Control device, method and program

Abstract

To quickly and accurately suppress operation fluctuation that can occur during disturbance generation in process control.SOLUTION: An automatic control system of a blast furnace process 1 includes a controller 2 and a disturbance observer 3, and executes model prediction control using a state space model representing a time change of a state variable x for an operation amount u, targeted on the blast furnace process 1 to be controlled. With the operation amount u as input, the blast furnace process 1 outputs a control amount z and an observation amount y including the control amount z, and input disturbance d1 is superposed on the operation amount u and output disturbance d2 is superposed on the observation amount y. With the operation amount u and the observation amount y on which the output disturbance d2 is superposed as input, the disturbance observer 3 outputs an estimated value x^ of the state variable, an estimated value d1^ of input disturbance, and an estimated value d2^ of the output disturbance. With a target value r and the output of the disturbance observer 3 as output, the controller 2 executes model prediction control to thereby output the operation amount u.SELECTED DRAWING: Figure 1

Description

[Technical Field]

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

[0002] In recent years, the use of inexpensive iron ore and production under low reducing agent ratio conditions have become popular in blast furnace operation. However, fluctuations in the quality of the raw materials charged to the blast furnace due to the use of inexpensive, low-quality raw materials can cause disturbances to blast furnace operation. As a result, fluctuations in blast furnace operation (hereinafter sometimes simply referred to as "operational fluctuations"), such as fluctuations in production volume and decreases in furnace heat, can lead to unstable furnace conditions. To achieve stable blast furnace operation, it is necessary to quickly suppress operational fluctuations that may occur when disturbances occur.

[0003] As technologies relating to blast furnace operation, Patent Documents 1 and 2 disclose a method for predicting the temperature of molten iron in a blast furnace using a physical model that can calculate the state inside the blast furnace in an unsteady state. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-24935 [Patent Document 2] Japanese Patent Application Publication No. 2019-19385 [Non-patent literature]

[0005] [Non-Patent Document 1] Adachi: Advanced System Identification for Control Using MATLAB, Tokyo Denki University Press, 2004 [Non-patent document 2] Hagiwara: Introduction to Digital Control, Corona Publishing, p.46, 1999 [Non-patent document 3] B.Moore:Principal component analysis in linear systems: Controllability, observability, and model reduction , IEEE Transactions on Automatic Control (Volume: 26, Issue: 1, February 1981) [Non-patent document 4] Ando et al.: Control System Design Using Numerical Analysis Methods, Corona Publishing, 1986 [Non-patent document 5] J. Maciejowski: Model Predictive Control, Tokyo Denki University Press, 2005 [Non-patent document 6] K. Murota: Matrices and Matroids for Systems Analysis, Springer-Verlag, Berlin,2000. Summary of the Invention [Problem to be solved by the invention]

[0006] In Patent Document 1, in order to compensate for process changes due to disturbances, the disturbances are addressed by going back over a predetermined time period in the past and adjusting the initial values ​​of the temperature distribution in the model and the model parameters. However, because process changes due to disturbances are compensated for only by adjusting the model initial values ​​and model parameters, there is a risk that control performance may deteriorate if a disturbance occurs that cannot be adjusted by changing the initial values ​​and model parameters (such as the influence of model parameters not subject to adjustment, the influence of reactions not considered in the model, or measurement errors). In addition, while the method assumes the use of a blast furnace mathematical model capable of non-steady-state calculations, numerical calculation models typically have an extremely high computational load and are difficult to calculate in a short control cycle. Therefore, this control cycle becomes a bottleneck, and there is a risk that operational fluctuations that occur when disturbances occur cannot be promptly suppressed.

[0007] In Patent Document 2, in order to correct the predicted value of the molten iron temperature calculated by the unsteady model, a regression equation is used in which the time change rate of operation indicators other than the molten iron temperature, such as the amount of solute carbon and the reducing agent ratio (RAR), is used as an explanatory variable. However, this method does not distinguish between disturbances that cause operational fluctuations in the blast furnace process, such as input disturbances superimposed on the operating end side of the blast furnace process, and output disturbances superimposed on the output end side of the blast furnace process, and there is a risk that the operational fluctuations that occur when these disturbances occur cannot be accurately suppressed.

[0008] The present invention has been made in view of the above-mentioned points, and an object of the present invention is to quickly and accurately suppress operational fluctuations that may occur when disturbances occur in process control. [Means for solving the problem]

[0009] The control device of the present invention comprises: Includes physical phenomena with different time constants A control device that performs model predictive control using a state space model that represents time changes in state variables, with a process as a control target, Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; before With the above manipulated variables and observed variables as input, Based on a design in which an input disturbance is superimposed on the manipulated variable and an output disturbance is superimposed or not superimposed on the observed variable for the state space model, estimates of the state variables; and The state space model is designed to be superimposed. Disturbance estimate and It has an estimation means that outputs death, the controlled variable includes an index for a predetermined physical phenomenon, the observable includes the controlled variable and an index for a physical phenomenon having a faster response than the predetermined physical phenomenon; the input disturbance includes an index of a factor that affects the index of the fast-response physical phenomenon included in the observable; It is characterized by: The control method of the present invention comprises: Includes physical phenomena with different time constants A control method for performing model predictive control using a state space model representing time changes in state variables for a process as a control target, comprising: Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; before With the above manipulated variables and observed variables as input, Based on a design in which an input disturbance is superimposed on the manipulated variable and an output disturbance is superimposed or not superimposed on the observed variable for the state space model, estimates of the state variables; and The state space model is designed to be superimposed. Disturbance estimate and Step to output death, the controlled variable includes an index for a predetermined physical phenomenon, the observable includes the controlled variable and an index for a physical phenomenon having a faster response than the predetermined physical phenomenon; the input disturbance includes an index of a factor that affects the index of the fast-response physical phenomenon included in the observable; It is characterized by: The program of the present invention is Includes physical phenomena with different time constants The process is the control target, and model predictive control is performed using a state space model that represents the time changes of state variables. To the computer execution Let A program for: Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; before With the above manipulated variables and observed variables as input, Based on a design in which an input disturbance is superimposed on the manipulated variable and an output disturbance is superimposed or not superimposed on the observed variable for the state space model, estimates of the state variables; and The state space model is designed to be superimposed. Disturbance estimate and As an estimation method to output The aforementioned Make your computer work 、 the controlled variable includes an index for a predetermined physical phenomenon, the observable includes the controlled variable and an index for a physical phenomenon having a faster response than the predetermined physical phenomenon; the input disturbance includes an index of a factor that affects an index of the fast-response physical phenomenon included in the observable. . [Effects of the Invention]

[0010] According to the present invention, the design allows input disturbances and output disturbances to be distinguished and superimposed, so that operational fluctuations that may occur when a disturbance occurs can be quickly and accurately suppressed in process control. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram for explaining an overview of a blast furnace process. [Figure 2] FIG. 1 is a diagram showing the relationship between a blast furnace process and manipulated variables, controlled variables, and observed variables. [Figure 3] FIG. 1 is a diagram showing the configuration of an automatic control system for a blast furnace process according to a first embodiment. [Figure 4] FIG. 10 is a diagram illustrating an example of the relationship between time and a controlled variable and an manipulated variable. [Figure 5]FIG. 1 illustrates an example of the configuration of an information processing device. [Figure 6] FIG. 1 shows the results of Example 1. [Figure 7] FIG. 6 is a diagram showing the configuration of an automatic control system for a blast furnace process according to a second embodiment. [Figure 8] FIG. 1 shows the results of Example 2. [Figure 9] FIG. 10 is a diagram for explaining an updating device in an automatic control system of a blast furnace process according to a third embodiment. [Figure 10] FIG. 10 is a characteristic diagram illustrating an example of the relationship between the amount of perturbation of an observation amount from a reference value and the amount of perturbation of a parameter from the reference value. [Figure 11] FIG. 10 is a characteristic diagram showing an example of the relationship between calculation steps and parameters of an indirect reduction reaction and the gas utilization rate, which is an observable quantity thereof. [Figure 12] FIG. 10 is a characteristic diagram showing the calculation results of the time response of the iron tapping rate and the molten iron temperature to the operation of changing the blast rate. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings. First Embodiment Figure 1 is a diagram illustrating an overview of the blast furnace process. In the blast furnace process, as shown in Figure 1, sintered ore and coke are charged into the top of a blast furnace 101 in alternating layers, and hot air and reducing agents such as pulverized coal are blown into the furnace through a blast tuyeres at the bottom. The hot air gasifies the pulverized coal and coke, and high-temperature reducing gases such as carbon monoxide and hydrogen rise up through the furnace, melting the sintered ore and removing oxygen. The molten iron comes into contact with the carbon in the coke and is reduced, resulting in molten iron containing just under 5% carbon, which accumulates in a basin at the bottom of the furnace. This molten iron is removed through a taphole located next to the bottom of the furnace and transported to the next steelmaking process. Gases such as coke oven gas (COG) and natural gas (NG) may also be blown into the furnace through a blast tuyeres (or other tuyeres).

[0013] In addition, in blast furnace operation, operation control indices are set, and the furnace conditions are controlled by monitoring the time transition of the operation control indices and executing operation operations for the blast furnace process at the operation variables corresponding to the operation control indices. Therefore, setting appropriate operation control indices is extremely important for stable operation of the blast furnace. Details of the operation control indices will be described later, but there are indices that should be maintained at set target values ​​(in this embodiment, control variables), indices that are mainly used to understand changes in the furnace conditions (in this embodiment, observation variables), and the like.

[0014] Figure 2 is a diagram showing the relationship between the blast furnace process 1 and the operation amount u, the control amount z, and the observation amount y. As shown in Figure 2, the blast furnace process 1 inputs the operation amount u and outputs the control amount z and the observation amount y. Furthermore, the state of the blast furnace process 1 is represented by x. The manipulated variable u includes one of the following (multiple selections are possible): blast volume, value related to pulverized coal volume (pulverized coal injection volume PCI or pulverized coal ratio PCR), value related to oxygen enrichment (oxygen enrichment volume or oxygen enrichment ratio), blast temperature, moisture content (blast moisture), value related to liquefied natural gas volume (LNG injection volume), coke ratio, and O / C (weight ratio of ore to coke). The controlled variable z includes the amount of pig iron produced and a furnace heat index. The furnace heat index includes one of the following (multiple selections allowed): molten iron temperature, Si in the molten iron, amount of carbon solution loss, amount of direct reduction reaction, heat balance above the tuyere, and cohesive zone level. The controlled variable z may be expressed as a linear combination of the observed variables y (described later). The observed quantity y includes the controlled quantity z (metal production rate and furnace thermal indicators), and also includes any of other furnace thermal indicators (furnace thermal indicators not treated as controlled quantity z), gas composition (e.g., CO concentration, CO2 concentration, N2 concentration, H2 concentration), gas utilization rate (e.g., CO utilization rate, H2 utilization rate), Tf (tuyere tip theoretical combustion temperature), furnace top temperature, heat flow ratio, and blast pressure (multiple selections possible).

[0015] Among the furnace heat indices described above, the heat balance above the tuyere is an index indicating the heat balance above the blast tuyere (hereinafter also referred to as the tuyere). Specifically, it is an index defined as the difference between the input heat quantity above the tuyere (i.e., generated at the tuyere level or above), which includes the sensible blast heat of the gas blown from the tuyere (also simply referred to as "sensible blast heat"), the heat of combustion of carbon burning in the raceway (also simply referred to as "carbon combustion heat"), and the reaction heat due to the indirect reduction reaction by carbon monoxide (also simply referred to as "indirect reduction heat"), and the output heat quantity above the tuyere, which includes the decomposition heat of blast moisture of the gas blown from the tuyere (also simply referred to as "blast moisture decomposition heat"), the decomposition heat of pulverized coal (also simply referred to as "pulverized coal decomposition heat"), the reaction heat due to the solution loss reaction, which is an endothermic reaction (also simply referred to as "solution loss reaction heat") and the reaction heat due to the direct reduction reaction (also simply referred to as "direct reduction heat"), and the reaction heat due to the hydrogen reduction reaction (i.e., the reduction reaction of iron ore by hydrogen) (also simply referred to as "hydrogen reduction heat").

[0016] [State-space model] In this embodiment, a discrete-time linear state space model of equation (1) that represents blast furnace operation is used to control the blast furnace process 1. The discrete-time linear state space model of equation (1) includes an equation that represents a discrete-time change in which the state (state variable) of the blast furnace process 1 changes from x(k) to x(k+1) when the current time is represented by k and an operating variable u(k) is applied at the current time. The discrete-time linear state space model of equation (1) also includes a relational expression between the estimated state variable x, the controlled variable z, and the observed variable y. For simplicity of explanation, the direct feedthrough term is omitted in equation (1), but may be considered as necessary.

[0017]

number

[0018] Here, we will explain how to create the discrete-time linear state space model of equation (1). It is assumed that the transfer function matrix G(s) of equation (2) that holds between each manipulated variable u and each observed variable y has been obtained. s is the Laplace operator, and G(s) is a transfer function matrix with p rows and m columns. p is the dimension of the observed variable y, and m is the dimension of the manipulated variable u.

[0019]

number

[0020] The transfer function matrix G(s) can be created, for example, by an open-loop or closed-loop identification method described in Non-Patent Document 1, etc., using data including periods during which the operation level (operating point) is changed during blast furnace operation. As a method that does not use data during operation, if a blast furnace mathematical model (blast furnace unsteady model) capable of unsteady calculations is available, the transfer function matrix can be determined by open-loop identification using time-series data obtained by changing several operating points of the blast furnace unsteady model. In either case, a linear model can be assumed because additivity of the operating variables generally holds for perturbations of the operating point during steady operation. Therefore, for example, the transfer function matrix can be determined for each operating variable.

[0021] Furthermore, the transfer function matrix G(s) obtained in this way is converted into an appropriate continuous-time state space representation such as a controllable canonical form by so-called state realization, and then time discretized using a method such as zero-order hold (keeping the input value constant between control periods), thereby creating a discrete-time linear state space model of equation (1). That is, the matrices A, B, C, C zcan be obtained. Furthermore, in order to apply a type of state estimation observer called a disturbance observer and model predictive control (described later), it is assumed that this discrete-time linear state-space model satisfies reachability and observability. In discrete-time linear state-space models, as described in Non-Patent Document 2 and elsewhere, reachability is a more accurate concept in discrete-time systems as a concept corresponding to the controllability of continuous-time state-space models. However, for convenience, this property is also referred to as controllability. Even if the discrete-time linear state-space model obtained does not have controllability and observability, a controllable and observable model can be constructed by applying balanced realization, for example, as described in Non-Patent Document 3. The calculation procedure for balanced realization is described in Chapter 5 of Non-Patent Document 4. Using the controllability Gramian and observability Gramian defined for a given state-space model, the Hankel singular values, which indicate the degree of controllability and observability, are calculated. A state-space model that satisfies controllability and observability can be obtained by truncating Hankel singular values ​​below a predetermined threshold. When the number of observables is large, simply creating a state-space model by time-discretizing the transfer function matrix G(s) may not result in a controllable and observable state-space model. The state-space model reconstructed by the balanced realization described above has a particularly remarkable effect when the dimensions of such observables are large.

[0022] The method for creating a discrete-time linear state space model is not limited to the method described above that involves creating a transfer function matrix. A discrete-time linear state space model may be obtained directly from input and output time-series data, without creating a transfer function matrix, by using a system identification method such as the subspace identification method (MOESP method or N4SID method) described in Non-Patent Document 1. Although a discrete-time linear state space model is used, if a continuous-time state space model is available, it may be time-discretized and used. Even if a nonlinear state space model is available, a linear state space model may be created by linear approximation around the operating point, or a linearized model may be used using a method such as Koopman mode decomposition or Carleman linearization.

[0023] [Automatic control system for blast furnace process including disturbance observer] FIG. 3 is a diagram showing the configuration of an automatic control system for a blast furnace process 1 according to the first embodiment. The automatic control system for the blast furnace process 1 includes a controller 2 and a discrete-time linear state space model (A, B, C, C) obtained by the above-described method. z matrix), and a controller 2 executes model predictive control for a blast furnace process 1 as a control target. The blast furnace process 1 is designed so that an input disturbance d1 can be superimposed on the manipulated variable u, and an output disturbance d2 can be superimposed on the observed variable y. The controller 2 receives the target value r and the estimated values ​​x^, d1^, and d2^ (for example, x^ is represented by a ^ above the x) that are the output of the disturbance observer 3 as inputs, and outputs the manipulated variable u by performing model predictive control. The disturbance observer 3 functions as the estimation means of the present invention, and acquires an observation value y including a controlled variable z output from the blast furnace process 1 when an operation value u is input to the blast furnace process 1, and outputs an estimated value x^ of a state variable and at least one of an estimated value d1^ of the input disturbance and an estimated value d2^ of the output disturbance, based on the operation value u and the observation value y superimposed with at least one of an input disturbance d1 and an output disturbance d2. In the example shown in Fig. 3, the disturbance observer 3 receives as input the operation value u and the observation value y superimposed with the output disturbance d2, and outputs an estimated value x^ of the state variable, an estimated value d1^ of the input disturbance, and an estimated value d2^ of the output disturbance.

[0024] Figure 3 shows a schematic representation of the locations of disturbances superimposed on the blast furnace process 1. The input disturbance d1 is a disturbance or measurement error superimposed on the inlet side (operating end) of the blast furnace process 1, and the output disturbance d2 is a disturbance or measurement error superimposed on the outlet side (output end) of the blast furnace process 1. In an actual blast furnace process, the disturbances do not necessarily occur exactly at the positions shown in Figure 3, but in practice, they can be considered to be close to either the input disturbance d1 or the output disturbance d2. For example, a decrease in reducibility due to the charged raw materials or a change in the heat of pulverized coal decomposition due to property variations can be considered as the input disturbance d1. Furthermore, variations in the measured temperature of the molten iron can be considered as the output disturbance d2. Furthermore, model errors that arise when approximating blast furnace phenomena with a discrete-time linear state space model can also be considered as the input disturbance d1 or the output disturbance d2.

[0025] Next, a method for constructing a disturbance observer 3 that takes into account the input disturbance d1 and output disturbance d2 superimposed on the blast furnace process 1 as state variables in a Luenberger-type state estimation observer will be described. The method for constructing the disturbance observer 3 is described, for example, in Non-Patent Document 5. Here, a more general case will be described in which the disturbance observer 3 is configured as in Equation (3), which includes an input disturbance and an output disturbance. In the following, an example will be described in which the disturbance to be estimated is a constant-value disturbance, but as will be described later, it is not limited to a constant-value disturbance.

[0026]

number

[0027] In addition, the observer gain (L x ,L d ) can be calculated using the discrete Riccati equation by the LQR design method, for example.

[0028] non-negative integer n d1 ,n d2 are the dimensions of the estimated input disturbance d1^ and the estimated output disturbance d2^, respectively, and their sum is n d Let (n d1 +n d2 =nd ) The dimension of the control variable z is r≦n d ≦p of the dimension of the observable y, but n d Explain as =p.

[0029] where B d1 (See equation (4)) is a matrix obtained by extracting from matrix B the column corresponding to the manipulated variable u, which is assumed to be superimposed by the input disturbance d1. d2 (See equation (5)) is a matrix determined according to the observation quantity y that is considered to be superimposed with the output disturbance d2. For example, if all disturbances are configured as input disturbances d1 and output disturbance d2 is not included, B d1 =B,C d2 =O. In addition, if all disturbances are configured as output disturbances d2 and input disturbances d1 are not included, B d1 =O,C d2 =I. In addition, B d1 and C d2 is a matrix that determines the influence of a disturbance on a state variable, and is not limited to the form associated with an input disturbance or an output disturbance as described above, but in this embodiment, it will be described in this form.

[0030]

number

[0031] matrix B d1 ,C d2 Specifically, in the construction of d1 +n d2 =n d (=p)(0≦n d1 ≦m, 0≦n d2 Any combination of input disturbance d1 and output disturbance d2 that satisfies (m+pn d )(m+p is the top row, n d (The numbers are written in the bottom row.) For example, m=2, p=4, n d = 4, there are 15 possible construction methods.

[0032] Thus, matrix B d1 ,C d2 is determined according to the assumed position of the disturbance superimposition. In other words, the selection of the position of the disturbance superimposition is determined by the matrix B d1 ,C d2 This corresponds to designing the following. matrix B d1 ,C d2 When prior knowledge is available regarding the position and degree of influence of disturbances superimposed on the blast furnace process 1, the disturbance observer 3 selects the superimposed position of the disturbance to be considered based on that knowledge, and the matrix B d1 ,C d2 If you do not have any prior knowledge, you can design (m+pn d ) combinations, the combination with the lowest worst performance of the evaluation function representing the control performance, that is, the combination with the highest robustness may be selected.

[0033] In the following, the estimated value d1^ of the input disturbance and the estimated value d2^ of the output disturbance are combined into one vector as in equation (6), and equation (3) is expressed as equation (7).

[0034]

number

[0035] In this embodiment, the disturbance observer 3 is constructed including the case where an observation quantity y other than the control quantity z can be used. For example, as described in Non-Patent Document 5, in the configuration of the disturbance observer, the control quantity z and the observation quantity y are the same, that is, C z =C, and then an observer is constructed that takes into account only the same number of output disturbances as the controlled variable z. However, in the present invention, by taking into account the optimal combination of input disturbances d1 and output disturbances d2, it is possible to obtain an effective suppression effect of operational fluctuations when a disturbance occurs.

[0036] Furthermore, the formulation of the disturbance configured by the disturbance observer 3 is not limited to a constant value disturbance, but may be a ramp disturbance. Also, it may have periodic dynamics such as a first-order lag system or a sine wave. The dynamics of such disturbances can generally be taken into account in the observer as in equation (9). For example, in equation (9), especially, A ξ =O,B ξ If we set I, it corresponds to the case of constant disturbance, and A ξ =I,B ξ = I corresponds to the case of a ramp disturbance.

[0037]

number

[0038] [Model Predictive Control] Control combining a disturbance observer 3 and model predictive control will be described. In model predictive control, feedback control is performed while calculating an optimization problem online. Details of model predictive control are provided in Non-Patent Document 5. In model predictive control, a state variable at the current time k is estimated by a state estimation observer (disturbance observer 3 in this embodiment), and this estimated value is used as the initial value of a predicted trajectory Receding Horizon. A predicted trajectory for future control operations is calculated based on the estimated value of the state estimation observer, which is the initial value. As shown in equation (10), the controller 2 calculates the error between the predicted trajectory for future control operations and the reference trajectory (target value r), and the series of changes in the manipulated variable u at future times, Δu^(k+1|k), Δu^(k+2|k), Δu^(k+H u Model predictive control is performed by calculating a series of changes in the optimal manipulated variables that minimizes a quadratic evaluation function V(k) for |k.

[0039]

number

[0040] x^(k+i|k) represents the estimated value x^ of the state variable at future time k+i for the series of changes in the optimal manipulated variable calculated at current time k, and z^(k+i|k) has the same meaning. Also, r(k+i|k) means the value of the target value r at future time k+i used in the calculation at current time k. Also, Δu^(k+i|k) means the value of the change in the optimal manipulated variable Δu^ at future time k+i used in the calculation at current time k. Figure 4 shows an example of the relationship between time and the controlled variable and manipulated variable. Figure 4(a) shows the relationship between time and r(k+i|k) and z^(k+i|k). Figure 4(b) shows the relationship between time and u(k)+Δu^(k+i|k).

[0041] In the following, for the sake of simplicity, equation (10) is rewritten as equation (11) using vectors and matrices.

[0042]

number

[0043] Here, as in equation (12), Z(k) is summarized in the form of a matrix and a vector.

[0044]

number

[0045] Using the input disturbance and output disturbance estimated by the disturbance observer 3, the tracking error ε(k), which is the error between the predicted trajectory and the reference trajectory, is defined as in equation (13) taking into account the influence of the input disturbance and the output disturbance. Defining the tracking error in this way makes it possible to perform target value control without generating steady-state error.

[0046]

number

[0047] Furthermore, by using properly defined matrices and vectors, the evaluation function V(k) in equation (10) can be expressed in a quadratic form as in equation (14).

[0048]

number

[0049] In model predictive control, a constraint may be set for at least one of the manipulated variable u and the controlled variable z. The constraints include, for example, upper and lower limit constraints on the manipulated variable u, upper and lower limit constraints on the controlled variable z, upper and lower limit constraints on the amount of change in the manipulated variable u, and upper and lower limit constraints on the amount of change in the controlled variable z. More specifically, the constraints on the controlled variable z include an upper limit constraint on the amount of iron production and a lower limit constraint on the furnace heat index. Furthermore, the constraints on the manipulated variable u include upper and lower limit constraints on the amount of change in the amount of blast rate and upper and lower limit constraints on the amount of change in the blast rate. The constraints are not limited to these, and any constraint that can be expressed by a linear inequality or equality may be used.

[0050] From the above, model predictive control is reduced to the problem of solving a quadratic programming problem consisting of a quadratic evaluation function and linear inequality constraints. Various solvers have been developed for quadratic programming problems, allowing for high-speed solutions, making it possible to achieve real-time calculation processing with short control cycles.

[0051] It is also possible to allow violations of some of the constraints. Constraints are divided into those that must be strictly met (hard constraints) and those that do not (soft constraints), and violations of soft constraints are allowed. For example, constraints on operation variables are hard constraints, while constraints on control variables are soft constraints. Priorities can also be set among soft constraints. For example, when it comes to constraints on the upper limit of the iron tapping rate and the lower limit of the iron temperature, it is possible to formulate a system in which the lower limit of the iron temperature takes precedence.

[0052] Furthermore, when the observable quantity y includes an index other than the controlled variable z, the evaluation function V(k) may further include a term of the sum of squares relating to the time differential or time difference of the future predicted value of the estimated value of the observable quantity, as shown in equation (15). In this way, the time change of the observable quantity can be taken into account in the evaluation function V(k), which is expected to improve the speed of response and make it possible to more quickly suppress operational fluctuations when disturbances occur.

[0053]

number

[0054] Fig. 5 is a diagram showing a configuration example of an information processing device 200 that functions as a control device for a blast furnace process 1. The information processing device 200 configures an automatic control system for the blast furnace process 1 that includes a controller 2 and a disturbance observer 3 as shown in Fig. 3, and performs model predictive control. The information processing device 200 includes a CPU 211 , a ROM 212 , a RAM 213 , an auxiliary storage device 214 , a display unit 215 , an operation unit 216 , a communication I / F (interface) 217 ​​, and a bus 218 . The CPU 211 controls the entire information processing device 200 using computer programs and data stored in the ROM 212 and RAM 213, thereby realizing the functions of the controller 2 and the disturbance observer 3 and executing model predictive control. The information processing device 200 may include one or more dedicated hardware components different from the CPU 211, and at least a portion of the processing by the CPU 211 may be executed by the dedicated hardware components. Examples of the dedicated hardware components include an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), and a DSP (Digital Signal Processor). The ROM 212 stores programs and the like that do not require modification. The RAM 213 temporarily stores programs and data supplied from the auxiliary storage device 214 and data and the like supplied from the outside via the communication I / F 217. The auxiliary storage device 214 is configured, for example, by a hard disk drive or the like, and stores various data such as image data and audio data. The display unit 215 is configured with, for example, a liquid crystal display or LEDs, and displays a GUI (Graphical User Interface) or the like for the user to operate the information processing device 200. The operation unit 216 is configured with, for example, a keyboard, a mouse, a joystick, a touch panel, and the like, and receives operations by the user to input various instructions to the CPU 211. The communication I / F 217 is used for communication with devices external to the information processing device 200. For example, when the information processing device 200 is connected to an external device via a wired connection, a communication cable is connected to the communication I / F 217. When the information processing device 200 has a function for wireless communication with an external device, the communication I / F 217 is equipped with an antenna. The bus 218 connects each unit of the information processing device 200 to transmit information. Although the display unit 215 and the operation unit 216 are illustrated as existing inside the information processing device 200, at least one of the display unit 215 and the operation unit 216 may be located outside the information processing device 200 as a separate device.

[0055] As described above, in the control of the blast furnace process, the input disturbance d1 and the output disturbance d2 are designed to be distinguishable and superimposable, so that operational fluctuations that may occur when a disturbance occurs can be quickly and accurately suppressed. [Example]

[0056] In blast furnace operation, measurements or analysis values ​​of molten iron temperature and molten iron components are obtained via the molten iron reservoir at the hearth, which is believed to have a certain time lag relative to the reaction timing in the blast furnace core. On the other hand, the gas composition at the furnace top, furnace top temperature, blast pressure, etc. are believed to change relatively quickly compared to analytical values ​​of molten iron components. Therefore, by using analytical values ​​such as gas composition measured at the furnace top other than molten iron temperature as the observable variable y, the effects of disturbances (especially input disturbances) occurring during operation can be quickly detected and operational fluctuations can be quickly suppressed when disturbances occur. Indicators that change relatively quickly and are suitable for use as the observable variable y are primarily those measurable at the furnace top, such as gas composition. In addition to gas composition, other indicators include gas utilization rate, carbon solution loss, direct reduction reaction rate, heat balance above the tuyere, furnace top temperature, heat flow ratio, and blast pressure.

[0057] In the examples, the controlled variables z were the pig iron tapping rate and molten iron temperature, and the controlled variables u were the blast rate, oxygen enrichment rate, and pulverized coal injection rate. Note that the oxygen enrichment rate was controlled in conjunction with the blast rate, with the oxygen enrichment rate held constant, so the actual degree of freedom of the controlled variables was two. Therefore, in the following, only the blast rate and pulverized coal injection rate will be described as controlled variables, and the oxygen enrichment rate will be omitted. In addition, the following description will be given assuming that the observable quantity y other than the controlled variable z is the gas composition (CO concentration, CO2 concentration). In this case, the dimension p of the observable quantity y is 4, the dimension r of the controlled variable z is 2, and the dimension m of the manipulated variable u is 2. [Controlled quantities] Iron production volume, furnace heat index (molten iron temperature) [Operation variables] Blow rate, pulverized coal injection rate, oxygen enrichment rate *Oxygen enrichment rate is linked to blow rate with the oxygen enrichment rate kept constant [Observed quantities] Amount of hot metal produced, molten iron temperature, gas composition (CO concentration, CO2 concentration)

[0058] (Construction of controllable and observable state-space models) The transfer function matrix between two control variables u (blast rate, pulverized coal injection rate) and four observation variables y (hot metal production rate, hot metal temperature, gas composition (CO concentration, CO2 concentration)) was calculated, and time discretized using zero-order hold with Δt = 20 min. Some of the Hankel singular values ​​of the state space model obtained were close to zero, so they were truncated at eighth order to construct a controllable and observable state space model.

[0059] (Building a disturbance observer) Next, the disturbance observer 3 is constructed. As described above, there is a degree of freedom in the superimposition position of the disturbance that can be considered by the disturbance observer 3. In this example, since the dimension p of the observable variable y is 4, a total of four input disturbances d1 and output disturbances d2 can be selected. Here, a case where a change in the heat of pulverized coal decomposition is assumed as the disturbance will be described. This disturbance affects the effect of pulverized coal contained in the manipulated variable u, and therefore corresponds to the input disturbance d1. Therefore, here, two input disturbances d1 are selected, and the remaining two disturbances are set as output disturbances d2, and two are selected from the four observable variables y. Based on preliminary studies, the disturbance observer 3 was constructed with the input disturbances d1 being the blast rate and pulverized coal injection rate, and the output disturbances d2 being the tapping rate and molten iron temperature, as highly robust options.

[0060] In the control combining the disturbance observer 3 constructed in this way and model predictive control, the suppression effect of operational fluctuations when a disturbance occurs was investigated by simulation. As a comparative example, the suppression effect of operational fluctuations when a disturbance occurs was investigated by simulation in a control system in which the disturbance observer 3 was constructed using only the tapping rate and molten iron temperature as the observed quantities y. Figure 6 shows the simulation results. Figure 6(a) shows the time-series change in the tapping rate, and (b) shows the time-series change in the molten iron temperature. The dotted line shows the characteristics of the example, and the solid line shows the characteristics of the comparative example. In the comparative example, the controlled variable and the observed variable were the same. That is, only the tapping rate and the molten iron temperature were used as the observed variables y, and a disturbance observer was constructed that considered only output disturbances. Although disturbances can be suppressed and the steady-state deviation can be reduced to zero in this case, when the disturbance observer 3 is constructed using more observed variables y, as in the example, it is possible to suppress operational fluctuations when a significant disturbance occurs. The reason for this is that by using an index (gas composition in the example) that changes reactionally relatively more quickly than the molten iron temperature and by also considering input disturbances, it is possible to detect and suppress disturbances more quickly.

[0061] In this way, in a process in which physical phenomena with different time constants coexist, such as a blast furnace process, the difference in the time constants of the time changes of multiple observables can be utilized. That is, by using another index as an observable prior to the time change of the index to be controlled, it becomes possible to quickly detect and suppress disturbances. In this embodiment, a blast furnace process has been described as an example of a process involving physical phenomena with different time constants, but the present invention is not limited to this. For example, the present invention can be applied to cases in which flow rates and pressures with faster response can be measured in temperature control of a thermal process that generally has a slow response. Furthermore, the present invention can also be applied more generally to a process with a cascade control structure having a master loop (major loop) and a slave loop (minor loop).

[0062] <Second embodiment> In the second embodiment, an example will be described in which, when a disturbance is measurable, a feedforward correction function for control operation is provided based on the measurable disturbance. Note that, in the following, detailed descriptions of the points in common with the first embodiment will be omitted, and the description will focus on the points that are different from the first embodiment. FIG. 7 is a diagram showing the configuration of an automatic control system for a blast furnace process 1 according to a second embodiment. The same components as those shown in FIG. 3 are given the same reference numerals, and detailed descriptions thereof will be omitted. In this embodiment, the controller 2 is provided with a measurable disturbance dm The controller 2 inputs the measurable disturbance d m It acts as a correction means to perform a feedforward correction of the control action based on the measurable disturbance d m These include the particle size and composition of the raw materials (ore, coke).

[0063] Hereinafter, the measurable disturbance d m The calculation of the optimal manipulated variable based on the above will be explained. In Figure 7, the measurable disturbance d m is shown superimposed on the blast furnace process 1, but the measurable disturbance d m are divided into those corresponding to input disturbances superimposed on the inlet side of the blast furnace process 1 and those corresponding to output disturbances superimposed on the outlet side of the blast furnace process 1. Here, the measurable disturbance d m The formulation will be explained assuming that is the input disturbance. In this case, the discrete-time linear state-space model is a model with a measurable disturbance d m Taking this into consideration, the equation becomes as shown in equation (16). Note that the control amount z is the same as in equation (1), so it is omitted here. In addition, among the various operational operations that are input conditions of the blast furnace, the operations that were not selected as the manipulated variable u for the discrete-time linear state space model are measurable disturbances d m It may be considered as such.

[0064]

number

[0065] In this case, the future predicted values ​​of the state variable x and the controlled variable z are expressed as shown in equation (17).

[0066]

number

[0067] where the predicted vector of measurable disturbances D m (k) is defined as in equation (18).

[0068]

number

[0069] The tracking error ε(k) can be calculated by subtracting the measurable disturbance obtained by equation (18).

[0070]

number

[0071] Thus, the measurable disturbance d m By correcting the tracking error ε(k) based on the above, the optimal manipulated variable is consequently given the effect of feedforward correction against disturbances. In addition, when the input disturbance and output disturbance estimated by the disturbance observer 3 are also taken into consideration, the tracking error calculated by equation (13) is further calculated as ΞD m Just subtract (k).

[0072] In addition, the measurable disturbance d m When is the output disturbance, the discrete-time linear state space model is as shown in equation (20). Note that the controlled variable z is the same as equation (1), so it is omitted here.

[0073]

number

[0074] In this case, the future predicted values ​​of the state variable x and the controlled variable z are expressed as shown in equation (21).

[0075]

number

[0076] The tracking error ε(k) can be calculated by subtracting the measurable disturbance as shown in equation (22).

[0077]

number

[0078] In this case, when the input disturbance and output disturbance estimated by the disturbance observer 3 are also taken into consideration, the tracking error calculated by equation (13) is further calculated as follows: m ×I)d m (k) (here The "x" in the description represents the Kronecker operator.

[0079] The formulation was explained above when considering both the measurable disturbance corresponding to the input disturbance and the measurable disturbance corresponding to the output disturbance. m,1 and the measurable disturbance d corresponding to the output disturbance m,2 can also be considered simultaneously.

[0080] Note that, whether it is an inlet position or an outlet position, as long as the main disturbance superimposed on that position can be measured, it is possible to configure the disturbance observer 3 described in the first embodiment in which the disturbance is set at a position other than that position. For example, if there is a measurable disturbance equivalent to the input disturbance, it is possible to design the disturbance observer 3 by reducing the input disturbance d1 and increasing the output disturbance d2. In this way, in the second embodiment, since the disturbance at a certain position can be measured, the disturbance observer 3 can be designed more effectively by using a sensor that measures the disturbance.

[0081] The formulation of disturbances is not limited to the above formulation, and dynamics may also be taken into consideration. For example, as in the first embodiment, process changes in response to measurable disturbances may be formulated using a state space model. Actual operation data including the period before and after the change point in the particle size of the raw ore measured during operation, or time response data obtained using a blast furnace mathematical model capable of non-steady-state calculations, may be used to create a transfer function matrix from this time series data, and a state space model may be created by state realization and time discretization. For example, if the interaction between the manipulated variable and the disturbance is ignored and additivity holds between the response of the control operation and the response of the measurable disturbance, then a formulation such as equation (26) is possible. Here, the disturbance is assumed to be superimposed on the output position, but this is not limiting.

[0082]

number

[0083] In this case, the disturbance observer 3 can be constructed as shown in equation (27).

[0084]

number

[0085] If there is a risk that observability will not be established in the above formulation, the observer gain L m = 0 and use the simpler form of equation (28).

[0086]

number

[0087] When the interaction between the manipulated variable and the disturbance cannot be ignored and additivity does not hold, for example, a model such as Equation (29) that takes the interaction effect into consideration, or a linearized model using a method such as Koopman mode decomposition or Carleman linearization, may be created.

[0088]

number

[0089] FIG. 8 shows the calculation results when particle size (average particle size of ore) is measured as a measurable disturbance. FIG. 8(a) shows the time series change in particle size, FIG. 8(b) shows the time series change in pulverized coal injection rate, and FIG. 8(c) shows the time series change in molten iron temperature. As shown in FIG. 8(a), the change in particle size, which is a measurable disturbance, is detected from sensor information. As shown in FIG. 8(b), a feedforward correction amount corresponding to the change in particle size is automatically calculated for the pulverized coal injection rate, which is the manipulated variable u (FB+FF control). As a comparative example, an example in which the controlled variable and the observed variable are set to be the same and only the feedback control described in the first embodiment is performed (FB control) is also shown. As a result, the molten iron temperature changes as shown in FIG. 8(c). As a comparative example, the cases of FB control and no control (maintaining the manipulated variable constant) are also shown. By providing a feedforward correction function for the control operation, the effect of suppressing operational fluctuations when a significant disturbance occurs is achieved.

[0090] <Third embodiment> In the third embodiment, an example will be described in which a function for updating a state space model (discrete-time linear state space model of Equation (1)) used for controlling the blast furnace process 1 is provided. In the first and second embodiments, short-term disturbance fluctuations can be compensated for, but the third embodiment aims to be able to compensate for relatively long-term operational fluctuations in addition to that. Note that, in the following, detailed description of commonalities with the first embodiment will be omitted, and differences from the first embodiment will be mainly described.

[0091] An update device 4 in an automatic control system of a blast furnace process according to a third embodiment will be described with reference to Fig. 9. Fig. 9(a) is a diagram showing an example of the functional configuration of the update device 4 in the automatic control system of a blast furnace process 1 according to the third embodiment. Fig. 9(b) is a flowchart showing an example of processing executed by the update device 4. The update device 4 updates the discrete-time linear state space model of equation (1) (hereinafter simply referred to as the state space model). For example, the information processing device 200 (see FIG. 5) that functions as a control device for the blast furnace process 1 may be configured to also function as the update device 4. The update device 4 functions as the update means referred to in the present invention. The update device 4 includes a determination unit 41 , a parameter fitting calculation unit 42 , and a linear approximation calculation unit 43 .

[0092] In step S101, the determination unit 41 determines whether or not a predetermined condition for updating the state space model is met. In this embodiment, a change in a predetermined operating condition of the blast furnace process 1 corresponds to the predetermined condition (hereinafter referred to as condition 1). The determination unit 41 detects a change in the predetermined operating condition based on the measured values ​​of the manipulated variable u and the observed variable y measured by various sensors. A change in the predetermined operating condition means that the instantaneous values ​​of the manipulated variable u and the observed variable y, or the time average value or standard deviation over a predetermined period, have changed by more than a predetermined amount. The predetermined period is set in advance based on actual data, etc. Factors that cause such a change include a change in the operating point (operating point) of the blast furnace process, a change in the reducibility of the ore (such as a change in the reducibility index RI), and a change in the process properties corresponding to a change in the internal state of the blast furnace, such as a change in permeability. Furthermore, the fact that a predetermined time has come meets a predetermined condition (hereinafter referred to as condition 2). The determination unit 41 detects that the predetermined time has come by, for example, communicating with a higher-level process computer. Furthermore, the occurrence of a predetermined event in the blast furnace process 1, such as an event such as a blast shutdown, satisfies a predetermined condition (hereinafter referred to as condition 3). The determination unit 41 detects the occurrence of the predetermined event by, for example, communicating with a higher-level process computer.

[0093] When the determination unit 41 determines that any one of the conditions 1 to 3 is satisfied, in step S102, the parameter fitting calculation unit 42 performs a parameter fitting calculation for the mathematical model used to create the state space model. In this embodiment, when the determination unit 41 determines that any one of conditions 1 to 3 is satisfied, the parameter fitting calculation unit 42 performs a parameter fitting calculation for the current blast furnace mathematical model. Conditions 1 to 3 are assumed to indicate changes in the parameters of the blast furnace mathematical model, so the parameter fitting calculation is performed. The parameters of the blast furnace mathematical model will be described later. Condition 1 also includes changes in the operating point (operating point) of the blast furnace process due to changes in target values ​​of operational parameters, etc. In this case, changes in the parameters of the blast furnace mathematical model may not necessarily occur. In such cases, the convergence determination condition in the parameter fitting calculation process described later is satisfied, and as a result, the parameter fitting calculation may be omitted.

[0094] In step S103, the linear approximation calculation unit 43 creates a state space model from the blast furnace mathematical model, and performs linear approximation calculation around the operating point. As described above, when the judgment unit 41 judges that any one of the conditions 1 to 3 is satisfied, the parameter fitting calculation unit 42 performs the parameter fitting calculation, and a state space model is created from the blast furnace mathematical model after the parameter fitting calculation. At that time, a linear approximation calculation is performed around the operating point. If the determining unit 41 determines that none of the conditions 1 to 3 is met, neither the parameter fitting calculation nor the linear approximation calculation is performed.

[0095] The updating device 4 configured as described above performs updating by using the state space model created by the linear approximation calculation unit 43 as a new state space model (replaces the current state space model with the new state space model). Although the update device 4 executes steps S101 to S103, some or all of the processes may be performed by a person.

[0096] The processing by the parameter fitting calculation unit 42 will be described below. First, the mathematical model of the blast furnace will be described. A blast furnace mathematical model defines a small region within the blast furnace and simulates the time evolution of behavior within the small region based on equations for physicochemical phenomena such as mass transfer, reactions, and heat transfer in the lumpy zone and cohesive zone. This model can be used to understand or predict furnace conditions based on blast furnace operating conditions and raw material properties. For example, a three-dimensional blast furnace mathematical model defines multiple meshes (small regions) by dividing the internal region of the blast furnace in the vertical, radial, and circumferential directions, and simulates behavior within each mesh. Various papers have been published on blast furnace mathematical models, but suitable examples include those described in Koji Takatani et al., "Three-dimensional Dynamic Simulator for Blast Furnace," ISIJ International, Vol. 39 (1999), No. 1, pp. 15-22, or Nishioka et al., "Development of a Blast Furnace Mathematical Model" (Nippon Steel & Sumitomo Metal Technical Report, No. 410, pp. 73-79, 2018). For example, by using a blast furnace mathematical model capable of performing non-steady-state calculations such as those described above, it is possible to calculate the time changes in the iron production rate (tons / day), coke ratio CR (kg / t), pulverized coal ratio PCR (kg / t), blast pressure (hPa), furnace top pressure (hPa), furnace top temperature (°C), gas utilization rate ηCO (%), and molten iron temperature (°C), for various boundary conditions such as the various model parameters used in the blast furnace mathematical model, the blast furnace raw material charging conditions, and the blast conditions.

[0097] These blast furnace mathematical models take into account the various chemical reactions that occur in the blast furnace, but to use the blast furnace mathematical model effectively, it is desirable to appropriately adjust the reaction rate constants and other model parameters of the various chemical reactions that occur in the blast furnace so that the measured values ​​of various physical quantities match the calculated values ​​of the model. Examples of measured values ​​of such physical quantities include blast pressure, H2 utilization rate, FeO concentration in the slag, CO2 utilization rate, carbon solution content (CSL), Si concentration in the molten iron, C concentration in the molten iron, molten iron temperature, iron production rate, and coke ratio. Furthermore, examples of parameters within a model that are candidates for adjustment include correction coefficients for reaction rate constants of chemical reactions such as ore reduction, coke gasification, SiO2 reduction, carburization, and hydrogen gas shift reactions, as well as correction coefficients for solid-liquid heat exchange coefficients and furnace void fraction correction parameters, as described in "Development of a Blast Furnace Mathematical Model by Nishioka et al." (Nippon Steel & Sumitomo Metal Technical Report, No. 410, pp. 73-79, 2018).

[0098] Next, the prerequisites for parameter fitting calculations will be described. A prerequisite for parameter fitting calculations may be, for example, stable operation of a blast furnace. During stable operation of a blast furnace, fluctuations in each operational parameter are relatively small, and the statistical properties of each operational parameter, such as the average value and variance, over a given period can be considered to be approximately constant. A specific method for determining such stable conditions may involve determining whether the instantaneous values ​​of parameters of interest, such as the iron tapping rate and molten iron temperature, fall within predetermined upper and lower limits, or whether the time-series average value, standard deviation, or coefficient of variation of these parameters falls within predetermined thresholds.

[0099] Next, we will discuss the correction coefficients for the reaction rate constants of chemical reactions and the parameters within the model, which are the targets of parameter fitting calculations. In this embodiment, eight types of parameters are subject to parameter fitting calculation: a correction coefficient for the reaction rate constant of the indirect reduction reaction (hereinafter simply referred to as the indirect reduction reaction; the same applies to other reactions), a coke gasification reaction, a SiO reduction reaction, a direct reduction reaction, a carburization reaction, a hydrogen gas shift reaction, a solid-liquid heat exchange coefficient, and a furnace void fraction correction parameter. Depending on the purpose, such as shortening the calculation time, it is possible to select only some of the eight parameters as the calculation targets, rather than all of the eight parameters. Examples of observable quantities to be fitted with the above eight parameters include, but are not limited to, the CO utilization rate (ηCO) or CSL, hot metal temperature, Si concentration in hot metal, slag FeO, C concentration in hot metal, H utilization rate (ηH2), and blast pressure. The observables selected for the parameter fitting calculation may be the same as the observables already listed above, or may be different. In this embodiment, different observables are selected.

[0100] Next, the algorithm for parameter fitting calculation will be described. In the parameter fitting calculation, a situation in which the blast furnace process is in a nearly steady state, as described above, is assumed. In such a case, a non-steady-state calculation of the blast furnace mathematical model is performed using actual data time-averaged from the current time back a predetermined time as input, and parameters are fitted so that the model calculated values ​​(referred to as steady-state calculated values) calculated up to the steady state match the actual values. Regarding the average time length, for example, in order to consider the consistency of the mass balance in the actual values ​​of the operational specifications, the time it takes for raw materials charged into the blast furnace to pass through the furnace can be set as one unit, and the time length can be set as an integer multiple of that. For example, the time length can be set as an integer multiple of about 8 hours as one unit. Specifically, in this parameter fitting calculation step, the simultaneous nonlinear equations of equation (30) are solved.

[0101]

number

[0102] Here, if the gradient ∂F / ∂p of the nonlinear function F with respect to the parameter p can be calculated, p* can be quickly found with a small number of iterations using an iterative algorithm such as Newton's method. However, in reality, the nonlinear function F corresponds to the calculation procedure of the blast furnace mathematical model, and it is difficult to accurately calculate the gradient ∂F / ∂p. Therefore, following the procedure described below, the nonlinear simultaneous equations in equation (30) are solved without using an accurately calculated gradient ∂F / ∂p.

[0103] In the parameter fitting calculation, the sensitivity of the eight parameters to be calculated is first calculated. The sensitivity is expressed by a relational expression of the scalar perturbation of each parameter and each observable around the reference value. In particular, the relationship between the perturbation Δy of the observable from the reference value and the perturbation Δp of the parameter from the reference value can be well approximated by a quadratic function or an exponential function. Equation (30) represents the relational expression when approximated by a quadratic function. Figure 10 is a characteristic diagram showing an example of the relationship between the perturbation Δy of the observable from the reference value and the perturbation Δp of the parameter from the reference value approximated by a quadratic function. This is calculated for all parameters and observables. In other words, a and b in the relational expression (31) are found for all parameters and observables.

[0104]

number

[0105] The specific calculation procedure for parameter fitting will now be described. In this embodiment, the parameter update amount is calculated so as to eliminate the target error δy of the observation amount. In this paragraph, too, the following description will be given assuming that δy and p defined below are both scalar amounts. Here, the target error δy of the observation amount is defined by equation (32), where p is the parameter value before updating.

[0106]

number

[0107] For example, as shown in equation (33), the formula for calculating the parameter update amount will be explained when the sensitivity relational expression is approximated by a quadratic function. The parameter update amount δp at each step is calculated from the formula for solving the quadratic equation using the perturbation amount δx (= p-p0) from the parameter reference value p0 at the time of sensitivity calculation and the target error δy of the observation amount. Note that since a often takes a small value, it is desirable to move the root sign to the denominator and rationalize it as in equation (33) to prevent cancellation of digits. The sign before the root sign is differentiated depending on the sign of a and the axis position of the parabola.

[0108]

number

[0109] There is roughly the following correspondence between each parameter and the observable: The observable corresponding to the indirect reduction reaction is the gas utilization rate ηCO or CSL. The observable corresponding to the coke gasification reaction is the gas utilization rate ηCO or CSL that was not selected as the observable corresponding to the indirect reduction reaction. The observable corresponding to the solid-liquid heat exchange coefficient is the hot metal temperature. The observable corresponding to the SiO2 reduction reaction is Si in the hot metal. The observable corresponding to the direct reduction reaction is slag FeO. The observable corresponding to the carburization reaction is C in the hot metal. The observable corresponding to the hydrogen gas shift reaction is the hydrogen utilization rate ηH2. The observable corresponding to the porosity correction parameter is the blast pressure. For each parameter, the parameter update calculation of Equation (32) is performed in order to eliminate the target error of the corresponding observation quantity, and the above parameter update calculation is repeated until the model calculation value converges to the actual value for all observation quantities, ultimately satisfying the predetermined convergence judgment condition.

[0110] Note that the above calculation procedure may result in slow convergence, oscillation, or divergence. To prevent this, the following procedure can be used to speed up and stabilize the parameter update calculation.

[0111] First, an 8x8 matrix is ​​defined, with observables arranged in the rows and parameters arranged in the columns. Each element stores 0 if the parameter has no effect on the observable, and 1 if the parameter has an effect. Here, "no effect of a parameter" refers to the case where the absolute value of the sensitivity of the observable to the parameter perturbation is smaller than a predetermined appropriate threshold, and "effect of a parameter" refers to the case where the absolute value of the sensitivity of the observable to the parameter perturbation is equal to or greater than a predetermined appropriate threshold. Furthermore, this matrix is ​​rearranged into a block lower triangular matrix by DM decomposition (Dulmage-Mendelsohn decomposition, described in Non-Patent Document 6). Rearranging using DM decomposition yields a block lower triangular matrix such as that shown in Equation (34). Here, the row-wise observation quantities are FeO in the slag, C concentration in the hot metal, H2 utilization rate, blast pressure, CO2 utilization rate, CSL, hot metal temperature, and Si in the hot metal, and the column-wise parameters are direct reduction reaction, carburization reaction, hydrogen gas shift reaction, porosity correction parameter, indirect reduction reaction, coke gas reaction, SiO2 reduction reaction, and solid-liquid heat exchange coefficient. This rearranged matrix is ​​referred to as a parameter influence matrix. The calculation order of the parameter adjustment calculations, which will be described later, can be determined based on the structure of the parameter influence matrix.

[0112]

number

[0113] In the parameter influence matrix of Eq. (34), the structure of the 4x4 matrix in the upper left corner indicates that the carburization reaction, direct reduction reaction, hydrogen gas shift reaction, and porosity correction parameter are not related to each other, and their influences can be considered independent. Therefore, these four can be updated simultaneously. Additionally, the structure of the 4x4 matrix in the lower right corner indicates that the indirect reduction reaction, SiO2 reduction reaction, solid-liquid heat exchange coefficient, and coke gasification reaction are all interconnected, and their respective effects cannot be considered independent. For this reason, these four parameters are updated in order rather than together. However, based on the formulation described below, these four can also be updated together, taking into account their interconnections. To summarize the above, for example, the following steps (1) to (5) constitute one cycle, and the parameter adjustment calculation may be carried out by repeating the cycle until the parameters converge. (1) Carburization reaction, direct reduction reaction, hydrogen gas shift reaction, and porosity correction parameter (2) Indirect reduction reaction (3) Coke gasification reaction (4) SiO2 reduction reaction (5) Solid-liquid heat exchange coefficient Note that the above (1) to (5) are in no particular order. 11A and 11B are diagrams for explaining an example of a parameter fitting calculation, in which (a) is a characteristic diagram showing an example of the relationship between the calculation step and the gas utilization rate η, which is an observation quantity of the indirect reduction reaction, and (b) is a characteristic diagram showing an example of the relationship between the calculation step and the parameters of the indirect reduction reaction. As shown in FIGS. 11A and 11B, the parameters converge so that the actual values ​​(target values) of the observation quantities match the calculated values.

[0114] Furthermore, based on the structure of the parameter influence matrix, the indirect reduction reaction, SiO2 reduction reaction, solid-liquid heat exchange coefficient, and coke gasification reaction are all interconnected, and their influences cannot be considered independent. Therefore, by taking these interconnections into consideration and solving the nonlinear programming problem in equation (35), (2), (3), (4), and (5) can be calculated together. By calculating them together, it is possible to reduce the calculation time. Given: weight Q i ,R i >0, error δ i , upper and lower limits Δ pi,min ,Δ pi,max , quadratic function approximation parameter a ij ,b ij Decision variable (parameter update amount): Δp i

[0115]

number

[0116] According to the procedure described above, model parameters that satisfy the simultaneous equations (Equation (30)), which are nonlinear equations, can be accurately calculated by using sensitivities approximated by quadratic functions, exponential functions, or the like, without using an accurately calculated gradient ∂F / ∂p. Note that this calculation involves fitting steady-state calculated values ​​of the blast furnace mathematical model, and since these steady-state calculated values ​​do not depend on the initial state at the start of the unsteady-state calculation of the blast furnace mathematical model, the calculation may be started in any initial state. In particular, in order to speed up the calculation, the calculation result at the end of the previous calculation may be set as the initial state. Furthermore, the sensitivity calculation that is performed at the beginning of the parameter fitting calculation may be performed each time a parameter fitting calculation is performed, or in order to shorten the calculation time, the calculation may be omitted each time and the sensitivity calculation result calculated in advance may be used instead.

[0117] The processing performed by the linear approximation calculation unit 43 will now be described. The unsteady calculation using the blast furnace mathematical model calculated by the parameter fitting is performed to obtain the operation variable u∈R m is the input of the blast furnace mathematical model, and the control variable z∈R r Observable y∈R including p An appropriate time response is calculated with the output being . For example, each manipulated variable can be changed stepwise independently to calculate the time response (step response). This time response data is fitted with a transfer function of a linear system such as a first-order or second-order lag system, and a transfer function matrix of size p × m is created with the observed variables arranged in the row direction and the manipulated variables arranged in the column direction. In addition to the manipulated variable u, the disturbance d∈R mdThe same calculation may be performed by giving the above and converting it into a transfer function matrix of size p × (m + md). In addition, the transfer function to be fitted may take into account the dead time corresponding to the influence delay of the manipulated variable. Here, the manipulated variables are, for example, the blast rate, a value related to the amount of pulverized coal, a value related to oxygen enrichment, the blast temperature, the moisture content, the amount of liquefied natural gas, the coke ratio or O / C, etc. The disturbances are, for example, the particle size and composition of the charged raw materials (ore, coke), etc. The observed variables are the controlled variables including the iron production rate and furnace heat indicators (hot metal temperature, Si in the molten metal, amount of carbon solution loss, amount of direct reduction reaction, heat balance above the tuyere, and cohesive zone level, etc.), as well as the gas composition, gas utilization rate, amount of carbon solution loss, amount of direct reduction reaction, heat balance above the tuyere, furnace top temperature, heat flow ratio, and blast pressure, etc.

[0118] Figure 12(a) is a characteristic diagram showing the calculation results of the time response of the iron production rate to the operation of changing the blast rate. As shown in Figure 12(a), the linear approximation calculation unit 43 calculates the time response of the iron production rate to the operation of changing the blast rate control amount. This time response can be approximated by a first-order lag transfer function, as shown by the dotted line in the figure.

[0119] Figure 12(b) is a characteristic diagram showing the calculation results of the time response of the molten iron temperature to the operation of changing the blast rate. As shown in Figure 12(b), the linear approximation calculation unit 43 calculates the time response of the molten iron temperature to the operation of changing the blast rate control amount. This time response can be approximated by a second-order lag transfer function, as shown by the dotted line in the figure.

[0120] If the transfer function matrix obtained in this way is state-realized in an appropriate canonical form such as a controllable canonical form and then time-discretized at an appropriate sampling time using a method such as zero-order hold, a discrete-time state-space model can be obtained. However, since the above transfer function matrix usually contains redundant dynamics, it is not possible to obtain a state-space model that satisfies controllability and observability in this state. Therefore, a balanced realization method is applied to construct a controllable and observable state-space model. The balanced realization method is a reduction method for MIMO (multiple-input, multiple-output) linear models, and can reduce the original model to a controllable and observable linear state-space model (see Non-Patent Document 3).

[0121] Here, we explain the reason for creating a controllable and observable linear state space model. Generally, for a controllable and observable linear state space model, a state estimation calculation method for variables in the model that uses observable quantities, such as a Luenberger observer, can be applied. However, a blast furnace mathematical model is a large-scale nonlinear model, and the establishment of properties equivalent to observability in a linear system is not guaranteed. Therefore, state estimation calculation methods such as a Luenberger observer cannot be applied as is. Similarly, the establishment of properties equivalent to controllability in a linear system is not guaranteed, so calculating the optimal manipulated variable in model predictive control is also difficult. In the case of blast furnace operation, linearity is generally established by limiting the operating range of the operating point, so it is effective to replace the blast furnace mathematical model in a relatively stable state with a linear approximation model. Furthermore, by using a linear approximation model that ensures observability and controllability through the above procedure, it becomes possible to calculate the optimal operating variables using state estimation calculations and model predictive control, making it possible to implement optimal automatic blast furnace control.

[0122] In this embodiment, when the judgment unit 41 judges that the conditions 1 to 3 are satisfied, the parameter fitting calculation unit 42 performs a parameter fitting calculation for the current blast furnace mathematical model. However, for example, if the blast furnace continues to be in a stable state, the parameter fitting calculation may be omitted, and representative parameters may be fixed, and then the linear approximation calculation unit 43 may perform a linear approximation calculation around the operating point. Furthermore, in order to update the state space model bumplessly after the linear approximation calculation is performed, calculation processing such as smoothing may be performed as appropriate immediately after the model update, and the state variables may be updated appropriately in advance.

[0123] The above-described embodiments are merely examples of specific implementations of the present invention, and the technical scope of the present invention should not be construed as being limited by these embodiments. In other words, the present invention can be implemented in various forms without departing from its technical concept or main features. The control function of the blast furnace process to which the present invention is applied can also be realized by supplying software (program) to a system or device via a network or various storage media, and having a computer in the system or device read and execute the program. The control function of the blast furnace process to which the present invention is applied can also be realized by a PLC (Programmable Logic Controller) or dedicated hardware such as an ASIC.

[0124] It can also be written as follows: [Claim 1] A control device that performs model predictive control using a state space model that represents time changes in state variables, with a process as a control target, Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; Assuming that the design allows input disturbances to be superimposed on the manipulated variable and the design allows output disturbances to be superimposed on the observed variable, a control device comprising: an estimation means for receiving as input the manipulated variable and the observed variable on which at least one of the input disturbance and the output disturbance is superimposed, and outputting an estimated value of the state variable and at least one of the estimated value of the input disturbance and the estimated value of the output disturbance. [Claim 2] 2. The control device according to claim 1, wherein an input to said estimation means includes an observation amount other than said controlled amount. [Claim 3] 3. The control device according to claim 1, wherein the process is a process including physical phenomena with different time constants. [Claim 4] 4. The control device according to claim 1, further comprising control means for executing the model predictive control using a target value and an output of the estimation means as inputs, thereby outputting the manipulated variable. [Claim 5] the process is a blast furnace process; the control variables include a production rate and a furnace heat index; 5. The control device according to claim 1, wherein the furnace heat indicators include any one of molten iron temperature, silicon in molten iron, amount of carbon solution loss, amount of direct reduction reaction, heat balance above the tuyere, and cohesive zone level. [Claim 6] the process is a blast furnace process; 6. The control device according to claim 1, wherein the manipulated variable includes any one of a value relating to the blast rate, a value relating to the amount of pulverized coal, a value relating to oxygen enrichment, a value relating to the blast temperature, a value relating to the moisture content, and a value relating to the amount of liquefied natural gas, a coke ratio, and an O / C ratio. [Claim 7] the process is a blast furnace process; 7. The control device according to claim 1, wherein the observation variables include the control variables and also include one of a gas composition, a gas utilization rate, an amount of carbon solution loss, an amount of direct reduction reaction, a heat balance above the tuyere, a furnace top temperature, a heat flow ratio, and a blast pressure that is not included in the control variables. [Claim 8] 8. The control device according to claim 1, wherein a constraint condition is set for at least one of the manipulated variable and the controlled variable. [Claim 9] 9. The control device according to claim 8, wherein violations of some of the constraints are permitted. [Claim 10] 10. The control device according to claim 9, wherein priorities are set among the part of the constraint conditions. [Claim 11] 11. The control device according to claim 1, wherein the sum of the dimensions of the estimated value of the input disturbance and the estimated value of the output disturbance is equal to or greater than the dimension of the controlled variable and equal to or less than the dimension of the observed variable. [Claim 12] 12. A control device according to any one of the preceding claims, characterized in that it comprises correction means for performing a feedforward correction of the control action on the basis of measurable disturbances. [Claim 13] 13. The control device according to claim 1, further comprising an update unit that updates the state space model when a predetermined condition is met. [Claim 14] 14. The control device according to claim 13, wherein the predetermined condition is at least one of a change in a predetermined operating condition of the process, a change in a predetermined property of the process, a predetermined time being reached, and an occurrence of a predetermined event in the process. [Claim 15] 15. The control device according to claim 13, wherein the updating means comprises parameter fitting calculation means for performing parameter fitting calculations on a mathematical model used to create the state space model. [Claim 16] 16. The control device according to claim 15, wherein the parameter fitting calculation means uses the mathematical model to perform sensitivity calculations on the parameters and updates the parameters based on the structure of a parameter influence matrix rearranged in a block lower triangular structure. [Claim 17] the process is a blast furnace process; the mathematical model is a blast furnace mathematical model; 17. The control device according to claim 15, wherein the parameter fitting calculation means performs parameter fitting calculations on at least one of a correction coefficient for each reaction rate constant of an indirect reduction reaction, a coke gasification reaction, a SiO reduction reaction, a direct reduction reaction, a carburization reaction, and a hydrogen gas shift reaction, a solid-liquid heat exchange coefficient, and a porosity correction parameter as a calculation target. [Claim 18] 18. The control device according to claim 15, wherein the updating means creates the state space model from the mathematical model before the parameter fitting calculation by the parameter fitting calculation means or the mathematical model after the parameter fitting calculation by the parameter fitting calculation means, and comprises linear approximation calculation means for performing linear approximation calculation around an operating point at that time. [Claim 19] the state-space model is a discrete-time linear state-space model; 19. The control device according to claim 18, wherein the linear approximation calculation means utilizes the mathematical model, fits a time response in which the manipulated variable is input and the controlled variable is output with a transfer function of a first-order lag system or a second-order lag system, constructs a transfer function matrix, and then applies a balanced realization method to construct the controllable and observable discrete-time linear state space model. [Claim 20] A control method for performing model predictive control using a state space model representing time changes in state variables for a process as a control target, comprising: Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; Assuming that the design allows input disturbances to be superimposed on the manipulated variable and the design allows output disturbances to be superimposed on the observed variable, a step of receiving as input the manipulated variable and the observed variable on which at least one of the input disturbance and the output disturbance is superimposed, and outputting an estimate of the state variable and at least one of the estimate of the input disturbance and the estimate of the output disturbance. [Claim 21] A program for performing model predictive control using a state space model that represents time changes in state variables for a process as a control target, Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; Assuming that the design allows input disturbances to be superimposed on the manipulated variable and the design allows output disturbances to be superimposed on the observed variable, a program for causing a computer to function as estimation means for receiving as input the manipulated variable and the observed variable on which at least one of the input disturbance and the output disturbance is superimposed, and outputting an estimated value of the state variable and at least one of the estimated value of the input disturbance and the estimated value of the output disturbance. [Explanation of symbols]

[0125] 1: Blast furnace process, 2: Controller, 3: Disturbance observer, 4: Update device, 41: Judgment unit, 42: Parameter fitting calculation unit, 43: Linear approximation calculation unit

Claims

1. A control device that performs model predictive control using a state space model that represents time changes in state variables, with a process including physical phenomena with different time constants as a control target, comprising: Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; an estimation means for receiving the manipulated variable and the observed variable as inputs, and outputting, based on a design of the state space model in which an input disturbance is superimposed on the manipulated variable and an output disturbance is superimposed or not superimposed on the observed variable, an estimate of the state variable and an estimate of the disturbance designed to be superimposed in the state space model; the controlled variable includes an index for a predetermined physical phenomenon, the observable includes the controlled variable and an index for a physical phenomenon having a faster response than the predetermined physical phenomenon; 10. A control device according to claim 9, wherein the input disturbance includes an index of a factor that affects an index for the fast-responding physical phenomenon included in the observable.

2. The process is a blast furnace process, the indicator for the predetermined physical phenomenon included in the controlled variable includes a molten iron temperature; The indicators for the fast-responding physical phenomena included in the observables include a heat balance above the tuyere; 2. The control system of claim 1, wherein the input disturbance includes a change in heat of pulverized coal decomposition that affects the heat balance above the tuyere.

3. 3. The control device according to claim 1, wherein a constraint condition is set for at least one of the manipulated variable and the controlled variable.

4. 4. The control device according to claim 3, wherein violations of some of the constraints are permitted.

5. 5. The control device according to claim 4, wherein priorities are set among the part of the constraint conditions.

6. 3. The control device according to claim 1, wherein the sum of the dimensions of the estimated value of the input disturbance and the estimated value of the output disturbance is equal to or larger than the dimension of the controlled variable and equal to or smaller than the dimension of the observed variable.

7. 3. A control system according to claim 1, further comprising correction means for performing a feedforward correction of the control action based on measurable disturbances.

8. 3. The control device according to claim 1, further comprising an update means for updating the state space model when a predetermined condition is met.

9. 9. The control device according to claim 8, wherein the predetermined condition is at least one of a change in a predetermined operating condition of the process, a change in a predetermined property of the process, a predetermined time being reached, and an occurrence of a predetermined event in the process.

10. 9. The control device according to claim 8, wherein the updating means comprises parameter fitting calculation means for performing parameter fitting calculations on a mathematical model used to create the state space model.

11. 11. The control device according to claim 10, wherein the parameter fitting calculation means uses the mathematical model to perform sensitivity calculations for parameters, determines an update calculation order for parameters based on a structure of a parameter influence matrix rearranged in a block lower triangular structure, and updates the parameters in accordance with the determined update calculation order.

12. the process is a blast furnace process; the mathematical model is a blast furnace mathematical model; The parameter fitting calculation means calculates the indirect reduction reaction, the coke gasification reaction, the SiO 2 11. The control device according to claim 10, wherein the parameter fitting calculation is performed on at least one of a correction coefficient for each reaction rate constant of a reduction reaction, a direct reduction reaction, a carburization reaction, and a hydrogen gas shift reaction, a solid-liquid heat exchange coefficient, and a porosity correction parameter as a calculation target.

13. 11. The control device according to claim 10, wherein the update means creates the state space model from the mathematical model before the parameter fitting calculation by the parameter fitting calculation means or the mathematical model after the parameter fitting calculation by the parameter fitting calculation means, and comprises linear approximation calculation means for performing linear approximation calculation around an operating point at that time.

14. the state-space model is a discrete-time linear state-space model; 14. The control device according to claim 13, wherein the linear approximation calculation means utilizes the mathematical model, fits a time response in which the manipulated variable is input and the controlled variable is output with a transfer function of a first-order lag system or a second-order lag system, constructs a transfer function matrix, and then applies a balanced realization method to construct the controllable and observable discrete-time linear state space model.

15. A control method for performing model predictive control using a state space model representing time changes in state variables, with a process including physical phenomena with different time constants as a control target, comprising: Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; a step of using the manipulated variable and the observed variable as inputs, and outputting, based on a design in which an input disturbance is superimposed on the manipulated variable and an output disturbance is superimposed or not superimposed on the observed variable, an estimated value of the state variable and an estimated value of the disturbance designed to be superimposed in the state space model, the controlled variable includes an index for a predetermined physical phenomenon, the observable includes the controlled variable and an index for a physical phenomenon having a faster response than the predetermined physical phenomenon; the input disturbance includes an index of a factor that affects the index of the fast-response physical phenomenon included in the observable; A control method comprising:

16. A program for causing a computer to execute model predictive control using a state space model that represents time changes in state variables, with a process including physical phenomena with different time constants as a control target, comprising: Observation quantities including control quantities output from the process when manipulated variables are input to the process are acquired; making the computer function as an estimation means that receives the manipulated variable and the observed variable as inputs, and outputs, based on a design in which an input disturbance is superimposed on the manipulated variable and an output disturbance is superimposed or not superimposed on the observed variable, an estimated value of the state variable and an estimated value of the disturbance that is designed to be superimposed in the state space model; the controlled variable includes an index for a predetermined physical phenomenon, the observable includes the controlled variable and an index for a physical phenomenon having a faster response than the predetermined physical phenomenon; The input disturbance includes an index of a factor that affects an index for the fast-responding physical phenomenon included in the observable.

Images