is constructed for the first time), one or more parameters arbitrarily selected from the operating data may be treated as initial learning data. In this case, even when the first learning model is constructed for the first time, it is possible to construct a learning model with less prediction error.

Returning to FIG. 7, in step S105, new learning data is created by adding the additional learning data selected in step S104 to the initial learning data, and the first learning model is reconstructed. As a result, the first learning model is constructed using new learning data obtained by adding additional learning data selected from the subsequently obtained driving data to the initial learning data used in the previous construction of the first learning model. Reconstruction can be performed.

Then, in step S106, the prediction error is calculated using the first learning model reconstructed in step S105. Calculation of the prediction error in step S106 is the same as in step S102 described above.

In step S107, similarly to step S103, it is determined whether or not the prediction error calculated in step S106 is equal to or less than the threshold ε. That is, it is determined whether the prediction error of the first learning model has been sufficiently improved by reconstruction. As a result, when the prediction error of the first learning model is improved to be equal to or less than the threshold value ε, the processing is terminated assuming that the prediction accuracy of the first learning model has been improved. On the other hand, if the prediction error of the first learning model is still larger than the threshold ε (step S107: NO), the process returns to step S104. That is, if the prediction error of the first learning model is not sufficiently improved even by reconstruction, additional learning data is selected again in step S104, and the learning data is reviewed, and then the first Construction of the learning model is performed repeatedly. Such reconstruction of the first learning model is repeated until the prediction error becomes equal to or less than the threshold value ε in step S107.

Here, the change in the predicted value of the first learning model due to the number of times the reconstruction is performed will be specifically described. FIG. 10 shows the distribution of the learning data used for reconstruction of the first learning model (the learning data of the SO ₂ concentration in the outflow gas which is the objective variable and the explanatory variable X used for the learning model) for each number of times of reconstruction. FIG. 11 is a diagram showing changes in predicted values of the first learning model reconstructed using the learning data shown in FIG. 10. FIG.

FIG. 10 shows how the number of data included in the learning data increases as the number of reconstruction implementations increases, as new additional learning data is selected in step S104. The prediction error of the first learning model reconstructed using such learning data decreases as the number of times of reconstruction increases, as shown in FIG. This indicates that the prediction accuracy of the first learning model is improved by appropriately selecting additional learning data each time reconstruction is performed.

Note that as the number of reconstruction operations increases, the prediction error of the first learning model converges to a predetermined value (near 0.7 in the example of FIG. 11). Therefore, in step S107, in addition to or instead of whether the predicted value of the first learning model is equal to or smaller than the threshold value, based on whether the prediction error has sufficiently converged, the iterative process after step S104 is terminated. A judgment may be made.

In this way, by selecting and adding the additional learning data to the learning data for building the first learning model, new learning data is created, and the first learning model is reconstructed using the learning data. will be At this time, the additional learning data added to the learning data is appropriately selected based on the degree of divergence from the initial learning data conventionally included in the learning data, thereby effectively reducing the prediction error of the first learning model. can. As a result, even if the prediction error of the first learning model is reduced for some reason, good prediction accuracy can be obtained by reconstructing the first learning model.

By reconstructing the first learning model with improved prediction accuracy in the first learning model construction unit 38, the control device 15 obtains the control target value for the circulation flow rate based on the predicted value of the first learning model. can be set with high precision. As a result, the circulation pump adjustment unit 33 can appropriately control the circulation flow rate by adjusting the number of circulation pumps 12 based on the control target value.

Reducing the prediction error by reconstructing such a learning model can be similarly performed for the second learning model handled by the second learning model construction unit 39. That is, when the prediction error of the second learning model calculated by the prediction error calculation unit 40 is equal to or less than the threshold, the additional learning data selection unit 42 adds the second learning model to the learning data for reconstructing the second learning model. Additional learning data to be added is selected, and the second learning model is reconstructed using new learning data including the additional learning data. At this time, the additional learning data added to the learning data is appropriately selected based on the degree of divergence from the initial learning data conventionally included in the learning data, thereby effectively reducing the prediction error of the second learning model. can. Thereby, even if the prediction error of the second learning model is reduced for some reason, good prediction accuracy can be obtained by reconstructing the second learning model.

By reconstructing the second learning model with improved prediction accuracy in the second learning model construction unit 39, the control device 15 can calculate the absorbent slurry supply amount based on the prediction value of the second learning model. Control target values can be set with high accuracy. As a result, the absorbent slurry supply controller 37 can suitably control the absorbent slurry supply amount by controlling the absorbent slurry supply amount control valve 23 based on the control target value.

Although CaCO ₃ is used as the SO ₂ absorbent in the above embodiment, it is not limited to CaCO ₃ . As an SO ₂ absorbent, for example, magnesium hydroxide (Mg(OH) ₂ ) or the like can be used.

The information processing device 52 shown in FIG. 12 as a device for executing each process in the control device 15 may be connected to the edge server 42 in a cloud environment or via a VPN so as to be electrically communicable. It is possible. In this case, the information processing device 52 includes the operation data receiving unit 30, the first relationship table creation unit 31, the circulation flow rate determination unit 32, the second relationship table creation unit 35, the absorbent slurry supply amount determination unit 36, the first learning model A construction unit 38, a second learning model construction unit 39, a prediction error calculation unit 40, and an additional learning data selection unit 42 are provided, and the control target value determined by the circulation flow rate determination unit 32 and the absorbent slurry supply amount determination unit 36 is controlled. By communicating with the circulation pump adjustment unit 33 and the absorbent slurry supply control unit 37 in the apparatus 15, the circulation pump and the supply amount of the absorbent may be controlled.
Further, the operation data receiving unit 30 may receive various kinds of operation data via the operation data relay unit 43 of the control device 15, or may receive various kinds of operation data from the operation data acquisition unit 20 as described above. good too.

In particular, when calculating in a cloud environment, from the viewpoint of security, there are cases where the control target values of the circulating pump and absorbent are not directly controlled, but only displayed. For example, there is a case where a driving index map generated in a cloud environment is transmitted and illustrated to a device (terminal 54) owned by the customer through a dedicated application, and the local driving index map is updated by the customer.
On the other hand, the information processing device 52 may also include a circulation pump control section 33 and an absorbent slurry supply control section 37 to remotely control the circulation pump and the supply amount of absorbent.
Furthermore, the information processing device 52 may be configured to execute each process in the information processing device 52 in response to a request from the terminal 54 .

 　In addition, it is possible to appropriately replace the components in the above-described embodiments with known components within the scope of the present disclosure, and the above-described embodiments may be combined as appropriate.

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

(1) A device according to one embodiment comprises:
A device for executing processing related to plant control based on prediction results using a learning model,
Additional learning data selection for selecting, as additional learning data, data having a large deviation from the learning data used to construct the learning model from the driving data when the prediction result using the learning model satisfies a predetermined condition. Department and
a learning model construction unit for reconstructing the learning model using new learning data including the learning data and the additional learning data;
Prepare.

According to the aspect (1) above, the additional learning data is selected and added to the learning data for building the learning model to create new learning data, and the learning model is reconstructed using the learning data. is done. At this time, by selecting additional learning data to be added to the new learning data so as to include data with a large degree of divergence from the learning data, the learning model can be reconstructed appropriately. Good control accuracy can be obtained by executing processing related to plant control based on the prediction result of the learning model appropriately reconstructed in this way.

(2) In another aspect, in the aspect of (1) above,
When a prediction result using the reconstructed new learning model satisfies a predetermined condition, the additional learning data selection unit selects data having a large divergence degree from the driving data not selected as the additional learning data. and the learning data constructing unit reconstructs the learning model using the new learning data including the additional learning data further selected by the additional learning data selecting unit to implement.

According to the aspect (2) above, when the prediction result using the reconstructed new learning model satisfies a predetermined condition, new additional learning data is added to the learning data to create new learning data. Then, the learning model is reconstructed again using the new learning data. By repeating the selection of additional learning data and the reconstruction of the learning model, the prediction error of the learning model can be sufficiently reduced.

(3) In another aspect, in the above aspect (1) or (2),
The additional learning data selection unit selects, as the additional learning data, average values of parameters included in the driving data over a predetermined period.

According to the aspect (3) above, by using the average value for the predetermined period as the additional learning parameter, it is possible to effectively reduce the amount of calculation when reconstructing the learning model while ensuring the learning accuracy.

(4) In another aspect, in any one aspect of (1) to (3) above,
The learning model construction unit reconstructs the learning model when the prediction result continuously satisfies the predetermined condition for a predetermined time or longer.

According to the aspect (4) above, the determination of whether or not the prediction result satisfies the predetermined condition is made based on whether or not the prediction result continuously satisfies the predetermined condition for a predetermined period of time. The prediction results may also vary depending on the operating state of the plant, and if short-term judgments are made, the learning model will be rebuilt frequently, and there is a risk that the burden of model management will increase. Therefore, by performing continuous determination over a predetermined period of time as in this aspect, the learning model can be appropriately reconstructed, and efficient model management becomes possible.

(5) In another aspect, in any one aspect of (1) to (4) above,
The learning data is data before construction of the learning model or data used for the previous construction.

According to the above aspect (5), the new learning data created by adding the additional learning data to the data before building the learning model or the learning data used for the previous building of the learning model is used. , a reconstruction of the learning model is performed.

(6) In another aspect, in any one aspect of (1) to (5) above,
The additional learning data selection unit selects the additional learning data from the operation data acquired during steady operation of the plant.

According to the aspect (6) above, the selection of the additional learning data is performed on the operation data acquired during the steady operation of the plant. For example, the operation data acquired during non-steady operation such as when an abnormality occurs in the plant, when the operation is started, when the operation is stopped, etc., is excluded from the selection of additional learning data, so that the prediction results of the learning model can be obtained appropriately. can be done. If the operating data includes data obtained during unsteady operation, the operating data may be excluded by performing preprocessing.

(7) In another aspect, in any one aspect of (1) to (6) above,
When the prediction result using the learning model satisfies a predetermined condition, it indicates when the prediction error of the predicted value obtained using the learning model satisfies the threshold.

According to the aspect (7) above, the determination of whether or not the prediction result satisfies the predetermined condition is made based on whether or not the prediction error of the predicted value obtained using the learning model satisfies the threshold. As a result, even if the prediction error of the learning model decreases for some reason, the learning model can be reconstructed using new learning data including the efficiently selected additional learning data, and good prediction accuracy can be achieved. can get. Good control accuracy can be obtained by executing processing related to plant control based on the predicted values of the learning model whose prediction accuracy has been improved in this way.

(8) In another aspect, in the aspect of (7) above,
The additional learning data selection unit selects a parameter to be included in the additional learning data from the driving data based on the degree of contribution to the predicted value.

According to the above aspect (8), by including some parameters selected from the operation data in the additional learning data, while ensuring the learning accuracy, the amount of calculation when reconstructing the learning model is effectively reduced. can.

(9) In another aspect, in the aspect of (7) or (8) above,
The plant is a wet-type flue gas desulfurization apparatus that performs desulfurization by bringing the exhaust gas generated in the combustion device and the absorbent circulated in the absorption tower into gas-liquid contact,
The predicted value is the sulfur dioxide concentration of the flue gas at the outlet of the absorber tower.

According to the above aspect (9), the learning model for predicting the sulfur dioxide concentration of the flue gas at the outlet of the absorption tower of the wet flue gas desulfurization system is rebuilt when the prediction error becomes larger than a predetermined value. By doing so, it is possible to suitably ensure the prediction accuracy of the learning model.

(10) In another aspect, in the aspect of (9) above,
A control target value for the circulation flow rate of the absorbent is determined based on the predicted value calculated by the learning model.

According to the above aspect (10), the prediction value is calculated using the learning model whose prediction error has been reduced by reconstruction, and the control target value for the circulation amount of the absorbent is determined based on the prediction value. , good control accuracy is obtained.

(11) In another aspect, in the above aspect (9) or (10),
A control target value for the amount of absorbent supplied to the absorber is determined based on the predicted value calculated by the learning model.

According to the above aspect (11), by calculating a predicted value using a learning model whose prediction error has been reduced by reconstruction, and determining the control target value of the absorbent supply amount based on the predicted value, Good control accuracy is obtained.

(12) A remote monitoring system according to one aspect,
A remote monitoring system comprising a terminal capable of communicating with a device for executing processing related to plant control based on prediction results using a learning model,
The device comprises:
In response to a request from the terminal, when the prediction result using the learning model satisfies a predetermined condition, data with a large deviation from the learning data used to construct the learning model from the driving data is selected as additional learning data. an additional learning data selection unit for
a learning model construction unit for reconstructing the learning model using learning data including the learning data and the additional learning data;
Prepare.

According to the above aspect (12), new learning data is created by selecting and adding additional learning data to learning data for building a learning model, and the learning model is rebuilt using the learning data. is done. At this time, by selecting additional learning data to be added to the new learning data so as to include data with a large degree of divergence from the learning data, the learning model can be reconstructed appropriately. Good control accuracy can be obtained by executing processing related to plant control based on the prediction result of the learning model appropriately reconstructed in this way.

(13) A method of controlling a device according to one aspect includes:
A control method for a device for executing processing related to plant control based on prediction results using a learning model,
an additional learning data selection step of selecting, as additional learning data, data having a large deviation from the learning data used to construct the learning model from the driving data when the prediction result using the learning model satisfies a predetermined condition; ,
a learning model building step of reconstructing the learning model using learning data including the learning data and the additional learning data;
Prepare.

According to the aspect (13) above, the additional learning data is selected and added to the learning data for building the learning model to create new learning data, and the learning model is reconstructed using the learning data. is done. At this time, by selecting additional learning data to be added to the new learning data so as to include data with a large degree of divergence from the learning data, the learning model can be reconstructed appropriately. Good control accuracy can be obtained by executing processing related to plant control based on the prediction result of the learning model appropriately reconstructed in this way.

(14) A method for controlling a remote monitoring system according to one aspect includes:
A control method for a remote monitoring system comprising a terminal capable of communicating with a device for executing processing related to plant control based on prediction results using a learning model,
In response to a request from the terminal, when the prediction result using the learning model satisfies a predetermined condition, data with a large deviation from the learning data used to construct the learning model from the driving data is selected as additional learning data. an additional learning data selection step for
a learning model building step of reconstructing the learning model using learning data including the learning data and the additional learning data;
Prepare.

According to the above aspect (14), new learning data is created by selecting and adding additional learning data to learning data for building a learning model, and the learning model is rebuilt using the learning data. is done. At this time, by selecting additional learning data to be added to the new learning data so as to include data with a large degree of divergence from the learning data, the learning model can be reconstructed appropriately. Good control accuracy can be obtained by executing processing related to plant control based on the prediction result of the learning model appropriately reconstructed in this way.

1 Combustion device 2 Pipe 3 Circulation pipe 5 Generator 10 Wet flue gas desulfurization device 11 Absorption tower 12 Circulation pump 13 Absorbent slurry supply unit 14 Gypsum recovery unit 15 Control device 16 Outflow pipe 17 Gas analyzer 20 Operation data acquisition unit 21 Absorbent slurry production facility 22 Absorbent slurry supply pipe 23 Absorbent slurry supply amount control valve 25 Gypsum separator 26 Gypsum slurry extraction pipe 27 Gypsum slurry extraction pump 30 Operation data receiving unit 31 First relationship table creation unit 32 Circulation Flow rate determination unit 33 Circulation pump adjustment unit 35 Second relationship table creation unit 36 Absorbent slurry supply amount determination unit 37 Absorbent slurry supply control unit 38 First learning model construction unit 39 Second learning model construction unit 40 Prediction error calculation unit 42 Additional learning data selection unit 43 Operation data relay unit 44 Remote monitoring system 48 Edge server 50 Remote monitoring device 52 Information processing device 54 Terminal 55 Information processing system

Claims (14)

1. A device for executing processing related to plant control based on prediction results using a learning model,
   Additional learning data selection for selecting, as additional learning data, data having a large deviation from the learning data used to construct the learning model from the driving data when the prediction result using the learning model satisfies a predetermined condition. Department and
   a learning model construction unit for reconstructing the learning model using new learning data including the learning data and the additional learning data;
   A device comprising:
2. When a prediction result using the reconstructed new learning model satisfies a predetermined condition, the additional learning data selection unit selects data having a large divergence degree from the driving data not selected as the additional learning data. and the learning data constructing unit reconstructs the learning model using the new learning data including the additional learning data further selected by the additional learning data selecting unit 11. The apparatus of claim 1, wherein the apparatus performs
3. The device according to claim 1 or 2, wherein the additional learning data selection unit selects, as the additional learning data, an average value of parameters included in the operation data over a predetermined period.
4. The apparatus according to any one of claims 1 to 3, wherein the learning model construction unit reconstructs the learning model when the prediction result continuously satisfies the predetermined condition for a predetermined time or longer.
5. 5. The apparatus according to any one of claims 1 to 4, wherein the learning data is data before building the learning model or data used for building the learning model last time.
6. The apparatus according to any one of claims 1 to 5, wherein the additional learning data selection unit selects the additional learning data from the operation data acquired during steady operation of the plant.
7. 7. The method according to any one of claims 1 to 6, wherein when a prediction result using said learning model satisfies a predetermined condition indicates when a prediction error of a predicted value obtained using said learning model satisfies a threshold. device.
8. The apparatus according to claim 7, wherein the additional learning data selection unit selects parameters to be included in the additional learning data from the driving data based on the degree of contribution to the predicted value.
9. The plant is a wet-type flue gas desulfurization apparatus that performs desulfurization by bringing the exhaust gas generated in the combustion device and the absorbent circulated in the absorption tower into gas-liquid contact,
   9. Apparatus according to claim 7 or 8, wherein the predicted value is the sulfur dioxide concentration of the flue gas at the outlet of the absorber tower.
10. The device according to claim 9, wherein a control target value for the circulation flow rate of the absorbent is determined based on the predicted value calculated by the learning model.
11. 11. The apparatus according to claim 9 or 10, wherein a control target value for the amount of absorbent supplied to said absorption tower is determined based on said predicted value calculated by said learning model.
12. A remote monitoring system comprising a terminal capable of communicating with a device for executing processing related to plant control based on prediction results using a learning model,
    The device comprises:
    In response to a request from the terminal, when the prediction result using the learning model satisfies a predetermined condition, data with a large deviation from the learning data used to construct the learning model from the driving data is selected as additional learning data. an additional learning data selection unit for
    a learning model construction unit for reconstructing the learning model using learning data including the learning data and the additional learning data;
    A remote monitoring system.
13. A control method for a device for executing processing related to plant control based on prediction results using a learning model,
    an additional learning data selection step of selecting, as additional learning data, data having a large deviation from the learning data used to construct the learning model from the driving data when the prediction result using the learning model satisfies a predetermined condition; ,
    a learning model building step of reconstructing the learning model using learning data including the learning data and the additional learning data;
    A device control method comprising:
14. A control method for a remote monitoring system comprising a terminal capable of communicating with a device for executing processing related to plant control based on prediction results using a learning model,
    In response to a request from the terminal, when the prediction result using the learning model satisfies a predetermined condition, data with a large deviation from the learning data used to construct the learning model from the driving data is selected as additional learning data. an additional learning data selection step for
    a learning model building step of reconstructing the learning model using learning data including the learning data and the additional learning data;
    A method of controlling a remote monitoring system, comprising:

Images