Control method, control device, and program

Abstract

This invention provides a control method that enables real-time optimization control of batch processes based on estimated values, while ensuring the same level of accuracy as simulation results from conventional process simulators when estimating real-time variables in batch processes, and reducing the time required to output calculation results per batch. [Solution] A control method for controlling a batch process, comprising: an estimation step of estimating real-time variables in a batch process based on measured values ​​using a machine learning model; and a provision step of displaying the estimated values ​​on a display unit and / or notifying a control unit that controls the batch process based on the estimated values.

Description

Technical Field

[0001] The present invention relates to a control method, a control device, and a program for controlling a batch process.

Background Art

[0002] Process manufacturing is used as a manufacturing means in industries such as chemistry, steel, pharmaceuticals, and food. Among them, a batch process that executes a series of chemical reactions or manufacturing processes at once is known. And, as an invention related to the control of the batch process, for example, there is one described in Patent Document 1.

[0003] Patent Document 1 describes a control method and the like aimed at facilitating the control of a batch process. The control method described in Patent Document 1 includes steps of generating an estimated value of an unmeasurable real-time variable of a batch process using a first principle model, providing the estimated value to a process control routine for controlling the batch process, and generating a signal based on the estimated value using the process control routine to control the batch process.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, in batch processes, the system state changes moment by moment, so conventional process simulators require calculations at predetermined time steps, and it takes time to output the calculation results for one batch (for example, 20 min / B). In that case, if the batch time is 120 min, conventional process simulators can only obtain simulation results a few times per batch, and it cannot be said that they can adequately optimize the batch process based on the simulation results. The same applies to the batch process control method etc. described in Patent Document 1.

[0006] Therefore, the present invention aims to provide a control method, a control device, and a program that, when estimating real-time variables in a batch process, ensure an accuracy comparable to that of the simulation results of conventional process simulators, reduce the time required to output calculation results per batch, and enable real-time level optimization control of the batch process based on estimated values. [Means for solving the problem]

[0007] As a result of diligent research to solve the above problems, the inventors have completed an invention that significantly reduces the time required to output calculation results per batch and enables real-time level optimization control of the batch process based on estimated values.

[0008] In other words, the present invention relates to the following invention. <1> An estimation step that uses a trained model (machine learning model) to estimate real-time variables in a batch process based on measured values, A providing step of displaying the estimated value on a display unit and / or notifying a control unit that controls the batch process based on the estimated value, A method for controlling batch processes, including those mentioned above. <2> The estimation step involves measuring information about the feed supplied to the apparatus and / or information about the state of the apparatus, and estimating the real-time variables in the batch process based on the measured values. <1> The control method described above. <3> The aforementioned machine learning model is a model constructed using a neural network suitable for learning time-series data, or a model constructed by combining the said neural network with arbitrary physical equations. <1> or <2> The control method described above. <4> The input data of the neural network is updated by backpropagation based on the computational history of the machine-learned model. <3> The control method described above. <5> The machine learning model is generated using one or more pieces of information from among the following as training data: information about the supplies supplied to the apparatus, information about the products generated within the apparatus, or information about the state of the apparatus. <1> from <4> A control method as described in any of the following. <6> The simulation results from the process simulator or measurement data from the actual machine are used as training data. <5> The control method described above. <7> The batch process includes at least one of a reaction, a phase change, or membrane separation. <1> from <6> A control method as described in any of the following. <8> An estimation unit that uses a machine learning model to estimate real-time variables in a batch process based on measured values, A providing unit that displays the estimated value on a display unit and / or notifies a control unit that controls the batch process based on the estimated value, A batch process control device having the following features. <9> Computers, An estimation unit that uses a machine learning model to estimate real-time variables in a batch process based on measured values, A providing unit that displays the estimated value on a display unit and / or notifies a control unit that controls the batch process based on the estimated value, A program that has the ability to operate as a control device for batch processes. [Effects of the Invention]

[0009] The present invention provides a control method that significantly reduces the time required to output calculation results per batch and enables real-time level optimization control of the batch process based on estimated values. [Brief explanation of the drawing]

[0010] [Figure 1] This diagram illustrates a conventional batch process control method using a process simulator. [Figure 2] This is a diagram illustrating a batch process control method according to the present invention. [Figure 3] This is a schematic functional block diagram of the control device according to the present invention. [Figure 4] This figure shows an example of a reservoir computing model. [Figure 5] This is a diagram illustrating the batch model of the epoxy resin formation reaction. [Figure 6] This is an illustrative diagram to explain how to create training data. [Figure 7] This is an illustrative diagram illustrating the estimation of real-time variables using the control device according to the present invention. [Figure 8] This figure shows the simulation results in this embodiment. [Modes for carrying out the invention]

[0011] The embodiments of the present invention will be described in detail below, but the description of the constituent elements described below is an example (representative example) of an embodiment of the present invention, and the present invention is not limited to the following unless its gist is changed. In this description, "calculation" of real-time variables and "estimation" of real-time variables are considered synonymous. When the expression "~" is used in this description, it is used to mean an expression that includes the numerical values ​​before and after it.

[0012] <Conventional Control Method for Batch Process> FIG. 1 is a diagram for explaining a conventional control method for a batch process using a process simulator. A process simulator is a simulator that optimizes and designs a process considering reaction engineering, heat transfer, mass balance, etc. For example, ASPEN (registered trademark) of Aspentech Japan Co., Ltd., AVEVA (registered trademark) Process Simulation of AVEVA Co., Ltd., UniSim (registered trademark) of Honeywell, Symmetry of Schlumberger, and CHEMCAD (registered trademark) of Chemstations are known.

[0013] First, referring to FIG. 1, the control of the batch process of a plant using the process simulator 20 will be described.

[0014] Step (a): First, based on the measured value of the measurement target, the PV value (input value) until the end point of the batch process is predicted. The measurement target is information about raw materials (flow rate), information about containers such as reactors and flasks in the plant where raw materials are charged (temperature, pressure), etc. Also, the measured value is measured using various measuring instruments such as flow meters, thermometers, and pressure gauges, or measured using instruments for sampling and analyzing the concentration of raw materials, etc. Note that the input value is a value that changes every moment, such as temperature.

[0015] Step (b): Next, using the process simulator 20, return to the input value generated until the end point of the batch process and calculate the real-time variable. A real-time variable is a value that changes in real time (every moment) in the container and is a parameter for which prediction of the value monitored in the batch process is required. Note that expressions such as monitoring parameters and key parameters generally used in batch processes are also synonymous with this.

[0016] Step (c): Next, optimize the input values ​​so that the target real-time variable reaches the desired value. For example, suppress (optimize) the increase in the input value over time so that the target real-time variable reaches the desired value.

[0017] Step (d): Then, based on the optimized input values, the SV value (target value) is determined. Based on the SV values ​​determined as described above, the plant's batch process is manually controlled so that the target real-time variable reaches the desired value.

[0018] Thus, considering the methods for controlling the target real-time variable to a desired value, the optimal control of a batch process requires the realization of real-time control or control close to it. In that case, the calculation of the real-time variable in (2) above must be sufficiently faster than the batch time (the time from the start to the end of the batch process).

[0019] However, batch processes, especially those involving reactions or phase changes, have the property that the properties within the system change integrally from start to finish. Therefore, the invention described in Patent Document 1 (control method using the first principle model) requires calculations to be repeated at short time steps from the initial state at the start of the batch process to calculate the properties at the end of the batch process, which results in an enormous calculation time for the real-time variables in (2) above.

[0020] Furthermore, even when using process simulators as described above, batch processes involving complex physical properties and operations such as reactions and distillation inevitably require several minutes to an hour of calculation time.

[0021] <Batch process control method according to the present invention> The batch process control method according to the present invention solves the problem of the time required to calculate such real-time variables by using the control device (control device 10) according to the present invention instead of the process simulator 20, as shown in Figure 2.

[0022] Furthermore, the control device according to the present invention has a machine learning model that can perform real-time variable calculations with an accuracy comparable to that of a process simulator. Therefore, first, the configuration of the control device according to the present invention will be described, and then the steps for generating the machine learning model and the steps for using the generated machine learning model will be described, respectively.

[0023] <Control device for batch processes according to the present invention> Figure 3 is a schematic functional block diagram of a control device according to the present invention. One embodiment of the control device 10 includes an input unit 11, an output unit 12, a storage unit 13, an estimation unit 14, a supply unit 15, and a display unit 16.

[0024] The input unit 11 and output unit 12 have the function of receiving data input from an external source (including an external device) and outputting data to an external source. Furthermore, the input unit 11 and output unit 12 also have the function of communicating with an external device via wired or wireless connection.

[0025] The memory unit 13 is a memory or database that stores data input from external sources, various data necessary for the operation of the control device 10, and programs (including programs for operating the computer as the control device 10). The generated machine learning model can also be stored in the memory unit 13.

[0026] The estimation unit 14 uses the machine learning model stored in the memory unit 13 to estimate the real-time variables, which will be described later.

[0027] The providing unit 15 displays the real-time variables estimated by the estimation unit 14 on the display unit 16, or notifies the control unit of an external device that controls the batch process based on the real-time variables estimated by the estimation unit 14.

[0028] The display unit 16 is, for example, a display, and the real-time variables estimated by the estimation unit 14 are displayed on the display by the providing unit 15.

[0029] In addition, the control device 10 has functions related to the control of the input unit 11 to the display unit 16, and various controls of the batch process.

[0030] Of course, the control device according to the present invention is not limited to the exemplified configuration and can be modified as appropriate. For example, the supply unit 15 can perform real-time variable estimation instead of the estimation unit 14, or some functions (for example, the storage unit 13) can be located outside the control device 10, such as in the cloud or external storage.

[0031] Furthermore, the control device 10 may have the function of controlling the batch process based on the real-time variables estimated by the estimation unit 14, rather than being an external device. The providing unit 15 may also display the real-time variables estimated by the estimation unit 14 not only on the display unit 16, but also on the display of an external terminal (such as a personal computer, tablet, or smartphone).

[0032] [Process for generating a machine learning-based model] A machine learning model is generated by training a neural network with training data. Alternatively, the neural network may be constructed by combining it with arbitrary physical equations. In this case, the input and output data of the neural network may be transformed using the arbitrary physical equations, or constraints based on those physical equations may be imposed, resulting in what is known as a PINN (Physics Informed Neural Network). The physical equations used should preferably be reaction rate equations or vapor-liquid equilibrium equations, such as those incorporated into process simulators.

[0033] (Neural network) While there are no specific types of neural networks, such as DNNs (Deep Neural Networks) and CNNs (Convolutional Neural Networks), it is desirable to use one that is particularly suitable for learning time-series data. Specifically, RNNs (Recurrent Neural Networks), LSTMs (Long Short-Term Memory), and ESNs (Echo State Networks) are preferable.

[0034] (Training data) Training data is preferably derived from simulation results from process simulators or processed measurement data from actual equipment, but is not limited to these. Training data can also include information about feedstocks in batch processes, equipment such as kettles and flasks, and products. Furthermore, there are no particular limitations on the information used for training data; the amount of raw materials, and in the case of resins, information such as the average molecular weight, weight-average molecular weight, and average degree of polymerization of the product, as well as the concentration of specific substances or terminal functional groups such as intermediate products of the reaction, and the amount of substances discharged from the system can also be used as training data.

[0035] Furthermore, in order to perform control using a neural network model, it is desirable to update the input data of the neural network and optimize the input by using backpropagation (including diachronic backpropagation) based on the computational history of the machine-learned model. By calculating the partial derivative of the error function with respect to the input data using this method, the change in error in response to changes in the input data can be determined, and the minimum value of the error function, i.e., the optimal input data, can be found using gradient descent or other methods. In the batch process control method according to the present invention, optimization using backpropagation can also be performed.

[0036] For example, if estimates obtained using a machine learning model indicate that the quality of the product will deviate from the specified range at the end of the batch, some intervention is necessary to prevent this deviation. In this case, the input data (e.g., pressure in the reactor) is adjusted so that the estimates of the machine learning model (e.g., impurity concentration) fall within the specified range, and the operators or control unit perform operations in the batch process to achieve this adjusted input data, thereby ensuring the quality of the product. Furthermore, for optimization, conventional optimization methods such as simulated annealing or genetic algorithms may be used.

[0037] [Process using machine learning models] Next, with reference to Figure 2, the control of a batch process in a plant using the control device 10 according to the present invention will be described.

[0038] Step (A): First, generate PV values ​​(input values) up to the end of the batch process based on the actual measured values ​​of the target.

[0039] Step (B): Next, the control device 10 calculates real-time variables based on the input values ​​predicted up to the end of the batch process. Specifically, the control device 10 (estimation unit 14) uses the input values ​​as input parameters from the input layer and performs calculations using a machine learning model, and outputs the calculation results (estimation results).

[0040] Step (C): Next, optimize the input values ​​so that the target real-time variable reaches the desired value. For example, suppress (optimize) the increase in the input value over time so that the target real-time variable reaches the desired value.

[0041] Step (D): Then, based on the optimized input values, the SV value (setpoint value) is determined.

[0042] Here, the control device 10 can display the real-time variables calculated in step (B) on the display unit 16 or the display of an external terminal. Then, for example, a plant worker can manually perform the control in steps (C) and (D) based on the displayed real-time variables.

[0043] Furthermore, the control device 10 can notify the control unit of an external device that controls the batch process of the real-time variables calculated in step (B) above. This external device (control unit) can then automatically perform the control in steps (C) and (D) above based on the notified real-time variables.

[0044] Furthermore, the control device 10 can, for example, use the estimation unit 14 to perform steps (C) and (D) above, and display and provide the determined SV value to an external terminal or external device.

[0045] The batch process control method according to the present invention significantly reduces the calculation time of real-time variables in step (B) by using the control device 10 instead of the process simulator 20. Therefore, the time required to control the batch process, including the subsequent steps (steps (C) and (D)), can be significantly reduced, enabling real-time control or near-real-time control. [Examples]

[0046] The present invention will be described in more detail below with reference to examples, but the present invention is not limited to the following examples unless its essence is changed.

[0047] In this embodiment, the "epoxy resin formation reaction" will be described.

[0048] [Process for generating a machine learning-based model] (Neural network) In this embodiment, the neural network employs an ESN, which is one of the reservoir computing techniques (see Figure 4).

[0049] (Training data) In this embodiment, the training data used was the simulation results from ASPEN®. More specifically, training data was created by conducting a case study of 300 cases using the batch model of the "epoxy resin formation reaction" from ASPEN® (see Figure 5).

[0050] Figure 6 is an illustrative diagram to explain the creation of training data. In this embodiment, the input data (input values) and output data (real-time variables) created as training data are as follows.

[0051] (Input data) (1) Information regarding raw materials • Amount of water contained in the raw material [kmol] • Amount of substance [kmol] of epichlorohydrin (ECH) • Amount of substance [kmol] of bisphenol A (BPA) • Amount of impurities in the raw material [kmol] • Flow rate of sodium hydroxide solution (NaOHaq) [kg / min] *Profile from the start to the end of the batch process

[0052] (2) Information regarding the container • Temperature inside the reactor [°C] *Profile from the start to the end of the batch process • Pressure inside the reactor • Temperature [°C] of the apparatus used to separate the reaction materials from the vapors produced during the reaction. • Pressure of the apparatus used to separate the reaction raw materials from the vapor produced during the reaction.

[0053] (Output data) (3) Information regarding the products • Amount of epoxy-terminated functional groups in the polymer [kmol] • Amount of substance [kmol] of bisphenol A (BPA) • Amount of terminal functional groups derived from polymer impurities [kmol] • Amount of 1,2-chlorohydrin-terminated functional groups in the polymer [kmol]

[0054] In this example, various data in the "Information on Raw Materials" and "Information on Containers" described above were normalized using the Euclidean distance (L2 norm), and all data except for the temperature inside the reactor and the catalyst drop (aqueous solution of sodium hydroxide (NaOHaq)) were input into the ESN as constant values ​​from the start to the end of the batch process.

[0055] The configuration of the ESN in this embodiment is as follows: (Input layer) • A uniform distribution with upper and lower limits of ±1.2 for all variables.

[0056] (Reservoir layer) • Number of nodes: 600 • Weight density: 0.8 • Spectral radius: 1.0 Leaky rate: 0.2 • Activation function: tanh

[0057] Furthermore, in this embodiment, a machine learning model was generated using the terminal functional group quantity derived from the impurity with the largest fluctuation in the output data obtained when generating training data using ASPEN® as the target variable (target real-time variable). In addition, while the target variable in this embodiment is one of the terminal functional group quantities derived from impurities, various output data can of course be used as the target variable through tuning, etc.

[0058] In this example, 300 cases (300 sets) of training data were generated, of which 270 sets were used for training and 30 sets for evaluation.

[0059] [Process using machine learning models] The following describes an example of controlling a batch process in a plant using the control method according to the present invention, with reference to Figures 2, 7, and 8. However, in order to explain the significant advantages over conventional batch process control methods using process simulators, we will mainly describe step (B) above.

[0060] In step (B) above, a control device 10 with a machine learning model was used to estimate real-time variables based on input values ​​predicted up to the end of the batch process. In this embodiment, the input values ​​were the same as those in the batch model of ASPEN®'s "epoxy resin production reaction" (see Figure 6) (see Figure 7).

[0061] (Estimation result) Figure 8 is a graph comparing the simulation results of ASPEN® for 30 sets (30 batches) used for model evaluation with the amount of terminal functional groups derived from impurities estimated by the control device 10. It is immediately clear that the line showing the simulation results of ASPEN® (see "Physical Sim.") and the line showing the estimation results of the control device 10 (see "Reservoir") almost overlap, and the root mean square error (RMSE) is a sufficiently small 0.0018, indicating that the control device 10 is able to reproduce the simulation results of ASPEN® very well. In other words, the control device 10 can be said to guarantee an estimation accuracy comparable to that of conventional process simulators.

[0062] On the other hand, the time required to estimate real-time variables was approximately 3 seconds for 30 batches when using the control device 10. In other words, it was about 0.1 s / B. In contrast, using ASPEN®, the time taken was approximately 20 min / B. Furthermore, when the estimation time was measured without launching the ASPEN® GUI, the time taken was approximately 11 min / B.

[0063] These results indicate that the control device 10 can achieve an estimated speed approximately 6,000 to 1,200 times faster than that of a conventional process simulator.

[0064] Thus, according to the batch process control method of the present invention, the control device 10 can be used to estimate the real-time variable (amount of terminal functional groups derived from impurities) in step (B) of the batch process (epoxy resin production reaction) in about 0.1 s / B. Therefore, input values ​​other than the initial supply amount of feed (raw material) (such as the temperature inside the reactor, the input value for catalyst dropping, or the temperature of the apparatus for separating reaction raw materials from the vapor generated during the reaction) can be changed (increased or decreased) during the batch process to optimize the real-time variable (amount of terminal functional groups derived from impurities).

[0065] As described above, the embodiment can be modified or tuned as needed. For example, the number of nodes or spectral radius of the ESN can be changed, the washout calculation can be changed, or a neural network other than an ESN can be adopted.

[0066] Based on the above, the batch process control method, etc., according to the present invention makes it possible to reduce the time required to output calculation results per batch while ensuring the same level of accuracy as the simulation results of conventional process simulators when estimating real-time variables in a batch process, and to enable real-time level optimization control of the batch process based on estimated values. [Explanation of Symbols]

[0067] 10 Control device 11 Input section 12 Output section 13 Storage section 14 Estimation part 15 Providing Department 16 Display 20 Process Simulator

Claims

1. An estimation step that uses a machine learning model to estimate real-time variables in a batch process based on measured values, A providing step of displaying the estimated value on a display unit and / or notifying a control unit that controls the batch process based on the estimated value, A method for controlling batch processes, including...

2. The control method according to claim 1, wherein the estimation step involves measuring information about the feed supplied to the apparatus and / or information about the state of the apparatus, and estimating the real-time variables in the batch process based on the measured values.

3. The control method according to claim 1 or 2, wherein the machine learning model is a model constructed using a neural network suitable for learning time-series data, or a model constructed by combining the neural network with arbitrary physical equations.

4. The control method according to claim 3, wherein the input data of the neural network is updated by backpropagation based on the computation history of the machine-learned model.

5. The control method according to claim 1 or 2, wherein the machine learning model is generated using one or more pieces of information selected from information about a supply provided to the device, information about products generated within the device, or information about the state of the device as training data.

6. The control method according to claim 5, wherein simulation results from a process simulator or measurement data from an actual machine are used as training data.

7. The control method according to claim 1 or 2, wherein the batch process includes at least one of a reaction, a phase change, or membrane separation.

8. An estimation unit that uses a machine learning model to estimate real-time variables in a batch process based on measured values, A providing unit that displays the estimated value on a display unit, and / or notifies a control unit that controls the batch process based on the estimated value, A batch process control device having the following features.

9. Computers, An estimation unit that uses a machine learning model to estimate real-time variables in a batch process based on measured values, A providing unit that displays the estimated value on a display unit, and / or notifies a control unit that controls the batch process based on the estimated value, A program that has the ability to operate as a control device for batch processes.

Images