Model identification device, program, and model identification method

Abstract

The model of the controlled object is identified accurately in parallel with the operation of the controlled object. [Solution] The first acquisition unit 11A sequentially acquires the outlet temperature T (controlled variable), which is the response of the controlled object 20 to the flow rate target value (operated variable) of the feedback control when the controlled object 20 is operated by feedback control. The second acquisition unit 11B sequentially acquires the predicted values ​​Tm1 to Tm3 of the outlet temperature T from each of the models M1 to M3, which are the responses of each of the models M1 to M3 when the flow rate target value is input to each of the models M1 to M3 that model the controlled object 20. The machine learning unit 11C repeatedly updates the model parameters MP1 to MP3 that characterize each of the models M1 to M3 by machine learning, based on the difference between the outlet temperature T and the predicted values ​​Tm1 to Tm3, under the learning conditions of those model parameters.

Description

Technical Field

[0001] The present invention relates to a model identification apparatus, a program, and a model identification method. Background Art

[0002] Technologies for modeling a control target in feedback control have been developed. As such a technology, Patent Document 1 discloses a technology for identifying a model of a control target by stopping the operation of the control target (feedback control by a controller), performing two-position control on the control target, and identifying the model of the control target. Prior Art Documents Patent Documents

[0003] Patent Document 1 Japanese Unexamined Patent Application Publication No. 2012-89004 Summary of the Invention Problems to be Solved by the Invention

[0004] In the above-mentioned Patent Document 1, when identifying a model, the operation of the control target must be temporarily stopped.

[0005] An object of the present invention is to accurately perform identification of a model of a control target in parallel with the operation of the control target. Means for Solving the Problems

[0006] The model identification device according to this invention comprises: a first acquisition unit that sequentially acquires a control quantity which is the response of the controlled object to an input variable of the feedback control when the controlled object is operated by feedback control; a second acquisition unit that sequentially acquires a predicted value of the control quantity which is the output of the model when the input variable of the controlled object is input to the model that models the controlled object; and a machine learning unit that identifies the model by repeatedly updating the model parameters that characterize the model using machine learning based on the difference between the sequentially acquired control quantity and the predicted value.

[0007] Furthermore, the program according to this invention causes a computer to function as the model identification device.

[0008] Furthermore, the model identification method according to this invention comprises: a first acquisition step of sequentially acquiring a control quantity which is the response of the controlled object to the manipulated variable of the feedback control when the controlled object is operated by feedback control; a second acquisition step of sequentially acquiring a predicted value of the control quantity which is the output of the model when the manipulated variable is input to a model that models the controlled object; and a machine learning step of identifying the model by repeatedly updating the model parameters that characterize the model by machine learning based on the difference between the sequentially acquired control quantity and the predicted value. [Effects of the Invention]

[0009] According to the present invention, the model of the controlled object can be identified with high accuracy in parallel with the operation of the controlled object. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows the configuration of a model identification device according to one embodiment of this invention. [Figure 2] Figure 1 is a block diagram showing an example configuration of a model identification device and a feedback control system. [Figure 3]Figure 3(A) is a graph showing the relationship between the time change of the outlet temperature of the controlled object and the time change of the model's predicted value. Figure 3(B) is a graph showing the time change of the gain of the model updated by machine learning. Figure 3(C) is a graph showing the time change of the time constant of the model updated by machine learning. [Figure 4] This graph superimposes the time evolution of the controlled outlet temperature and the time evolution of the predicted values ​​of the three models, using a common vertical and horizontal axis. [Figure 5] This block diagram shows an example configuration of a model identification device and feedback control system related to a modified model. [Modes for carrying out the invention]

[0011] The embodiments of this invention will be described in detail below with reference to the drawings.

[0012] As shown in Figure 1, the model identification device 10 according to this embodiment is configured to identify a model of the controlled object 20, which is the target of feedback control. This identification is performed in parallel with the operation of the controlled object 20 (the operation of the controlled object 20 in its intended use, also called full operation), that is, with the feedback control applied to the controlled object 20. Furthermore, when the model identification device 10 identifies a model with good accuracy, it tunes the control parameters of the feedback control (the control parameters of the primary controller 31, described later) based on the model parameters of this model.

[0013] Now, let's describe the controlled object 20. The controlled object 20 is a plant that includes a fuel supply mechanism 21 and a heating furnace 22.

[0014] The fuel supply mechanism 21 supplies fuel to the heating furnace 22. The fuel supply mechanism 21 comprises a pipe 21A connected to the heating furnace 22 and a valve 21B located in the middle of the pipe 21A. The pipe 21A forms a supply path that supplies fuel from a fuel tank (not shown) to the heating furnace 22. The valve 21B opens and closes the supply path formed in the pipe 21A. The valve 21B adjusts the flow rate of fuel supplied to the heating furnace 22 by the degree of its opening. A flow meter S1 for measuring the flow rate of fuel supplied to the heating furnace 22 is provided in the middle of the pipe 21A.

[0015] The heating furnace 22 heats the fluid to be processed, such as crude oil, supplied from the piping L1 connected to the heating furnace 22. The heating furnace 22 heats the fluid to be processed by burning fuel supplied from the fuel supply mechanism 21. After heating, the fluid to be processed is supplied from the piping L2 connected to the heating furnace 22 to a distillation column (not shown) for processing. A thermometer S2 is installed in the middle of the piping L2 to measure the temperature of the heated fluid to be processed flowing through the piping L2. This temperature can also be called the outlet temperature of the heating furnace 22, and will be referred to as the outlet temperature below.

[0016] In the controlled device 20 configured as described above, the flow rate of fuel to the heating furnace 22 is adjusted by the opening degree of valve 21B, thereby adjusting the amount of fuel burned in the heating furnace 22, i.e., the amount of heating of the fluid to be processed. As a result, the temperature of the fluid to be processed after heating, i.e., the outlet temperature, is adjusted. The operation of the controlled device 20 is controlled by controller 30. Controller 30 uses the outlet temperature measured by thermometer S2 as a feedback value and controls the opening degree of valve 21B so that this outlet temperature becomes a preset target value (hereinafter also referred to as the temperature target value).

[0017] The controller 30 includes a primary controller 31 and a secondary controller 32, and performs feedback control via cascade control on the controlled object 20. The controller 30 is configured of, for example, a DCS (Distributed Control System), a PLC (Programmable Logic Controller), or a computer. The controller 30 may be a combination of a plurality of computers that perform distributed processing. The primary controller 31 and the secondary controller 32 may be configured as separate devices (PLC, computer).

[0018] The primary controller 31 uses the outlet temperature measured by the thermometer S2 as a feedback value, derives a target value of fuel flow rate to the heating furnace 22 (hereinafter also referred to as a flow rate target value) that matches this outlet temperature with the target temperature, and performs feedback control (here, PID (Proportional-Integral-Differential) control) that supplies the derived flow rate target value to the secondary controller 32. The flow rate target value is a manipulated variable for operating the controlled object 20 in the feedback control performed by the primary controller 31.

[0019] The secondary controller 32 uses the fuel flow rate measured by the flowmeter S1 as a feedback value, and performs feedback control (here, PID control) that operates the valve 21B so that this flow rate matches the flow rate target value from the primary controller 31. The operation of the valve 21B is performed by supplying a manipulated variable indicating the opening degree of the valve 21B to the valve 21B. The manipulated variable is, for example, a value obtained by adding (or subtracting) the change in opening degree derived through feedback control to the current opening degree of the valve 21B. The valve 21B operates such that its own opening degree becomes the input manipulated variable, that is, the input opening degree.

[0020] The model identification apparatus 10 is configured of various computers such as a personal computer. The model identification apparatus 10 comprises: a processor 11 such as a CPU (Central Processing Unit); a storage 12, which is a non-volatile storage device that stores a program 12P executed or used by the processor 11 and various data; and a main memory 13 such as a RAM (Random Access Memory) that provides a work area for the processor 11. The processor 11 may be configured to include an ASIC (Application Specific Integrated Circuit) and / or an FPGA (Field-Programmable Gate Array). The program 12P only needs to be stored in a computer-readable non-transitory medium such as the storage 12. The model identification apparatus 10 further comprises: an operation device 14 to which a user's operation is input; a display device 15 that displays various images; and an I / O (Input / Output) 16 that relays information (signals) transmitted and received between the processor 11 and the outside of the model identification apparatus 10.

[0021] By executing the program 12P, the processor 11 operates as the models M1 to M3, the first acquisition unit 11A, the second acquisition unit 11B, the machine learning unit 11C, and the tuning unit 11D shown in FIG. 2.

[0022] Models M1 to M3 are models that model the controlled object 20. Models M1 to M3 take a flow rate target value from the primary controller 31 as input and output a predicted value of the outlet temperature as a response to the input flow rate target value. The transfer function G1(s) of model M1 is given by equation (1) below. The transfer function G2(s) of model M2 is given by equation (2) below. The transfer function G3(s) of model M3 is given by equation (3) below. K1 to K3 are gains, and τ1 to τ3 are time constants. K1 to K3 and τ1 to τ3 are model parameters that characterize each of the models M1 to M3. Specifically, the model parameter MP1 of model M1 includes two parameters: K1 and τ1. The model parameter MP2 of model M2 includes two parameters: K2 and τ2. The model parameter MP3 of model M3 includes two parameters: K3 and τ3. Models M1 to M3 are defined as first-order lag systems. Each of these parameters is stored in storage 12 and used. In equations (1) to (3), "*" indicates multiplication (and so on in the other equations). G1(s) = K1 / (1 + τ1*s) ... (1) G2(s) = K2 / (1+τ2*s) ... (2) G3(s) = K3 / (1+τ3*s) ... (3)

[0023] Model parameters MP1 to MP3 are repeatedly updated by machine learning, as described below. The storage 12 stores the K1 to K3 and τ1 to τ3 of each of the repeatedly updated historical model parameters MP1 to MP3 as time series data. The most recent K1 to K3 and τ1 to τ3 from the time series data are used for models M1 to M3. Each of the model parameters MP1 to MP3 has different learning conditions. The learning conditions include conditions that limit the numerical range that a particular type of parameter can take (the numerical range that can be updated by machine learning, as described below) to a specific subset of ranges. For example, the numerical range that one or two of K1 to K3 can take may be limited to a certain range (which may be a range consisting of multiple numbers or a range consisting of a single number). The remaining numerical range of K1 to K3 may be limited to other certain ranges, or there may be no limitation on the numerical range. For example, the numerical range of K1 may be fixed to a first value, the numerical range of K2 may be set to a range greater than or equal to a second value, and the numerical range of K3 may not be limited. For example, the numerical range for K1 could be set to be greater than or equal to the third value and less than the fourth value, the numerical range for K2 could be set to be greater than or equal to the fourth value and less than the fifth value, and K3 could be fixed at the fifth value. This could also be applied to τ1 to τ3.

[0024] The first acquisition unit 11A sequentially acquires the outlet temperature (control amount) periodically measured by the thermometer S2 during feedback control of the controlled object 20, that is, while the controlled object 20 is in operation. This outlet temperature will also be referred to as the outlet temperature T below. The first acquisition unit 11A stores the sequentially acquired outlet temperatures T in the storage 12 in chronological order. As a result, the storage 12 stores chronological data of the outlet temperatures T. The storage destination for the outlet temperatures T may be the main memory 13 instead of the storage 12. The same applies to other data.

[0025] The second acquisition unit 11B sequentially acquires the flow rate target values ​​(operated variables) that the primary controller 31 periodically outputs to the secondary controller 32 during feedback control of the controlled object 20, that is, while the controlled object 20 is in operation. The second acquisition unit 11B inputs each sequentially acquired flow rate target value to each of the models M1 to M3 and acquires the predicted value of the outlet temperature (controlled variable) T, which is the output of each model. Hereinafter, the predicted value output from model M1 will also be called predicted value Tm1, the predicted value output from model M2 will be called predicted value Tm2, ​​and the predicted value output from model M3 will be called predicted value Tm3. The second acquisition unit 11B sequentially stores the sequentially acquired predicted values ​​Tm1 to Tm3 in the storage 12. As a result, the storage 12 stores the time-series data of the predicted values ​​Tm1 to Tm3.

[0026] The first acquisition unit 11A and the second acquisition unit 11B operate synchronously with each other, sequentially acquiring the outlet temperature T and predicted values ​​Tm1 to Tm3 at the same time. Therefore, the outlet temperature T and predicted values ​​Tm1 to Tm3 stored in the storage 12 at the same time correspond to a common manipulated variable. As another example, the first acquisition unit 11A may acquire the manipulated variable and the outlet temperature T, and supply the acquired manipulated variable and outlet temperature T to the second acquisition unit 11B. In this case, the second acquisition unit 11B holds the manipulated variable and outlet temperature T from the first acquisition unit 11A, and inputs the manipulated variable to each of the models M1 to M3 to acquire predicted values ​​Tm1 to Tm3. The second acquisition unit 11B stores the acquired predicted values ​​Tm1 to Tm3 and the outlet temperature T in the storage 12, associating them with each other. As a result, pairs of outlet temperature T and predicted values ​​Tm1 to Tm3 at the same time are stored in the storage 12 in chronological order.

[0027] The machine learning unit 11C performs machine learning periodically. For example, as machine learning, the machine learning unit 11C continuously or autonomously searches for (learns) model parameters using an exploratory algorithm in response to a given evaluation function or reward. As an example, as shown in Figure 3(A), the machine learning unit 11C searches for model parameters using an exploratory algorithm based on an evaluation function, reward, etc., that minimizes the difference between the time series data of the outlet temperature T and the predicted value Tm1 stored in the storage 12 for a certain period (window interval) 101 going back from the present, i.e., the difference between the outlet temperature T and the predicted value Tm1 during that certain period 101. The unit periodically repeats an update process to update the model parameters MP1 (i.e., K1 and τ1) with the searched parameter values. The difference may be the difference between each time series data of the outlet temperature T and the predicted value Tm1, the difference at a predetermined timing, or the difference in various statistical values ​​of the outlet temperature T and the predicted value Tm1. Examples of machine learning using such an exploratory algorithm include Bayesian optimization, evolutionary algorithms, and reinforcement learning. The dotted rectangle representing a fixed period 102 in Figure 3(A) is the fixed period in the update process preceding fixed period 101. The update is performed under learning conditions. As described above, if the range of values ​​that can be taken in the update is limited to a certain range by the learning conditions, the update process updates the values ​​within that certain range. The model parameters MP1 K1 and τ1, which are updated periodically during the update process, are sequentially stored in the storage 12, forming time-series data showing the time change of K1 and time-series data showing the time change of τ1. The machine learning unit 11C also periodically repeats this update process for models M2 and M3 (model parameter MP1 becomes model parameter MP2 or MP3, Tm1 becomes Tm2 or Tm3, and K1 and τ1 become K2 and τ2, or K3 and τ3).

[0028] As shown in Figure 3(A), at the beginning of updating the model parameters MP1 (K1 and τ1) of model M1 (see the period on the left side of the graph in Figure 3(A)), the difference between the outlet temperature T and the predicted value Tm1 is large, and therefore, the prediction accuracy of the predicted value Tm1 of model M1 is not good. Subsequently, as the updates of K1 and τ1 progress, the difference between the outlet temperature T and the predicted value Tm1 decreases, and the prediction accuracy of model M1 improves. In other words, by repeatedly updating K1 and τ1, the machine learning unit 11C can identify model M1, which is a model of the controlled object 20 with good prediction accuracy. Similarly, by repeatedly updating the model parameters MP2 (K2 and τ2) and model parameters MP3 (K3 and τ3), respectively, the machine learning unit 11C can identify models M2 and M3, which are also good predictive models. Note that the learning conditions for models M1 to M3 are different as described above. Therefore, the time it takes to identify a model with good prediction accuracy for the predicted value may differ for models M1 to M3. In this embodiment, among models M1 to M3, the model that shows the fastest improvement in prediction accuracy is adopted as the model for the controlled object 20, and the control parameters of the PID control of the primary controller 31 are tuned based on the model parameters of that model. This tuning is performed by the tuning unit 11D. The tuning is performed using a known tuning method, such as the IMC-PID tuning rule.

[0029] The tuning unit 11D monitors the prediction accuracy of each model M1 to M3 and detects the model among M1 to M3 whose prediction accuracy has exceeded a predetermined standard. An example of this detection is described below.

[0030] The tuning unit 11D monitors the time-series data of the outlet temperature T and predicted values ​​Tm1 to Tm3 in the storage 12, and each time new outlet temperature T and predicted values ​​Tm1 to Tm3 are stored in the storage 12, it sequentially derives the difference ΔT1 between the outlet temperature T and predicted value Tm1, the difference ΔT2 between the outlet temperature T and predicted value Tm2, ​​and the difference ΔT3 between the outlet temperature T and predicted value Tm3. The differences ΔT1 to ΔT3 may be the difference between the outlet temperature T itself and the predicted values ​​Tm1 to Tm3 themselves, but as another example, they may be derived as the difference between the filtered value of the outlet temperature T and the filtered values ​​of each of the predicted values ​​Tm1 to Tm3. As an example, the differences ΔT1 to ΔT3 may be derived as the difference between the moving average of the outlet temperature T and the moving averages of each of the predicted values ​​Tm1 to Tm3. The tuning unit 11D detects the model among the models M1 to M3 that outputs the predicted value of the difference ΔT1 to ΔT3 if the difference falls below a preset threshold Th1 for n consecutive times (where n is an integer greater than or equal to 1), or if there is a difference. The tuning unit 11D then derives the control parameters for the PID control of the primary controller 31 based on the current model parameters of the detected model. For example, if the tuning unit 11D detects model M1 as the difference ΔT1, it derives the control parameters for the PID control of the primary controller 31 based on K1 and τ1 of this model M1. The method for deriving the control parameters will be described later.

[0031] As shown in Figures 3(A) to (C), as the difference between the outlet temperature T and the predicted value Tm1 decreases, K1 and τ1 also converge to a certain value. The same applies to models M2 and M3. Therefore, as another specific example of model detection, the tuning unit 11D may monitor the time-series data of K1 to K3 and τ1 to τ3 included in the model parameters MP1 to MP3 of the storage 12. In this case, each time new K1 to K3 and τ1 to τ3 are stored in the storage 12, the tuning unit 11D derives the difference ΔK1 to ΔK3 and Δτ1 to Δτ3, which are the difference between the latest data (which may be a moving average) and the previous or multiple previous data (which may be a moving average). When the derived ΔK1 and Δτ1, ΔK2 and Δτ2, and ΔK3 and Δτ3 fall below the preset thresholds Th2 and Th3 for n consecutive times, the tuning unit 11D detects the set of models that fell below the thresholds as models with high prediction accuracy. For example, the tuning unit 11D detects model M2 as a model with high prediction accuracy when the sequentially derived ΔK2 and Δτ2 fall below the preset thresholds Th3 and Th4 for n consecutive times, respectively. Alternatively, the tuning unit 11D may calculate the standard deviations of the differences ΔK1 to ΔK3 and Δτ1 to Δτ3 based on the time series data for the most recent predetermined period, and detect a model with a low standard deviation as a model with high prediction accuracy. The tuning unit 11D derives the control parameters for the PID control of the primary controller 31 based on the current model parameters of the detected model.

[0032] The tuning unit 11D may also detect a model specified by the user of the model identification device 10 (the person tuning the PID control parameters). For example, the user compares the graphs of the time series data of the predicted values ​​Tm1 to Tm3 for models M1 to M3, displayed on the display device 15 of the model identification device 10, with the graph of the time series data of the outlet temperature T. These graphs are displayed by the tuning unit 11D. The user identifies a model with high prediction accuracy by comparing the graphs. The user specifies the identified model via the operating device 14. The tuning unit 11D may detect the model specified by the user as a model with high prediction accuracy. After performing the detection of a model with high prediction accuracy based on the user's model specification multiple times, the tuning unit 11D may proceed to the automatic detection of a model with high prediction accuracy. In this case, the trend of the model specified by the user may be learned through machine learning, and automatic detection may be performed using the machine learning results.

[0033] Here, the control parameters for the PID control of the primary controller 31 are described. The primary controller 31 performs PID control using, for example, the calculation formula shown in equation (4) below. The control parameters for PID control include the parameters Kp, Ti, and Td. As described above, models M1 to M3 are first-order lag systems, and the tuning unit 11D derives Kp using equation (5). In equation (5), τ and K become τ1 and K1 when the identified model is model M1, τ2 and K2 when the identified model is model M2, and τ3 and K3 when the identified model is model M3. Tcl is a value equivalent to the time constant in the closed-loop response of the feedback control by the primary controller 31. This value is set by the user of the model identification device 10 (the person tuning the PID control parameters). This value determines the strength of the PID control. Ti is the integral time constant, where Ti=τ1 when the identified model is model M1, Ti=τ2 when it is model M2, and Ti=τ3 when it is model M3. Td is the differential time constant, where Td=0. The PID control formula is arbitrary and can be a more practical formula derived from equation (4). Also, as described later, models M1 to M3 may be models other than first-order lag systems. In that case, the PID control formula will also be changed. Formulas for deriving the control parameters are also provided according to the PID control formula and model. C(s)=Kp{1+1 / (Ti*s)+Td*s} ··· (4) Kp = τ / K * Tcl ... (5)

[0034] As described above, the control parameters for PID control (here, parameters Kp, Ti, and Td) can be derived based on the model parameters of any of the models M1 to M3 detected as models with high prediction accuracy. For example, if the tuning unit 11D detects model M1 as a model with high prediction accuracy, it derives the control parameters (Kp, Ti, and Td) based on the current model parameters MP1 (K1 and Δτ1) of model M1. The tuning unit 11D sets the derived control parameters (Kp, Ti, and Td) to the primary controller 31. As a result, these new parameters are set as the control parameters for feedback control by the primary controller 31; in other words, the control parameters used by the primary controller 31 for feedback control are tuned.

[0035] If the tuning unit 11D automatically detects a model with high prediction accuracy as described above, it may display information about that model (for example, a graph of time-series data of the model's predicted values ​​and a graph of time-series data of the outlet temperature T) on the display device 15 of the model identification device 10. In this case, the user looks at the displayed information and inputs an instruction via the operating device 14 whether or not to derive control parameters, that is, to tune the control parameters as described later. The tuning unit 11D may perform tuning of the control parameters on the condition that an instruction to perform tuning has been received.

[0036] As described above, the model identification device 10 according to this embodiment comprises a first acquisition unit 11A, a second acquisition unit 11B, and a machine learning unit 11C. The first acquisition unit 11A sequentially acquires the outlet temperature T (controlled variable), which is the response of the controlled object 20 to the flow rate target value (operated variable) of the feedback control when the controlled object 20 is operated by feedback control. The second acquisition unit 11B sequentially acquires N predicted values ​​Tm1 to Tm3 of the outlet temperature T from each of the N models M1 to M3 (here, N=3, but N can be an integer of 2 or more) that model the controlled object 20, when the flow rate target value is input to each of the N models M1 to M3. Each of the N models M1 to M3 is characterized by model parameters MP1 to MP3, each having different learning conditions. The machine learning unit 11C repeatedly updates N model parameters MP1 to MP3 that characterize N models M1 to M3, each under the learning conditions of those model parameters, based on the difference (difference between time series data) between the sequentially acquired outlet temperature T (this time series data in the above example) and N sequentially acquired predicted values ​​Tm1 to Tm3 (these time series data in the above example). For example, the machine learning unit 11C identifies model M1 by repeatedly updating the model parameter MP1 that characterizes model M1, each under the learning conditions of that model parameter MP1, based on the difference (difference between time series data) between the sequentially acquired outlet temperature T and the sequentially acquired predicted value Tm1 during a predetermined period.

[0037] With this configuration, the model parameters MP1 to MP3 of models M1 to M3 are repeatedly updated by machine learning based on the difference between the outlet temperature T and N predicted values ​​Tm1 to Tm3. As a result, models M1 to M3 are identified in parallel with the operation of the controlled object 20 (operation that achieves the original purpose of the controlled object 20). This gradually improves the model parameters MP1 to MP3, and models M1 to M3 with good prediction accuracy are identified. Therefore, according to this embodiment, the identification of models M1 to M3 of the controlled object 20 can be performed accurately in parallel with the operation of the controlled object 20, that is, without stopping the operation of the controlled object 20.

[0038] While the specific examples of machine learning are arbitrary, as mentioned above, machine learning may include search-type algorithms such as Bayesian optimization, evolutionary algorithms, and reinforcement learning. In this embodiment, as a search-type algorithm, learning is performed to update the model parameters MP1 to MP3 so as to minimize the difference between the sequentially acquired time series data of outlet temperature T and each of the N sequentially acquired time series data of predicted values ​​Tm1 to Tm3. Through this learning, the model parameters MP1 to MP3 are repeatedly updated to minimize the aforementioned difference, thereby identifying models M1 to M3 that model with greater accuracy.

[0039] As described above, the learning condition for the model parameters of the first model among the N models M1 to M3 can be set to limit the numerical range of a predetermined type of parameter among the model parameters of this first model to a first range. Furthermore, the learning condition for the model parameters of the second model among the N models M1 to M3 can be set to limit the numerical range of the predetermined type of parameter of this second model to a second range different from the first range, or to leave the numerical range unrestricted. With such a configuration, various models can be prepared, making it easier to obtain a model with good accuracy.

[0040] In addition to or instead of the above limitations on numerical ranges, the N models M1 to M3 may have different model structures (i.e., different types of model parameters). For example, each of models M1 to M3 may be any of the following: a first-order lag system model, a first-order lag + dead time system model, or a second-order lag system model. For example, models M1 and M2 may be first-order lag systems, and model M2 may be a first-order lag + dead time system model. The type of one or more parameters that make up the model parameters will differ depending on the system of the model. For example, in a first-order lag + dead time system model, dead time L is also included in the model parameters. Different types of parameters mean that different parameters are updated, and therefore the above learning conditions are also different. Thus, the learning conditions for the model parameters of the third model and the learning conditions for the model parameters of the fourth model among the N models may be for updating different types of model parameters. Another example of such a third and fourth model is a model with the same model structure but different updatable parameters. The types of parameters and methods for deriving them for feedback control may also vary depending on the system.

[0041] As described above, the model identification device 10 includes a tuning unit 11D. The tuning unit 11D detects one of the N models M1 to M3 that has output a predicted value among the N predicted values ​​Tm1 to Tm3 whose prediction accuracy is higher than a predetermined standard, and tunes the control parameters of the feedback control by the primary controller 31 based on the current model parameters of the detected model. As a result, accurate feedback control is performed after tuning.

[0042] As described above, the tuning unit 11D derives new control parameters for feedback control based on the current model parameters of the detected model, and sets the derived new control parameters in the primary controller 31, thereby tuning the control parameters of the feedback control by the primary controller 31. In this way, accurate feedback control is performed after tuning.

[0043] Models M1 to M3 may be divided into reference models and control parameter update models. For example, the tuning unit 11D may periodically tune the control parameters of the feedback control by the primary controller 31 based on the model parameter MP1 of model M1 (a predetermined model for control parameter update) among the models M1 to M3. Furthermore, the tuning unit 11D presents the information of other models, models M2 and M3, which are reference models, to the user by displaying the information on the display device 15. The presented information includes at least one of the following: (1) predicted values ​​Tm2 and Tm3 output by models M2 and M3, which are reference models, and acquired by the second acquisition unit 11B (for example, trend data (time series data) for a period going back a predetermined time from the present); and (2) the difference between the predicted values ​​Tm2 and Tm3 output by models M2 and M3 and acquired by the second acquisition unit 11B and the outlet temperature T (control quantity) acquired by the first acquisition unit 11A (for example, trend data (time series data) for a period going back a predetermined time from the present). Regarding (1) above, a graph of at least one time series data of predicted values ​​Tm1 to Tm3 (predicted values ​​of the control parameter update model and predicted values ​​of other models) and a graph of time series data of the outlet temperature T may be presented in a comparable manner (for example, superimposed display with the same vertical and horizontal axes). The graphs presented in a comparable manner should include a graph of the predicted values ​​output by the current control parameter update model and a graph of the predicted values ​​output by at least one reference model. This allows the user to intuitively understand whether the prediction accuracy of the reference model is better or worse than the prediction accuracy of the current control parameter update model. The user may specify which of the predicted values ​​Tm1 to Tm3 to present. Figure 4 shows a graph of time series data of predicted values ​​Tm1 to Tm3 and the outlet temperature T at a different time than that shown in Figure 3. Regarding (2) above, at least one of the time series data of the difference between each of the time series data of the predicted values ​​Tm1 to Tm3 (predicted values ​​of the model for updating control parameters and predicted values ​​of other models) and the time series data of the outlet temperature T may be presented, or at least two may be presented in a comparable manner (for example, superimposed display with the vertical and horizontal axes being common).The graphs presented for comparison should preferably include graphs showing the difference between the predicted value output by the current control parameter update model and the outlet temperature T, and graphs showing the difference between the predicted value output by at least one reference model and the outlet temperature T. This allows the user to intuitively understand whether the prediction accuracy of the reference model is better or worse than that of the current control parameter update model. The user may specify which differences to present. The differences may be the difference between the two at a predetermined timing, or the difference in statistical values ​​(means) based on the time-series data of both over a certain period. The user looks at the presented information and, for example, identifies a model M2 or M3 (in the example in Figure 4, model M2 outputting the predicted value Tm2) that appears to have better prediction accuracy than the current control parameter update model, and specifies that model via the operating device 14 as the model to be used for tuning the control parameters of the feedback control. When the tuning unit 11D detects that the user has specified model M2 or M3, it may begin periodically tuning the control parameters of the feedback control based on the model parameters MP2 or MP3 of model M2 or M3 specified by the user, instead of the model parameter MP1 of model M1. This allows the control parameters to be suitably tuned using the model selected by the user. However, the above periodic tuning is not required. Tuning using the model parameters of the model used for updating the control parameters may be performed at any time specified by the user. Furthermore, when the user selects one of the reference models as described above, the user or the system may, at any time thereafter, use the selected reference model as the model for updating the control parameters and perform tuning using the model parameters of that model.

[0044] The controlled object 20 can be anything that is subject to feedback control. For example, the controlled object 20 may be configured as an oil plant with crude oil as the fluid to be processed, or it may be configured as another plant. Furthermore, the feedback control may be PI control. In addition, the feedback control to be tuned may be feedback control by a secondary controller 32, etc. The model is also constructed to match the feedback control. Thus, the object of the model (i.e., the controlled object) and the feedback control to be tuned using the model are arbitrary and are not limited to the examples of the above embodiment.

[0045] The number of models can be any number other than 3. For example, the number of models can be just one, model M1. In this case, the first acquisition unit 11A, the second acquisition unit 11B, the machine learning unit 11C, and the tuning unit 11D perform the following processing. The effects described above can also be obtained as appropriate with this configuration. There may or may not be limitations on the numerical range of the multiple parameters that constitute the model parameters of model M1.

[0046] The first acquisition unit 11A sequentially acquires the outlet temperature T, which is the response of the controlled object 20 to the flow rate target value, which is the manipulated variable of the feedback control, when the controlled object 20 is operated by feedback control. The second acquisition unit 11B sequentially acquires the predicted value Tm1 of the outlet temperature T, which is the output of model M1 when the flow rate target value is input to model M1, a model of the controlled object 20. The machine learning unit 11C identifies model M1 by repeatedly updating model parameters MP1 that characterize model M1 using machine learning (such as reinforcement learning) based on the difference between the sequentially acquired outlet temperature T and the predicted value Tm1. The tuning unit 11D may tune the control parameters of the feedback control based on the current model parameters MP1 of model M1 when it detects that the prediction accuracy of the predicted value Tm1 of model M1 has become higher than a predetermined standard.

[0047] Furthermore, in addition to feedback control, feedforward control may be performed. In such a case, a feedforward controller may be provided that controls the manipulated variable in accordance with observed disturbances (for example, changes in the amount of crude oil supplied to the heating furnace 22). These observed disturbances also affect the outlet temperature T. The model identification device 10 also forms a model of these observed disturbances, and the output of this model is added to the predicted values ​​Tm1 to Tm3.

[0048] If the control quantity, such as the outlet temperature T, is greatly affected by noise, the control quantity from the controlled object 20 (here, the heating furnace 22) may be acquired by the first acquisition unit 11A as a filtered control quantity via a predetermined noise filter F shown in Figure 5. The machine learning unit 11C updates the model parameters MP1 to MP3 of models M1 to M3 by machine learning in the same manner as described above, based on the output of models M1 to M3 acquired by the second acquisition unit 11B, i.e., the predicted values ​​of the control quantities, and the filtered control quantities. When the tuning unit 11D tunes the control parameters of the PID control based on the model parameters of the model with the highest prediction accuracy among models M1 to M3, it applies the inverse filter F of the noise filter F to the model with the highest prediction accuracy. -1 The formula is applied (for example, by multiplication). The tuning unit 11D tunes the control parameters of the feedback control based on the model parameters of the model after the formula has been applied.

[0049] (Scope of the present invention) The present invention is not limited to the embodiments described above. For example, the present invention includes various modifications to the above embodiments and modifications that can be understood by those skilled in the art within the scope of the technical concept of the present invention. The configurations listed in the above embodiments and modifications can be combined as appropriate to the extent that they do not contradict each other. It is also possible to delete any of the above configurations.

[0050] (Note) Configurations based on the above embodiments and modifications are described below as examples. Any partial configuration of the above embodiments and modifications may be applied to each of these configurations. Furthermore, parts of each configuration may be combined. (Note 1) A first acquisition unit sequentially acquires control quantities that are the response of the controlled object to the manipulated quantity of the feedback control when the controlled object is being operated by feedback control, A second acquisition unit sequentially acquires the predicted value of the controlled variable, which is the output of the model when the manipulated variable is input to the model that models the controlled object. A machine learning unit identifies the model by repeatedly updating the model parameters that characterize the model using machine learning, based on the difference between the sequentially acquired control quantities and the predicted values. A model identification device equipped with the following features. (Note 2) The system further includes a tuning unit that, when it detects that the prediction accuracy of the predicted value of the model has become higher than a predetermined standard, tunes the control parameters of the feedback control based on the current model parameters of the model. The model identification device described in Appendix 1. (Note 3) The tuning unit tunes the control parameters of the feedback control by deriving new control parameters for the feedback control based on the current model parameters and setting the derived new control parameters to the controller that performs the feedback control. The model identification device described in Appendix 2. (Note 4) The second acquisition unit sequentially acquires N predicted values ​​of the control variable, which are the responses of each of the N models (where N is an integer of 2 or more) that model the controlled object, including the aforementioned model, when the manipulated variable is input to each of the N models. Each of the N models is characterized by model parameters having different learning conditions. The machine learning unit identifies each of the N models by repeatedly updating the N model parameters that characterize each of the N models using machine learning, based on the difference between the controlled variable and each of the N predicted values, under the learning conditions for those model parameters. A model identification device as described in any of the appendices 1 to 3. (Note 5) The learning condition for the model parameters of the first model among the N models is to limit the numerical range of a predetermined type of parameter among the model parameters of the first model to a first range. The learning condition for the model parameters of the second model among the N models is to either limit the numerical range of the predetermined type of parameters of the second model to a second range different from the first range, or not limit the numerical range. The model identification device described in Appendix 4. (Note 6) Of the N models mentioned above, the learning conditions for the model parameters of the third model and the learning conditions for the model parameters of the fourth model are updated using different types of model parameters. A model identification device as described in Appendix 4 or 5. (Note 7) The third model and the fourth model have different model structures. The model identification device described in Appendix 6. (Note 8) The system further includes a tuning unit that detects one of the N models that outputs a predicted value whose prediction accuracy is higher than a predetermined standard, and tunes the control parameters of the feedback control based on the current model parameters of the detected model. A model identification device as described in any of the appendices 4 to 7. (Note 9) The system includes a tuning unit that periodically tunes the control parameters of the feedback control based on the model parameters of a predetermined model among the N models, The aforementioned tuning unit is The user is presented with at least one of the following: the predicted value output by a model other than the predetermined model among the N models and acquired by the second acquisition unit, and the difference between the predicted value output by the other model and acquired by the second acquisition unit and the control quantity acquired by the first acquisition unit. When it is detected that the user has specified the other model, the control parameters of the feedback control are periodically tuned based on the model parameters of the other model instead of the model parameters of the predetermined model. A model identification device as described in any of the appendices 4 to 8. (Note 10) The machine learning described above is a learning process that updates the model parameters to minimize the difference between the time series data of the controlled variable acquired sequentially over a predetermined period and the time series data of the predicted value acquired sequentially over the predetermined period. A model identification device as described in any of the appendices 1 to 9. (Note 11) A program that causes a computer to function as a model identification device as described in one of the appendices 1 to 10. (Note 12) A first acquisition step involves sequentially acquiring control quantities, which are the response of the controlled object to the manipulated quantity of the feedback control when the controlled object is being operated by feedback control, A second acquisition step involves sequentially acquiring predicted values ​​of the controlled variable, which are the output of the model when the manipulated variable is input to the model that models the controlled object. A machine learning step to identify the model by repeatedly updating the model parameters that characterize the model using machine learning based on the difference between the sequentially acquired control variables and the predicted values, A model identification method having the following characteristics. Furthermore, this model identification method is also an invention of a method for producing a model, or a device capable of functioning as such a model. [Explanation of symbols]

[0051] 10...Model identification device, 11...Processor, 11A...First acquisition unit, 11B...Second acquisition unit, 11C...Machine learning unit, 11D...Tuning unit, 12...Sequential storage, 12...Storage, 12P...Program, 13...Main memory, 14...Operating device, 15...Display device, 20...Controlled object, 21...Fuel supply mechanism, 21A...Piping, 21B...Valve, 22...Furnace, 30...Controller, 31...Primary controller, 32...Secondary controller, 101...Fixed period (window section), 102...Fixed period, F...Noise filter, L1...Piping, L2...Piping, M1...Model, M2...Model, M3...Model, S1...Flow meter, S2...Thermometer.

Claims

1. A first acquisition unit sequentially acquires control quantities that are the response of the controlled object to the manipulated quantity of the feedback control when the controlled object is being operated by feedback control, A second acquisition unit sequentially acquires the predicted value of the control variable, which is the output of the model when the manipulated variable is input to the model that models the control target, A machine learning unit identifies the model by repeatedly updating the model parameters that characterize the model using machine learning, based on the difference between the sequentially acquired control quantities and the predicted values. A model identification device equipped with the following features.

2. The system further includes a tuning unit that, when it detects that the prediction accuracy of the predicted value of the model has become higher than a predetermined standard, tunes the control parameters of the feedback control based on the current model parameters of the model. The model identification device according to claim 1.

3. The tuning unit tunes the control parameters of the feedback control by deriving new control parameters for the feedback control based on the current model parameters and setting the derived new control parameters to the controller that performs the feedback control. The model identification device according to claim 2.

4. The second acquisition unit sequentially acquires N predicted values ​​of the control variable, which are the responses of each of the N models (where N is an integer of 2 or more) that model the controlled object, including the model mentioned above, when the manipulated variable is input to each of the N models. Each of the N models is characterized by model parameters having different learning conditions. The machine learning unit identifies each of the N models by repeatedly updating the N model parameters that characterize each of the N models through machine learning, based on the difference between the controlled variable and each of the N predicted values, under the learning conditions for those model parameters. The model identification device according to claim 1.

5. The learning condition for the model parameters of the first model among the N models is to limit the numerical range of a predetermined type of parameter among the model parameters of the first model to a first range. The learning condition for the model parameters of the second model among the N models is to either limit the numerical range of the predetermined type of parameters of the second model to a second range different from the first range, or not limit the numerical range. The model identification device according to claim 4.

6. Of the N models mentioned above, the learning conditions for the model parameters of the third model and the learning conditions for the model parameters of the fourth model are updated using different types of model parameters. The model identification device according to claim 4.

7. The third model and the fourth model have different model structures. The model identification device according to claim 6.

8. The system further includes a tuning unit that detects one of the N models that outputs a predicted value whose prediction accuracy is higher than a predetermined standard, and tunes the control parameters of the feedback control based on the current model parameters of the detected model. The model identification device according to claim 4.

9. The system includes a tuning unit that periodically tunes the control parameters of the feedback control based on the model parameters of a predetermined model among the N models, The aforementioned tuning unit is The user is presented with at least one of the following: the predicted value output by a model other than the predetermined model among the N models and acquired by the second acquisition unit, and the difference between the predicted value output by the other model and acquired by the second acquisition unit and the control quantity acquired by the first acquisition unit. When it is detected that the user has specified the other model, the control parameters of the feedback control are periodically tuned based on the model parameters of the other model instead of the model parameters of the predetermined model. The model identification device according to claim 4.

10. The machine learning described above is a learning process that updates the model parameters to minimize the difference between the time series data of the controlled variable acquired sequentially over a predetermined period and the time series data of the predicted value acquired sequentially over the predetermined period. The model identification device according to claim 1.

11. A program that causes a computer to function as the model identification device described in claim 1.

12. A first acquisition step involves sequentially acquiring control quantities, which are the response of the controlled object to the manipulated quantity of the feedback control when the controlled object is being operated by feedback control, A second acquisition step involves sequentially acquiring predicted values ​​of the controlled variable, which are the output of the model when the manipulated variable is input to the model that models the controlled object. A machine learning step to identify the model by repeatedly updating the model parameters that characterize the model using machine learning based on the difference between the sequentially acquired control variables and the predicted values, A model identification method having the following characteristics.

Images