Prediction device, learning device, prediction method, and prediction program

Abstract

To improve the accuracy of predictions when forecasting plant process values. [Solution] The prediction device stores a physical model for performing simulations of the behavior of a plant's processing steps, which includes: an adjusted physical model whose parameters are adjusted based on a first measurement taken in the processing step; an unobservable state variable in the processing step calculated by operating the adjusted physical model based on a second measurement taken in the processing step that is different from the first measurement; and a trained prediction model constructed by training the prediction model using a third measurement taken in the processing step that is different from the second measurement. The device operates the adjusted physical model, calculates the state variable based on the second measurement taken in the processing step, operates the trained prediction model, and predicts the process values ​​of the plant based on the calculated state variable and the third measurement taken in the processing step.

Description

Technical Field

[0001] The present disclosure relates to a prediction device, a learning device, a prediction method, and a prediction program.

Background Art

[0002] There is known a prediction technique for predicting a process value of an incinerator based on image data (measurement values) obtained by photographing the inside of the incinerator.

Prior Art Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the image data (measurement values) obtained by photographing the inside of the incinerator only fragmentarily represents the behavior of the incineration process of the incinerator. Therefore, it is difficult to accurately predict the process value of the incinerator.

[0005] An object of the present disclosure is to improve the prediction accuracy when predicting the process value of a plant.

Means for Solving the Problems

[0006] The prediction device according to the first aspect of the present disclosure is a physical model that executes a simulation of the behavior of a process of a plant, the adjusted physical model in which parameters are adjusted by a first measurement value measured in the process, and ​The system stores: an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step; and a trained predictive model constructed by training the predictive model using a third measurement value different from the second measurement value measured in the processing step, which predicts the process values ​​of the plant. The adjusted physical model is operated to calculate state variables based on the second measurement taken in the processing step, and the trained prediction model is operated to predict the process values ​​of the plant based on the calculated state variables and the third measurement taken in the processing step.

[0007] A second aspect of this disclosure is a prediction device relating to the first aspect, The state variables calculated by running the adjusted physical model and the process values ​​predicted by running the trained prediction model are visualized.

[0008] A third aspect of this disclosure is a prediction device relating to the first or second aspect, The aforementioned adjusted physical model is The first measurement value measured in the processing step is input to the physical model, and the output value calculated by the physical model is generated by adjusting the parameters of the physical model using a Kalman filter so that the output value calculated by the physical model approaches the output value measured in the processing step.

[0009] A fourth aspect of this disclosure is a prediction device relating to any of the first to third aspects, The aforementioned trained predictive model is The predictive model is constructed by using as input data an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, and using the process values ​​measured in the processing step of the plant as ground truth data.

[0010] A fifth aspect of this disclosure is a prediction device relating to any of the first to fourth aspects, The control values ​​of the plant are searched so that the process values ​​predicted by running the trained prediction model approach the newly set target values.

[0011] A sixth aspect of this disclosure is a prediction device relating to any of the first to fifth aspects, The aforementioned physical model is a grate combustion model, a gas-phase combustion model, or a boiler heat acquisition and power generation model.

[0012] The prediction device relating to the seventh aspect of this disclosure is A physical model for simulating the behavior of a waste incinerator during the incineration process, specifically when the waste moves on a movable grate and undergoes the processes of drying, thermal decomposition, and combustion in sequence to achieve complete combustion, comprising an adjusted grate combustion model, which is an adjusted physical model whose parameters are adjusted based on first measured values ​​taken during the incineration process, and A physical model for simulating the behavior of combustible gas generated by combustion in the incineration process of the incinerator when it is combusted by supplied air, comprising: an adjusted gas-phase combustion model, which is an adjusted physical model whose parameters have been adjusted based on first measured values ​​taken in the incineration process; A physical model for simulating the behavior of the incineration process of the aforementioned incinerator, specifically when steam generated by the absorption of heat from exhaust gas in a waste heat boiler rotates a steam turbine and generates electricity, wherein the adjusted boiler heat recovery / power generation model is an adjusted physical model whose parameters have been adjusted based on first measured values ​​taken in the incineration process, The system stores a trained predictive model that predicts the process values ​​of the incinerator, which is constructed by training the predictive model using an unobservable state variable in the incineration process, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the incineration process of the incinerator, and a third measurement value different from the second measurement value measured in the incineration process. The adjusted physical model is operated to calculate state variables based on second measurements taken in the incineration process, and the trained predictive model is operated to predict the process values ​​of the incinerator based on the calculated state variables and third measurements taken in the incineration process.

[0013] The learning device relating to the eighth aspect of this disclosure is An adjustment unit that adjusts the parameters of a physical model used to simulate the behavior of a plant's processing steps using first measured values ​​taken during the processing steps, and generates an adjusted physical model. The system includes a learning unit that trains a predictive model using an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, and constructs a trained predictive model.

[0014] The ninth aspect of this disclosure is a learning device relating to the eighth aspect, The adjustment unit is, The adjusted physical model is generated by inputting the first measurement value measured in the processing step into the physical model and adjusting the parameters of the physical model using a Kalman filter so that the output value calculated by the physical model approaches the output value measured in the processing step.

[0015] A tenth aspect of this disclosure is a learning device relating to the eighth or ninth aspect, The aforementioned learning unit, The learned prediction model is constructed by using, as input data, an unobservable state variable in the processing step and a third measurement value different from the second measurement value, which are calculated by operating the adjusted physical model based on the second measurement value different from the first measurement value measured in the processing step, and performing learning of the prediction model with a process value measured in the processing step of the plant as correct data.

[0016] The prediction method according to the eleventh aspect of the present disclosure is An adjusted physical model that is a physical model for simulating the behavior of a processing step of a plant and whose parameters are adjusted by a first measurement value measured in the processing step, and A learned prediction model that predicts a process value of the plant, which is constructed by performing learning of the prediction model using an unobservable state variable in the processing step and a third measurement value different from the second measurement value, which are calculated by operating the adjusted physical model based on the second measurement value different from the first measurement value measured in the processing step. The prediction method of a prediction device that stores the above is The computer of the prediction device executes a process of operating the adjusted physical model, calculating a state variable based on the second measurement value measured in the processing step, operating the learned prediction model, and predicting the process value of the plant based on the calculated state variable and the third measurement value measured in the processing step.

[0017] The prediction program according to the twelfth aspect of the present disclosure is An adjusted physical model that is a physical model for simulating the behavior of a processing step of a plant and whose parameters are adjusted by a first measurement value measured in the processing step, and The computer of the prediction device stores a learned prediction model for predicting the process value of the plant, which is constructed by operating a prediction model using an unobservable state variable in the processing step, which is calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step. Operate the adjusted physical model, calculate a state variable based on a second measurement value measured in the processing step, operate the learned prediction model, and predict the process value of the plant based on the calculated state variable and a third measurement value measured in the processing step.

Advantages of the Invention

[0018] According to the present disclosure, the prediction accuracy when predicting the process value of a plant can be improved.

Brief Description of the Drawings

[0019] [Figure 1] It is a diagram showing an example of the system configuration in the learning phase of the target plant. [Figure 2] It is a diagram showing an example of the hardware configuration of the learning device. [Figure 3A] It is a diagram showing an example of the configuration of an incinerator. [Figure 3B] It is a diagram showing an example of a plurality of physical models for performing a simulation of the behavior of the incineration process of an incinerator. [Figure 4] It is a diagram showing an example of the functional configuration of the learning device in the parameter adjustment phase. [Figure 5] It is a first diagram for explaining the outline of the parameter adjustment process. [Figure 6] It is a second diagram for explaining the outline of the parameter adjustment process. [Figure 7] It is a third diagram for explaining the outline of the parameter adjustment process. [Figure 8]This figure shows an example of the functional configuration of the learning device during the predictive model construction phase. [Figure 9] This figure shows a specific example of the learning process performed by a learning device during the predictive model construction phase. [Figure 10] This is a flowchart showing the processing flow during the learning phase. [Figure 11] Figure 1 shows an example of the system configuration during the prediction phase of the target plant. [Figure 12] This figure shows an example of the hardware configuration of a prediction device. [Figure 13] The first figure shows an example of the functional configuration of the prediction device during the prediction phase. [Figure 14] This figure shows a specific example of prediction processing by a prediction device during the prediction phase. [Figure 15] This figure shows an example of a monitoring screen for an incinerator. [Figure 16] This is the first flowchart illustrating the processing flow during the prediction phase. [Figure 17] The second figure shows an example of the system configuration during the prediction phase of the target plant. [Figure 18] The second figure shows an example of the functional configuration of the prediction device during the prediction phase. [Figure 19] This is the second flowchart illustrating the processing flow during the prediction phase. [Modes for carrying out the invention]

[0020] Each embodiment will be described below with reference to the attached drawings. In this specification and the drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant descriptions will be omitted.

[0021] [First Embodiment] <System configuration during the learning phase of the target plant> First, we will describe the system configuration in the learning phase of a target plant to which the learning device according to the first embodiment is applied. Figure 1 is a diagram showing an example of the system configuration in the learning phase of a target plant.

[0022] As shown in Figure 1, the target plant 10 to which the learning device 120 according to the first embodiment is applied comprises the target equipment 100 and the plant control system 110. In the first embodiment, the target plant 10 is, for example, a general waste treatment plant intended for waste disposal.

[0023] The target equipment 100 is, for example, an incinerator, which transmits measured values ​​taken during operation to the plant control system 110 and performs incineration based on control values ​​transmitted from the plant control system 110.

[0024] The plant control system 110 includes a monitoring and control unit 111. The monitoring and control unit 111 monitors the target equipment 100 in operation based on measurements transmitted from the target equipment 100. The monitoring and control unit 111 also calculates control values ​​to operate the target equipment 100 so that the process values ​​used to control the processing process (in the case of the target equipment 100 being an incinerator, the incineration process) from the measurements transmitted from the target equipment 100 match the target values. Furthermore, the monitoring and control unit 111 transmits the calculated control values ​​to the target equipment 100.

[0025] The learning device 120 generates a physical model for simulating the behavior of the processing steps of the target equipment 100. The learning device 120 acquires measurement values ​​taken during the processing steps of the target equipment 100 from the plant control system 110, and generates an adjusted physical model by adjusting the parameters using the acquired measurement values. The phase in which the learning device 120 adjusts the parameters of the physical model and generates an adjusted physical model is called the "parameter adjustment phase."

[0026] Furthermore, the learning device 120 generates a predictive model that predicts process values ​​used to control the processing steps of the target equipment 100. The learning device 120 constructs a trained predictive model by training the predictive model using unobservable state variables calculated by the adjusted physical model and measured values ​​obtained from the plant control system 110. The phase in which the learning device 120 trains the predictive model to construct a trained predictive model is referred to as the "predictive model construction phase."

[0027] If the target equipment 100 is an incinerator, the behavior of the incineration process is simulated using multiple physical models (for example, a grate combustion model, a gas-phase combustion model, a boiler heat absorption / power generation model, etc. Details will be described later).

[0028] Thus, the learning device 120 is • To simulate the behavior of the processing steps of the target equipment 100, multiple physical models are generated. • The generated physical models are adjusted using measurements taken during the processing steps of the target equipment 100.

[0029] As a result, the learning device 120 can generate multiple adjusted physical models that can perform highly accurate simulations of the behavior of the processing steps of the target equipment 100.

[0030] Furthermore, the learning device 120 is By performing simulations of the processing behavior of the target equipment 100 using multiple adjusted physical models, highly accurate unobservable state variables are calculated. In addition to the measured values ​​taken in the processing steps of the target equipment 100, a predictive model that predicts process values ​​is trained by using highly accurate, unobservable state variables.

[0031] As a result, the learning device 120 can construct a trained predictive model that can predict process values ​​with higher accuracy compared to a trained predictive model that was trained based only on measurements taken in the processing steps of the target equipment 100.

[0032] In other words, according to the first embodiment, it is possible to improve the prediction accuracy when predicting process values ​​used to control the processing steps of the target equipment 100.

[0033] <Hardware configuration of the learning device> Next, the hardware configuration of the learning device 120 will be described. Figure 2 shows an example of the hardware configuration of the learning device.

[0034] As shown in Figure 2, the learning device 120 includes a processor 201, memory 202, auxiliary storage device 203, connection device 204, communication device 205, and drive device 206. The processor 201, memory 202, auxiliary storage device 203, connection device 204, communication device 205, and drive device 206 of the learning device 120 are interconnected via a bus 207.

[0035] The processor 201 has various computing devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 201 reads various programs (for example, learning programs, etc.) into memory 202 and executes them.

[0036] Memory 202 has main memory devices such as ROM (Read Only Memory) and RAM (Random Access Memory). The processor 201 and memory 202 form a so-called computer, and the computer realizes various functions by having the processor 201 execute various programs read from memory 202.

[0037] The auxiliary storage device 203 stores various programs and various information used when those programs are executed by the processor 201.

[0038] The connection device 204 is a device for connecting the learning device 120 to an example of an external device, such as the operating device 211 and the display device 212.

[0039] The communication device 205 is a device for communicating with various devices (for example, the plant control system 110) via a network.

[0040] The drive device 206 is a device for setting the recording medium 213. The recording medium 213 here includes media that record information optically, electrically, or magnetically, such as CD-ROMs, flexible disks, and magneto-optical disks. The recording medium 213 may also include semiconductor memory that records information electrically, such as ROMs and flash memory.

[0041] The various programs to be installed on the auxiliary storage device 203 are installed, for example, when the distributed recording medium 213 is set in the drive device 206 and the various programs recorded on the recording medium 213 are read. Alternatively, the various programs to be installed on the auxiliary storage device 203 may be installed by downloading them from the network via the communication device 205.

[0042] <Outline configuration of the incinerator> Next, we will describe the general configuration of an incinerator, which is an example of the target equipment 100. Figure 3A is a diagram showing an example of the configuration of an incinerator, and it shows an example of the configuration of an incinerator called a stoker-type incinerator.

[0043] As shown in Figure 3A, in a stoker-type incinerator, the waste that is fed in moves on a movable grate. Air is supplied to the movable grate, and the waste moving on the grate undergoes complete combustion by sequentially going through the processes of drying, thermal decomposition, and combustion.

[0044] The incinerated ash generated by combustion is sent to an ash pit (not shown). The exhaust gas generated by combustion (including exhaust gas generated from the combustion of combustible gases) is absorbed in a waste heat boiler and then sent to a filtration-type dust collector (not shown). The steam generated by the heat absorption in the waste heat boiler is sent to a steam turbine and used for power generation.

[0045] In this embodiment, the following three physical models are used to simulate the behavior of the incineration process of the stoker-type incinerator shown in Figure 3A. • Grate combustion model: A physical model that simulates the behavior of a completely combusted waste material in the region indicated by reference numeral 310, where the waste material moves along a movable grate and undergoes the processes of drying, thermal decomposition, and combustion in sequence. • Gas-phase combustion model: A physical model that simulates the behavior of combustible gas generated by combustion in the region indicated by reference numeral 320, when it is combusted by supplied air. • Boiler heat recovery / power generation model: A physical model that simulates the behavior when a steam turbine rotates and generates electricity using steam generated by heat absorption in a waste heat boiler from exhaust gas in the region indicated by reference numeral 330.

[0046] Each physical model is represented by a predetermined relational expression, which calculates the desired information by performing the simulations described above. Specific examples include the relational expression disclosed in Japanese Patent Application No. 2023-182295, or the relational expression disclosed in Japanese Patent Application No. 2023-182691.

[0047] A grate combustion model is composed of one, two, or all of the following models: a water evaporation model, a volatile matter release model, and a fixed carbon combustion model, and is constructed according to the type of information to be calculated. This type of information includes, for example, the amount of water evaporation (amount of water vapor), the amount of combustible gas generated, and the amount of fixed carbon combustion. However, the models that constitute a grate combustion model are not limited to these.

[0048] The same applies to the gas-phase combustion model and the boiler heat recovery / power generation model, which are represented by predetermined relational equations that allow the desired information to be calculated by performing the simulations described above. These predetermined relational equations include, but are not limited to, those that calculate information such as the amount of combustible gas burned, the amount of heat recovered by the boiler, and the amount of power generated.

[0049] <Overview of the physical model> Next, we will describe the outlines of the three physical models mentioned above. Figure 3B shows an example of multiple physical models used to simulate the behavior of the incineration process in an incinerator.

[0050] As shown in Figure 3B, the grate combustion model 310M includes: Examples of measurements taken during the incineration process of an incinerator include the composition of the waste being fed in, the amount of waste being fed in, the amount of air supplied, the movement of the movable grate, etc. The amount of combustible gas burned, etc., was calculated by running a simulation using the gas-phase combustion model 320M. The following is entered.

[0051] The grate combustion model 310M performs the simulation based on this input information. This allows the grate combustion model 310M to calculate observable and unobservable state variables in the incinerator. Figure 3B shows an example where the unobservable state variables are: • Composition of waste put in, • Distribution of waste dropped into the furnace, • Weight of waste per grate, • Waste temperature per grate, • Radiant heat from the fire, • Amount of water vapor, • Amount of combustible gas generated, • Composition of flammable gases, • Fixed carbon combustion amount, • Amount of waste moved, This shows how the calculations were performed.

[0052] In this context, observable information refers to information for which both measured values ​​obtained during the incineration process of the incinerator and calculated values ​​obtained by the grate combustion model 310M exist. Observable information includes the composition of the waste being fed in, the amount of waste being fed in, the amount of air supplied, and the movement of the movable grate, as well as, although not shown in Figure 3B, the temperature of the combustible gas, for example. On the other hand, unobservable state variables (information) refer to information for which calculated values ​​obtained by the grate combustion model 310M exist, but for which no measured values ​​are obtained during the incineration process of the incinerator.

[0053] Some of these state variables (e.g., water vapor amount, combustible gas generation amount, fixed carbon combustion amount) calculated by the grate combustion model 310M are input into the gas-phase combustion model 320M.

[0054] The 320M gas-phase combustion model includes: • Some of the state variables input by the grate combustion model 310M (e.g., water vapor amount, combustible gas generation amount, fixed carbon combustion amount), Examples of measurements taken during the incineration process of an incinerator include the amount of air supplied to burn the combustible gas generated by combustion, The following is entered.

[0055] The gas-phase combustion model 320M performs the simulation based on this input information. This allows the gas-phase combustion model 320M to calculate observable information and unobservable state variables in the incinerator. Figure 3B shows an example where unobservable state variables are: • Amount of combustible gas burned, • Composition of post-combustion gas, • Amount of heat removed from the furnace body, This shows how the calculations were performed.

[0056] In this context, observable information refers to information for which both measured values ​​obtained during the incineration process of the incinerator and calculated values ​​obtained by the gas-phase combustion model 320M exist. Observable information includes the amount of air mentioned above, as well as, although not shown in Figure 3B, combustion control temperature, gas concentration, etc. On the other hand, unobservable state variables (information) refer to information for which calculated values ​​obtained by the gas-phase combustion model 320M exist, but for which measured values ​​obtained during the incineration process of the incinerator do not exist.

[0057] Some of these state variables (e.g., the amount of combustible gas burned) calculated by the gas-phase combustion model 320M are input into the grate combustion model 310M and the boiler heat recovery / power generation model 330M.

[0058] The boiler heat recovery / power generation model 330M includes: • Some of the state variables input by the gas-phase combustion model 320M (e.g., amount of combustible gas burned), Examples of measurements taken during the incineration process of an incinerator include exhaust gas flow rate, exhaust gas temperature, etc. (not shown in Figure 3B), The following is entered.

[0059] The boiler heat recovery / power generation model 330M performs the simulation based on this input information. This allows the boiler heat recovery / power generation model 330M to calculate observable information and unobservable state variables in the incinerator. In the example in Figure 3B, the unobservable state variables are: • Heat recovery efficiency, • Amount of heat removed from the furnace body, These are calculated and, as observable information, • Boiler heat output, • Amount of power generation, This shows how the calculations were performed.

[0060] In this context, observable information refers to information for which both measured values ​​obtained during the incineration process of the incinerator and calculated values ​​obtained by the boiler heat recovery / power generation model 330M exist. In other words, observable information includes the boiler heat recovery amount, power generation amount, and the exhaust gas flow rate and exhaust gas temperature mentioned above. On the other hand, unobservable state variables (information) refer to information for which calculated values ​​obtained by the boiler heat recovery / power generation model 330M exist, but measured values ​​obtained during the incineration process of the incinerator do not exist.

[0061] <Functional configuration of the learning device (parameter tuning phase)> Next, the functional configuration of the learning device 120 in the parameter adjustment phase will be described. Figure 4 is a diagram showing an example of the functional configuration of the learning device in the parameter adjustment phase. As described above, a learning program is installed in the learning device 120, and when this program is executed, the learning device 120 functions as a measurement value acquisition unit 400 and a parameter adjustment unit 410 in the parameter adjustment phase.

[0062] The measurement value acquisition unit 400 acquires measurement values ​​measured during the incineration process of the incinerator from the plant control system 110 and stores the acquired measurement values ​​in the measurement value storage unit 450.

[0063] The parameter adjustment unit 410 includes a grate combustion model 310M, an adjustment unit 420, a gas phase combustion model 320M, an adjustment unit 430, a boiler heat recovery / power generation model 330M, and an adjustment unit 440.

[0064] The adjustment unit 420 adjusts the parameters of the grate combustion model 310M based on the observable output values ​​calculated by the grate combustion model 310M running a simulation and the measured values ​​stored in the measured value storage unit 450.

[0065] The adjustment unit 430 adjusts the parameters of the gas-phase combustion model 320M based on the observable output values ​​calculated by the gas-phase combustion model 320M running a simulation and the measured values ​​stored in the measured value storage unit 450.

[0066] The adjustment unit 440 adjusts the parameters of the boiler heat recovery / power generation model 330M based on the observable output values ​​calculated by the boiler heat recovery / power generation model 330M running a simulation and the measured values ​​stored in the measured value storage unit 450.

[0067] <Details of parameter adjustment processing by the learning device's parameter adjustment unit> Next, we will explain in detail the parameter adjustment process performed by the parameter adjustment unit 410 of the learning device 120 during the parameter adjustment phase.

[0068] (1) Details of the grate combustion model and parameter adjustment process by the adjustment unit First, we will explain the details of the parameter adjustment process by the grate combustion model 310M and the adjustment unit 420, which are included in the parameter adjustment unit 410. Figure 5 is the first diagram illustrating the overview of the parameter adjustment process.

[0069] As shown in Figure 5, among the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​necessary for running the simulation by the grate combustion model 310M (for example, the composition of the waste to be fed in and the amount of waste to be fed in) are input to the grate combustion model 310M as input values ​​(an example of the first measurement value). As a result, the grate combustion model 310M runs the simulation under default parameters. At this time, the grate combustion model 310M outputs to the adjustment unit 420 an observable output value (calculated value of water evaporation) calculated by running the simulation, which corresponds to the measurement values ​​other than those used as input values ​​among the measurement values ​​stored in the measurement value storage unit 450 (for example, the calculated value of water evaporation, which is an output value corresponding to measurement values ​​other than the composition of the waste to be fed in and the amount of waste to be fed in).

[0070] Furthermore, as shown in Figure 5, among the measurement values ​​stored in the measurement value storage unit 450, • Measurement values ​​other than those used as input values ​​(for example, measurement values ​​other than the composition of the waste being put in, the amount of waste being put in, such as the amount of water evaporation), • Measured values ​​(measured water evaporation amounts) corresponding to the observable output values ​​(calculated water evaporation amounts) calculated by the grate combustion model 310M, This is input to the adjustment unit 420 as an output value (an example of the first measured value). Note that measurement values ​​other than those used as input values ​​refer to measurement values ​​measured during the incineration process of the incinerator for which there are calculated values ​​calculated by the grate combustion model 310M.

[0071] The adjustment unit 420 includes a Kalman filter 421. The Kalman filter 421 is • Observable output values ​​(calculated values ​​of water evaporation) calculated using the grate combustion model 310M, • Of the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​other than the measurement value used as the input value (measurement value of water evaporation amount) The parameters of the grate combustion model 310M are updated to match. This generates the adjusted grate combustion model 310M'.

[0072] Although not shown in Figure 5, when the grate combustion model 310M performs a simulation, state variables (e.g., amount of combustible gas burned) are input in addition to the first measured value. However, at the stage of adjusting the parameters of the grate combustion model 310M, the state variables calculated by the adjusted grate combustion model 310M' have not yet been calculated. Therefore, it is assumed here that predetermined values ​​for the state variables are input to the grate combustion model 310M.

[0073] (2) Details of the gas-phase combustion model and parameter adjustment process by the adjustment unit Next, the details of the parameter adjustment process by the gas-phase combustion model 320M and the adjustment unit 430, which are included in the parameter adjustment unit 410, will be explained. Figure 6 is a second diagram illustrating the overview of the parameter adjustment process.

[0074] As shown in Figure 6, among the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​necessary for running the simulation by the gas-phase combustion model 320M (for example, the measurement value of the amount of air) are input to the gas-phase combustion model 320M as input values ​​(an example of the first measurement value). In addition, state variables (for example, the amount of water evaporation (amount of water vapor)) calculated by running the simulation with the adjusted grate combustion model 310M' are input to the gas-phase combustion model 320M. As a result, the gas-phase combustion model 320M runs the simulation under default parameters. At this time, the gas-phase combustion model 320M outputs to the adjustment unit 430 the observable output values ​​(calculated values ​​of combustion control temperature) calculated by running the simulation, which are output values ​​corresponding to measurement values ​​other than those used as input values ​​among the measurement values ​​stored in the measurement value storage unit 450 (for example, calculated values ​​of combustion control temperature, which are output values ​​corresponding to measurement values ​​other than the amount of air).

[0075] Furthermore, as shown in Figure 6, among the measurement values ​​stored in the measurement value storage unit 450, • Measurement values ​​other than those used as input values ​​(for example, measurement values ​​of combustion control temperature other than air volume), • Measured values ​​(measured combustion control temperature) corresponding to the observable output value (calculated value of combustion control temperature) calculated by the gas-phase combustion model 320M, This is input to the adjustment unit 430 as an output value (an example of the first measured value). Note that measurement values ​​other than those used as input values ​​refer to measurement values ​​measured during the incineration process of the incinerator for which there are calculated values ​​calculated by the gas-phase combustion model 320M.

[0076] The adjustment unit 430 includes a Kalman filter 431. The Kalman filter 431 is • Observable output values ​​(calculated values ​​of combustion control temperature) calculated using the gas-phase combustion model 320M, • Of the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​other than those used as input values ​​(measurement values ​​of combustion control temperature) are included, The parameters of gas-phase combustion model 320M are updated to match. This generates the adjusted gas-phase combustion model 320M'.

[0077] (3) Details of the boiler heat recovery / power generation model and parameter adjustment process by the adjustment unit Next, we will explain the details of the parameter adjustment process by the boiler heat recovery / power generation model 330M and the adjustment unit 440, which are included in the parameter adjustment unit 410. Figure 7 is a third diagram illustrating the overview of the parameter adjustment process.

[0078] As shown in Figure 7, among the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​necessary for running the simulation by the boiler heat recovery / power generation model 330M (for example, the exhaust gas flow rate) are input to the boiler heat recovery / power generation model 330M as input values ​​(an example of the first measurement value). In addition, state variables (for example, the amount of combustible gas combustion) calculated by running the simulation with the adjusted gas-phase combustion model 320M' are input to the boiler heat recovery / power generation model 330M. As a result, the boiler heat recovery / power generation model 330M runs the simulation under default parameters. At this time, the boiler heat recovery / power generation model 330M outputs to the adjustment unit 440 the observable output values ​​(calculated boiler heat recovery values) calculated by running the simulation, which are the output values ​​corresponding to measurement values ​​other than those used as input values ​​among the measurement values ​​stored in the measurement value storage unit 450 (for example, the calculated boiler heat recovery amount, which is the output value corresponding to measurement values ​​other than the exhaust gas flow rate).

[0079] Furthermore, as shown in Figure 7, among the measurement values ​​stored in the measurement value storage unit 450, • Measurement values ​​other than those used as input values ​​(for example, measurement values ​​of boiler heat acquisition, which are measurement values ​​other than exhaust gas flow rate), • Measured values ​​(measured boiler heat acquisition values) corresponding to the observable output values ​​(calculated boiler heat acquisition values) calculated by the boiler heat acquisition / power generation model 330M. This is input to the adjustment unit 440 as an output value (an example of the first measured value). Note that measurement values ​​other than those used as input values ​​refer to measurement values ​​measured during the incineration process of the incinerator for which there are calculated values ​​calculated by the boiler heat recovery / power generation model 330M.

[0080] The adjustment unit 440 includes a Kalman filter 441. The Kalman filter 441 is • Observable output values ​​(calculated boiler heat gain) calculated using the boiler heat recovery / power generation model 330M, • Of the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​other than those used as input values ​​(measurement values ​​of boiler heat acquisition), Update the parameters of boiler heat recovery / power generation model 330M so that they match. This will generate the adjusted boiler heat recovery / power generation model 330M'.

[0081] <Functional configuration of the learning device (predictive model construction phase)> Next, the functional configuration of the learning device 120 in the prediction model construction phase will be described. Figure 8 shows an example of the functional configuration of the learning device in the prediction model construction phase. As mentioned above, a learning program is installed in the learning device 120, and when this program is executed, the learning device 120 functions as a measurement value acquisition unit 400 and a learning unit 800 in the prediction model construction phase. Of these, the measurement value acquisition unit 400 has already been explained using Figure 4, so its explanation will be omitted here.

[0082] The learning unit 800 includes a tuned grate combustion model 310M', a tuned gas-phase combustion model 320M', a tuned boiler heat acquisition / power generation model 330M', and a prediction model learning unit 810.

[0083] The adjusted grate combustion model 310M' is, • Read the measurement values ​​necessary for running the simulation from the measurement value storage unit 450. • Obtain the state variables from the previous cycle required for running the simulation from the adjusted gas-phase combustion model 320M'. The simulation is executed. This causes the tuned grate combustion model 310M' to calculate state variables. The tuned grate combustion model 310M' also outputs the state variables necessary for the tuned gas-phase combustion model 320M' to run the simulation. Furthermore, the tuned grate combustion model 310M' outputs the state variables necessary for the prediction model learning unit 810 to train the prediction model.

[0084] Note that the state variables from one cycle ago refer to the state variables calculated during the simulation execution one cycle ago, when the simulation is run at a predetermined interval.

[0085] The tuned vapor phase combustion model 320M' is, • Read the measurement values ​​necessary for running the simulation from the measurement value storage unit 450. • Obtain the state variables from the previous cycle required for running the simulation from the adjusted grate combustion model 310M'. The simulation is executed. This causes the tuned gas-phase combustion model 320M' to calculate state variables. The tuned gas-phase combustion model 320M' also outputs the state variables necessary for the tuned boiler heat recovery / power generation model 330M' to run the simulation. Furthermore, the tuned gas-phase combustion model 320M' outputs the state variables necessary for the prediction model learning unit 810 to train the prediction model to run the prediction model learning unit 810.

[0086] The tuned boiler heat recovery / power generation model 330M' is, • Read the measurement values ​​necessary for running the simulation from the measurement value storage unit 450. • Obtain the state variables from the previous cycle required for running the simulation from the adjusted gas-phase combustion model 320M'. The simulation is executed. This causes the tuned boiler heat recovery / power generation model 330M' to calculate state variables. The tuned boiler heat recovery / power generation model 330M' also outputs to the predictive model learning unit 810 the state variables from the calculated state variables that are necessary for the predictive model learning unit 810 to train the predictive model.

[0087] The prediction model learning unit 810 obtains the state variables necessary for training the prediction model from the tuned grate combustion model 310M', the tuned gas-phase combustion model 320M', and the tuned boiler heat acquisition / power generation model 330M'. The prediction model learning unit 810 also reads the measured values ​​necessary for training the prediction model from the measured value storage unit 450.

[0088] The prediction model learning unit 810 uses the acquired state variables and read-out measurement values ​​as training data to train the prediction model and construct a trained prediction model.

[0089] <Details of the learning process by the predictive model learning unit of the learning device> Next, we will describe the details of the learning process performed by the predictive model learning unit 810 of the learning device 120 during the predictive model construction phase. Figure 9 is a diagram showing a specific example of the learning process performed by the learning device during the predictive model construction phase. As shown in Figure 9, during the predictive model construction phase, the predictive model learning unit 810 has a predictive model 900 and a comparison / modification unit 910.

[0090] As shown in Figure 9, among the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​necessary for running the simulation using the adjusted grate combustion model 310M' are input to the adjusted grate combustion model 310M' as input values ​​(an example of the second measurement value). In addition, some of the unobservable state variables calculated by the adjusted gas-phase combustion model 320M' one cycle earlier are input to the adjusted grate combustion model 310M'. As a result, the adjusted grate combustion model 310M' runs the simulation and calculates the unobservable state variables.

[0091] Some of the unobservable state variables calculated by the adjusted grate combustion model 310M' are input to the prediction model 900 as input data. Some of the unobservable state variables calculated by the adjusted grate combustion model 310M' are input to the adjusted gas-phase combustion model 320M' as input values.

[0092] As shown in Figure 9, among the measurement values ​​stored in the measurement value storage unit 450, the measurement values ​​necessary for running the simulation using the adjusted gas-phase combustion model 320M' are input to the adjusted gas-phase combustion model 320M' as input values ​​(an example of a second measurement value). In addition, some of the unobservable state variables calculated by the adjusted grate combustion model 310M' one cycle earlier are also input to the adjusted gas-phase combustion model 320M'. As a result, the adjusted gas-phase combustion model 320M' runs the simulation and calculates the unobservable state variables.

[0093] Some of the unobservable state variables calculated by the tuned gas-phase combustion model 320M' are input to the prediction model 900 as input data. In addition, some of the unobservable state variables output by the tuned gas-phase combustion model 320M' are input to the tuned boiler heat recovery / power generation model 330M' as input values.

[0094] As shown in Figure 9, among the measured values ​​stored in the measured value storage unit 450, the measured values ​​necessary for running the simulation using the adjusted boiler heat recovery / power generation model 330M' are input to the adjusted boiler heat recovery / power generation model 330M' as input values ​​(an example of a second measured value). In addition, some of the unobservable state variables calculated by the adjusted gas-phase combustion model 320M' one cycle earlier are also input to the adjusted boiler heat recovery / power generation model 330M'. As a result, the adjusted boiler heat recovery / power generation model 330M' runs the simulation and calculates the unobservable state variables.

[0095] Some of the unobservable state variables calculated by the adjusted boiler heat gain / power generation model 330M' are input to the prediction model 900 as input data.

[0096] As shown in Figure 9, among the measurement values ​​stored in the measurement value storage unit 450, measurement values ​​other than those used as input values, which are used as input data for the prediction model 900 (an example of a third measurement value), are input to the prediction model 900 as input data.

[0097] As shown in Figure 9, the prediction model 900 is • Some of the unobservable state variables calculated by the adjusted grate combustion model 310M', • Some of the unobservable state variables calculated by the adjusted gas-phase combustion model 320M', • Some of the unobservable state variables calculated by the adjusted boiler heat gain / power generation model 330M', • A third measurement that differs from the second measurement, The input predicts the process values ​​used to control the incineration process. The process values ​​predicted by the prediction model 900 are input to the comparison / modification unit 910. The comparison / modification unit 910 receives the process values ​​used to control the incineration process from the measured values ​​stored in the measured value storage unit 450 as ground truth data.

[0098] As a result, the comparison / modification unit 910 calculates the error between the process value predicted by the prediction model 900 and the process value input as the ground truth data, and updates the model parameters of the prediction model based on the calculated error. As a result, the prediction model learning unit 810 constructs a trained prediction model.

[0099] <Processing flow by the learning device during the learning phase> Next, we will explain the processing flow by the learning device 120 during the learning phase. Figure 10 is a flowchart showing the processing flow during the learning phase.

[0100] In step S1001, the learning device 120 acquires measurement values ​​from the plant control system 110.

[0101] In step S1002, the learning device 120 adjusts the parameters of the physical model using the first measurement value and generates an adjusted physical model. The learning device 120 stores the generated adjusted physical model in the auxiliary storage device 203.

[0102] In step S1003, the learning device 120 performs a simulation using the generated adjusted physical model and calculates an unobservable state variable based on a second measurement that differs from the first measurement.

[0103] In step S1004, the learning device 120 uses the calculated unobservable state variables and a third measurement value different from the second measurement value as input data, and process values ​​as ground truth data to train a predictive model and construct a trained predictive model.

[0104] In step S1005, the learning device 120 stores the constructed trained prediction model in the auxiliary storage device 203.

[0105] <System configuration during the prediction phase of the target plant> Next, the system configuration in the prediction phase of a target plant to which the prediction device according to the first embodiment is applied will be described. Figure 11 is a first diagram showing an example of the system configuration in the prediction phase of a target plant.

[0106] As shown in Figure 11, the target plant 10 to which the prediction device 1100 according to the first embodiment is applied comprises target equipment 100 and a plant control system 110. Note that the target equipment 100 and plant control system 110 shown in Figure 11 are the same as the target equipment 100 and plant control system 110 shown in Figure 1, so their description is omitted here.

[0107] The prediction device 1100 includes a tuned physical model and a trained prediction model. The prediction device 1100 acquires measurement values ​​measured in the processing steps of the target equipment 100 from the plant control system 110, and predicts process values ​​to be used for controlling the processing steps of the target equipment 100 based on the acquired measurement values.

[0108] <Hardware configuration of the prediction device> Next, the hardware configuration of the prediction device 1100 will be described. Figure 12 shows an example of the hardware configuration of the prediction device.

[0109] As shown in Figure 12, the prediction device 1100 includes a processor 1201, memory 1202, auxiliary storage device 1203, connection device 1204, communication device 1205, and drive device 1206. The processor 1201, memory 1202, auxiliary storage device 1203, connection device 1204, communication device 1205, and drive device 1206 of the prediction device 1100 are interconnected via a bus 1207.

[0110] Since the hardware configuration of the prediction device 1100 is the same as that of the learning device 120 shown in Figure 2, we will omit the explanation here.

[0111] <Functional Configuration of the Prediction Device> Next, the functional configuration of the prediction device 1100 will be described. Figure 13 is the first diagram showing an example of the functional configuration of the prediction device in the prediction phase. The prediction device 1100 has a prediction program installed, and when this program is executed, the prediction device 1100 functions as a measurement value acquisition unit 400 and a prediction unit 1300. Of these, the measurement value acquisition unit 400 has already been explained using Figure 4, so its explanation will be omitted here.

[0112] The prediction unit 1300 includes a tuned grate combustion model 310M', a tuned gas-phase combustion model 320M', a tuned boiler heat recovery / power generation model 330M', a trained prediction model 900', and a visualization unit 1310.

[0113] The adjusted grate combustion model 310M' is, • The measurement values ​​necessary for running the simulation are obtained from the measurement value acquisition unit 400. • The state variables from the previous cycle required for running the simulation are obtained from the adjusted gas-phase combustion model 320M'. State variables are calculated by running the simulation. As a result, the tuned grate combustion model 310M' outputs the state variables necessary for running the simulation by the tuned gas-phase combustion model 320M', and the trained prediction model 900' also outputs the state variables necessary for predicting process values.

[0114] The tuned vapor phase combustion model 320M' is, • The measurement values ​​necessary for running the simulation are obtained from the measurement value acquisition unit 400. • Obtain the state variables from the previous cycle required for running the simulation from the adjusted grate combustion model 310M'. The state variables are calculated by running the simulation. As a result, the tuned gas-phase combustion model 320M' outputs the state variables necessary for running the simulation by the tuned boiler heat recovery / power generation model 330M', and the trained prediction model 900' also outputs the state variables necessary for predicting process values.

[0115] The tuned boiler heat recovery / power generation model 330M' is, • The measurement values ​​necessary for running the simulation are obtained from the measurement value acquisition unit 400. • Obtain the state variables from the previous cycle required for running the simulation from the adjusted gas-phase combustion model 320M'. The state variables are calculated by running the simulation. This allows the tuned boiler heat gain / power generation model 330M' to output the state variables necessary for the trained prediction model 900' to predict process values.

[0116] The pre-trained predictive model 900' is, • State variables output from the adjusted grate combustion model 310M', adjusted gas-phase combustion model 320M', and adjusted boiler heat recovery / power generation model 330M', • The measurement values ​​obtained by the measurement value acquisition unit 400, Based on this, we predict the process value.

[0117] The visualization unit 1310 visualizes state variables calculated by the adjusted grate combustion model 310M', the adjusted gas-phase combustion model 320M', and the adjusted boiler heat recovery / power generation model 330M'. The visualization unit 1310 also visualizes process values ​​predicted by the trained prediction model 900' and process values ​​included in the measurement values ​​acquired by the measurement value acquisition unit 400.

[0118] <Details of the prediction process by the prediction unit of the prediction device> Next, we will explain the details of the prediction process performed by the prediction unit 1300 of the prediction device 1100. Figure 14 is a diagram showing a specific example of the prediction process performed by the prediction device during the prediction phase.

[0119] As shown in Figure 14, among the measurement values ​​acquired by the measurement value acquisition unit 400, the measurement values ​​necessary for running the simulation using the adjusted grate combustion model 310M' are input to the adjusted grate combustion model 310M' as input values ​​(an example of the second measurement value). In addition, some of the unobservable state variables calculated by the adjusted gas-phase combustion model 320M' one cycle earlier are input to the adjusted grate combustion model 310M'. As a result, the adjusted grate combustion model 310M' runs the simulation and calculates the unobservable state variables.

[0120] Some of the unobservable state variables calculated by the adjusted grate combustion model 310M' are input to the trained prediction model 900' as input data. Similarly, some of the unobservable state variables calculated by the adjusted grate combustion model 310M' are input to the adjusted gas-phase combustion model 320M' as input values. Furthermore, the unobservable state variables calculated by the adjusted grate combustion model 310M' are output to the visualization unit 1310.

[0121] As shown in Figure 14, among the measurement values ​​acquired by the measurement value acquisition unit 400, the measurement values ​​necessary for running the simulation using the adjusted gas-phase combustion model 320M' are input to the adjusted gas-phase combustion model 320M' as input values ​​(an example of the second measurement value). In addition, some of the unobservable state variables calculated by the adjusted grate combustion model 310M' one cycle earlier are input to the adjusted gas-phase combustion model 320M'. As a result, the adjusted gas-phase combustion model 320M' runs the simulation and calculates the unobservable state variables.

[0122] Some of the unobservable state variables calculated by the tuned gas-phase combustion model 320M' are input to the trained prediction model 900' as input data. Similarly, some of the unobservable state variables calculated by the tuned gas-phase combustion model 320M' are input to the tuned boiler heat recovery / power generation model 330M' as input values. Furthermore, the unobservable state variables calculated by the tuned gas-phase combustion model 320M' are output to the visualization unit 1310.

[0123] As shown in Figure 14, among the measurements acquired by the measurement acquisition unit 400, the measurements necessary for running the simulation using the adjusted boiler heat recovery / power generation model 330M' are input to the adjusted boiler heat recovery / power generation model 330M' as input values ​​(an example of the second measurement). In addition, some of the unobservable state variables calculated by the adjusted gas-phase combustion model 320M' one cycle earlier are input to the adjusted boiler heat recovery / power generation model 330M'. As a result, the adjusted boiler heat recovery / power generation model 330M' runs the simulation and calculates the unobservable state variables.

[0124] Some of the unobservable state variables calculated by the adjusted boiler heat recovery / power generation model 330M' are input to the trained prediction model 900' as input data. In addition, the unobservable state variables calculated by the adjusted boiler heat recovery / power generation model 330M' are output to the visualization unit 1310.

[0125] As shown in Figure 14, among the measurement values ​​acquired by the measurement value acquisition unit 400, • Measurement values ​​other than those used as input values, • Measured values ​​used as input data for the trained predictive model 900' (an example of the third measured value), This is input as input data into the trained predictive model 900'.

[0126] Furthermore, as shown in Figure 14, the trained prediction model 900' is, • Some of the unobservable state variables calculated by the adjusted grate combustion model 310M', • Some of the unobservable state variables calculated by the adjusted gas-phase combustion model 320M', • Some of the unobservable state variables calculated by the adjusted boiler heat gain / power generation model 330M', • A third measurement that differs from the second measurement, The input predicts process values ​​used to control the incineration process. The process values ​​predicted by the trained prediction model 900' are output to the visualization unit 1310. The visualization unit 1310 also receives process values ​​included in the measurement values ​​acquired by the measurement value acquisition unit 400.

[0127] As a result, the visualization unit 1310, • Unobservable state variables calculated by the adjusted grate combustion model 310M', • Unobservable state variables calculated by the adjusted gas-phase combustion model 320M' • Unobservable state variables calculated by the adjusted boiler heat gain / power generation model 330M' • Process values ​​predicted by the trained predictive model 900', • Process values ​​measured in the plant control system 110, To visualize it.

[0128] <Incinerator monitoring screen> Next, we will describe the incinerator monitoring screen displayed by the visualization unit 1310. Figure 15 shows an example of the incinerator monitoring screen. Due to space limitations, Figure 15 only shows the process value, state variable A, and state variable B as visualized by the visualization unit 1310.

[0129] Of these, graph 1510 displays the process values ​​predicted by the trained prediction model 900' and the process values ​​measured by the plant control system 110. In graph 1510, the horizontal axis represents time and the vertical axis represents the process values.

[0130] Graph 1520 displays state variable A, which is one of the unobservable state variables calculated by, for example, the adjusted grate combustion model 310M'. In Graph 1520, the horizontal axis represents time, and the vertical axis represents the state variable.

[0131] Graph 1530 displays state variable B, which is one of the unobservable state variables calculated by, for example, the adjusted gas-phase combustion model 320M'. In Graph 1530, the horizontal axis represents time, and the vertical axis represents the state variable.

[0132] <Processing flow in the prediction phase> Next, we will explain the processing flow by the prediction device 1100 during the prediction phase. Figure 16 is a first flowchart showing the processing flow during the prediction phase.

[0133] In step S1601, the prediction device 1100 acquires measurement values ​​from the plant control system 110.

[0134] In step S1602, the prediction device 1100 operates the adjusted physical model to calculate an unobservable state variable from a second measurement that is different from the first measurement.

[0135] In step S1603, the prediction device 1100 uses a portion of the calculated unobservable state variables and a third measurement value different from the second measurement value as input data to operate the trained prediction model and predict the process value.

[0136] In step S1604, the prediction device 1100 visualizes the calculated unobservable state variables, the predicted process values, and the measured process values.

[0137] In step S1605, the prediction device 1100 determines whether to continue processing. If it determines to continue processing (i.e., if the answer in step S1605 is YES), it returns to step S1601.

[0138] On the other hand, if it is determined in step S1605 to terminate the process (if the answer is NO in step S1605), the process in the prediction phase is terminated.

[0139] <Summary> As is clear from the above description, the learning device 120 according to the first embodiment is The parameters of the physical model used to simulate the behavior of the processing steps of the target plant are adjusted using the first measured values ​​obtained during the processing steps of the target plant, and an adjusted physical model is generated. • By operating a pre-adjusted physical model based on a second measurement value, which differs from the first measurement value, measured in the processing step of the target plant, unobservable state variables in the processing step are calculated. Furthermore, a predictive model is trained using the calculated state variables and a third measurement value, which differs from the second measurement value, measured in the processing step, to construct a trained predictive model.

[0140] As a result, the learning device 120 according to the first embodiment makes it possible to construct a trained predictive model that can predict process values ​​with high accuracy compared to a trained predictive model that is trained based only on measurements taken in the processing steps of the target equipment 100.

[0141] Furthermore, the prediction device 1100 according to the first embodiment is • Run the adjusted physical model to calculate state variables based on a second measurement taken during the processing step, then run the trained predictive model to predict the process values ​​of the target plant based on the calculated state variables and a third measurement taken during the processing step.

[0142] Thus, by using an unobservable state variable when predicting the process values ​​of the target plant, the prediction accuracy when predicting the process values ​​of the target plant can be improved according to the first embodiment.

[0143] [Second Embodiment] The prediction device 1100 according to the first embodiment described above is configured to visualize the process values ​​predicted by the trained prediction model 900' in the visualization unit 1310. In contrast, the prediction device according to the second embodiment searches for control values ​​that realize newly set target values ​​based on the process values ​​predicted by the trained prediction model 900'. The second embodiment will now be described, focusing on the differences from the first embodiment described above.

[0144] <System configuration during the prediction phase of the target plant> First, we will describe the system configuration in the prediction phase of a target plant to which the prediction device according to the second embodiment is applied. Figure 17 is a second diagram showing an example of the system configuration in the prediction phase of a target plant.

[0145] As shown in Figure 17, the target plant 10 to which the prediction device 1700 according to the second embodiment is applied comprises target equipment 100 and a plant control system 110. Note that the target equipment 100 and plant control system 110 shown in Figure 17 are the same as the target equipment 100 and plant control system 110 shown in Figure 11, so their description is omitted here.

[0146] The prediction device 1700 has a pre-tuned physical model and a trained prediction model. When a newly set target value is transmitted from the plant control system 110, the prediction device 1700 searches for a control value that will achieve that target value.

[0147] Specifically, the prediction device 1700 calculates the measured value after the change in the control value as a simulation value, and uses the calculated simulation value to predict the process value. The prediction device 1700 searches for a control value that achieves the newly set target value by changing the control value so that the predicted process value matches the transmitted target value.

[0148] The prediction device 1700 transmits the detected control values ​​to the plant control system 110. This allows the plant control system 110 to immediately operate the target equipment 100 using the detected control values, even if new target values ​​are set.

[0149] <Functional Configuration of the Prediction Device> Next, the functional configuration of the prediction device 1700 will be described. Figure 18 is a second diagram showing an example of the functional configuration of the prediction device in the prediction phase. The prediction device 1700 has a prediction program installed, and when this program is executed, the prediction device 1700 functions as a difference calculation unit 1810, an optimization unit 1820, a simulation value calculation unit 1830, and a prediction unit 1300. Of these, the prediction unit 1300 has already been explained using Figure 13, so its explanation will be omitted here.

[0150] The difference calculation unit 1810 obtains a newly set target value from the plant control system 110. The difference calculation unit 1810 obtains the process value predicted by the prediction unit 1300. The difference calculation unit 1810 calculates the difference between the obtained target value and the predicted process value and notifies the optimization unit 1820 of the difference value.

[0151] The optimization unit 1820 calculates control values ​​for operating the target equipment 100 based on the difference value notified by the difference calculation unit 1810. The optimization unit 1820 notifies the simulation value calculation unit 1830 of the calculated control values. Furthermore, if the difference value notified by the difference calculation unit 1810 falls below a predetermined threshold, the optimization unit 1820 notifies the plant control system 110 of the control values ​​at which the difference value was calculated.

[0152] The simulation value calculation unit 1830 obtains control values ​​from the optimization unit 1820, calculates the measured values ​​obtained when the target equipment 100 is operated using the acquired control values, and inputs them to the prediction unit 1300.

[0153] As a result, the prediction unit 1300 uses the input simulation values ​​to operate the tuned grate combustion model 310', the tuned gas-phase combustion model 320', the tuned boiler heat recovery / power generation model 330', and the trained prediction model 900' to predict process values.

[0154] <Processing flow in the prediction phase> Next, we will explain the processing flow by the prediction device 1700 during the prediction phase. Figure 19 is a second flowchart showing the processing flow during the prediction phase.

[0155] In step S1901, the prediction device 1700 obtains a newly set target value from the plant control system 110.

[0156] In step S1902, the prediction device 1700 calculates the difference between the predicted process value and the acquired target value.

[0157] In step S1903, the prediction device 1700 determines whether the calculated difference value is less than or equal to a predetermined threshold. If it is determined in step S1903 that the difference is not less than or equal to the predetermined threshold (i.e., the answer in step S1903 is NO), the device proceeds to step S1904.

[0158] In step S1904, the prediction device 1700 changes the control value according to the difference value.

[0159] In step S1905, the prediction device 1700 calculates the measured value when controlled with the modified control value as the simulation value.

[0160] In step S1906, the prediction device 1700 predicts process values ​​based on the simulation values ​​and returns to step S1902.

[0161] On the other hand, if it is determined in step S1903 that the value is below a predetermined threshold (i.e., if the answer in step S1903 is YES), the process proceeds to step S1907.

[0162] In step S1907, the prediction device 1700 transmits the control value it determined to be below a predetermined threshold to the plant control system 110.

[0163] <Summary> As is clear from the above description, the prediction device 1700 according to the second embodiment is The system searches for control values ​​for the target plant so that the predicted process values ​​approach the newly set target values.

[0164] As a result, according to the second embodiment, even if a new target value is set in the plant control system 110, the target equipment 100 can be operated immediately using the searched control value.

[0165] [Third Embodiment] In the embodiments described above, the learning device and the prediction device were described as separate devices, but the learning device and the prediction device may be the same device. Also, in the embodiments described above, the learning device and the prediction device were described as being provided separately from the plant control system, but some or all of the functions of the learning device and some or all of the functions of the prediction device may be implemented in the plant control system.

[0166] Furthermore, although the learning device and the prediction device were described in each of the above embodiments as being implemented by a single device, they may also be implemented by multiple devices. For example, each functional unit implemented by the learning device and the prediction device may be implemented in a distributed manner across multiple devices.

[0167] It should be noted that the present invention is not limited to the configurations shown in the above embodiments, including combinations with other elements. These aspects can be modified without departing from the spirit of the present invention and can be appropriately determined according to their application. [Explanation of Symbols]

[0168] 10: Target Plant 100: Target equipment 110: Plant control system 120: Learning device 310M: Grate combustion model 310M': Adjusted grate combustion model 320M: Gas-phase combustion model 320M': Adjusted gas-phase combustion model 330M: Boiler heat recovery / power generation model 330M': Adjusted boiler heat recovery / power generation model 400: Measurement value acquisition unit 410: Parameter adjustment unit 420~440: Adjustment section 800: Learning Department 810: Predictive Model Learning Unit 900: Predictive Model 1100: Prediction device 1300: Prediction Department 1310 :Visualization part 1700: Prediction device 1810: Difference calculation part 1820: Optimization Department 1830: Simulation value calculation unit

Claims

1. A physical model for performing simulations of the behavior of a plant's processing steps, comprising a modified physical model whose parameters are adjusted based on first measured values ​​taken in the processing steps, The system stores a trained predictive model that predicts the process values ​​of the plant, which is constructed by training a predictive model using an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step. The adjusted physical model is operated to calculate state variables based on the second measurement taken in the processing step, the trained prediction model is operated to predict the process values ​​of the plant based on the calculated state variables and the third measurement taken in the processing step. Prediction device.

2. The state variables calculated by running the aforementioned adjusted physical model and the process values ​​predicted by running the aforementioned trained prediction model are visualized. The prediction device according to claim 1.

3. The aforementioned adjusted physical model is The first measurement value measured in the processing step is input to the physical model, and the output value calculated by the physical model is generated by adjusting the parameters of the physical model using a Kalman filter so that the output value calculated by the physical model approaches the output value measured in the processing step. The prediction device according to claim 1.

4. The aforementioned trained predictive model is The predictive model is constructed by using the following as input data: an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, and using the process values ​​measured in the processing step of the plant as ground truth data. The prediction device according to claim 1.

5. The trained prediction model is run to search for control values ​​for the plant so that the predicted process values ​​approach newly set target values. The prediction device according to claim 1.

6. The aforementioned physical model is a grate combustion model, a gas-phase combustion model, or a boiler heat acquisition and power generation model. The prediction device according to claim 1.

7. A physical model for simulating the behavior of a waste incinerator during the incineration process, specifically when the waste moves on a movable grate and undergoes the processes of drying, thermal decomposition, and combustion in sequence to achieve complete combustion, comprising an adjusted grate combustion model, which is an adjusted physical model whose parameters are adjusted based on first measured values ​​taken during the incineration process, and A physical model for simulating the behavior of combustible gas generated by combustion in the incineration process of the incinerator, when it is combusted by supplied air, comprising an adjusted gas-phase combustion model, which is an adjusted physical model whose parameters have been adjusted based on first measured values ​​taken in the incineration process, A physical model for simulating the behavior of the incineration process of the aforementioned incinerator, specifically when steam generated by the absorption of heat from exhaust gas in a waste heat boiler rotates a steam turbine and generates electricity, wherein the adjusted physical model is a modified boiler heat recovery / power generation model whose parameters have been adjusted based on first measured values ​​taken in the incineration process, and The system stores a trained predictive model that predicts the process values ​​of the incinerator, which is constructed by training the predictive model using an unobservable state variable in the incineration process, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the incineration process of the incinerator, and a third measurement value different from the second measurement value measured in the incineration process. The adjusted physical model is operated to calculate state variables based on the second measurement taken in the incineration process, the trained prediction model is operated to predict the process values ​​of the incinerator based on the calculated state variables and the third measurement taken in the incineration process. Prediction device.

8. An adjustment unit that adjusts the parameters of a physical model used to simulate the behavior of a plant's processing steps using first measured values ​​taken during the processing steps, and generates an adjusted physical model. A learning unit that trains a predictive model using an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, and constructs a trained predictive model. A learning device having

9. The adjustment unit is, The adjusted physical model is generated by inputting the first measurement value measured in the processing step into the physical model, and then adjusting the parameters of the physical model using a Kalman filter so that the output value calculated by the physical model approaches the output value measured in the processing step. The learning device according to claim 8.

10. The aforementioned learning unit, The trained predictive model is constructed by using as input data an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, and by training the predictive model using the process values ​​measured in the processing step of the plant as ground truth data. The learning device according to claim 8.

11. A physical model for performing simulations of the behavior of a plant's processing steps, comprising a modified physical model whose parameters are adjusted based on first measured values ​​taken in the processing steps, A prediction method for a prediction device that stores a trained prediction model for predicting process values ​​of a plant, which is constructed by training a prediction model using an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, the trained prediction model for predicting process values ​​of the plant, the trained prediction model for which the trained prediction model is constructed, the trained prediction model for predicting process values ​​of the plant, the trained prediction device stores a prediction device that stores a prediction device that stores a prediction device that stores a prediction device that stores a trained prediction model for predicting process values ​​of the plant, the trained prediction model for which the trained prediction model is constructed by training the prediction model using an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, the trained prediction model for predicting process values ​​of the plant, the prediction device stores a prediction device that The adjusted physical model is operated to calculate state variables based on the second measurement taken in the processing step, the trained prediction model is operated to predict the process values ​​of the plant based on the calculated state variables and the third measurement taken in the processing step. A prediction method in which the computer of the prediction device performs the processing.

12. A physical model for performing simulations of the behavior of a plant's processing steps, comprising a modified physical model whose parameters are adjusted based on first measured values ​​taken in the processing steps, A computer in a prediction device that stores a trained prediction model for predicting the process values ​​of the plant, which is constructed by training a prediction model using an unobservable state variable in the processing step, calculated by operating the adjusted physical model based on a second measurement value different from the first measurement value measured in the processing step, and a third measurement value different from the second measurement value measured in the processing step, The adjusted physical model is operated to calculate state variables based on the second measurement taken in the processing step, the trained prediction model is operated to predict the process values ​​of the plant based on the calculated state variables and the third measurement taken in the processing step. A prediction program designed to execute a process.

Images