State prediction device, refining control system, state prediction method, and refining control method

Abstract

This state prediction device comprises a control unit that: acquires a refining condition prior to a refining treatment and an expected value for a manipulated variable; calculates the degree of similarity between the refining condition prior to the refining treatment and the expected value for the manipulated variable, and a past refining condition actual value and past manipulated variable actual value in past refining operations; generates a prediction model that associates the past refining condition actual value and the past manipulated variable actual value with past refining result actual values; determines a parameter of the prediction model by solving an optimization problem, where an evaluation function that has the degree of similarity as a weight serves as the evaluation function for evaluating prediction errors in the prediction model; and calculates refining results including a plurality of post-refining treatment predicted values from among a predicted value for a molten metal component, a predicted value for a molten metal temperature, and a predicted value for a slag component by inputting, into the prediction model, the refining condition prior to the refining treatment and the expected value for the manipulated variable.

Description

State prediction device, refining control system, state prediction method, and refining control method

[0001] This disclosure relates to a state prediction device, a refining control system, a state prediction method, and a refining control method.

[0002] In steel mills, molten iron extracted from blast furnaces is processed in refining facilities such as pre-treatment facilities, converters, and secondary refining facilities to adjust its composition and temperature. Refining facilities remove impurities from the molten metal and raise its temperature by adding auxiliary materials to the ladle or furnace and blowing in oxygen. This process plays a crucial role in steel quality control and cost-efficiency in manufacturing.

[0003] In smelting facilities, information about the molten metal composition and temperature obtained during processing is limited. Therefore, operational quantities such as the amount of acid supplied and the amount of auxiliary materials added must be appropriately determined at the start of the smelting process, after predicting the molten metal composition and temperature after smelting. Consequently, accurately predicting the molten metal composition and temperature after smelting is extremely important in smelting processes.

[0004] For example, Patent Document 1 proposes a method for estimating the phosphorus concentration after refining by calculating the dephosphorization rate based on a statistical model. Also, for example, Patent Document 2 proposes a method for accurately predicting the phosphorus concentration after refining by generating a model based on past performance under refining conditions similar to those of the target material.

[0005] Japanese Patent Publication No. 2024-5899, Japanese Patent No. 5821656

[0006] Patent Document 1 exemplifies the use of a multiple regression model as a statistical model. However, since refining reactions often involve nonlinear relationships, it is considered difficult to estimate the phosphorus concentration with high accuracy when using a multiple regression model.

[0007] Patent Document 2 generates a model based on past results of refining conditions similar to those of the prediction target, so it is considered that phosphorus concentration can be predicted with high accuracy even when there is a non-linear relationship. However, Patent Document 2 only assumes a case where there is one prediction target. Refining equipment generally has a plurality of control targets such as molten metal components, molten metal temperature, and slag components, but Patent Document 2 does not assume prediction for a plurality of control targets.

[0008] An object of the present disclosure is to provide a state prediction device, a refining control system, a state prediction method, and a refining control method that enable highly accurate prediction of refining results for a plurality of targets.

[0009] [1] A state prediction device that predicts a refining result in refining equipment, the device comprising: a control unit configured to: acquire refining conditions before a refining process for a prediction target, the refining conditions including a measured value of a molten metal component and a measured value of a molten metal temperature before the refining process for the prediction target, and a planned value of an operation amount including an auxiliary raw material input amount and an oxygen feed amount in the refining process for the prediction target; calculate a similarity between the set of the refining conditions before the refining process and the planned value of the operation amount, and a set of an actual value of past refining conditions and an actual value of a past operation amount in past refining results; generate a prediction model that correlates the actual value of the past refining conditions, the actual value of the past operation amount, and an actual value of a past refining result; determine a parameter of the prediction model by solving an optimization problem using an evaluation function with the similarity as a weight as an evaluation function for evaluating a prediction error of the prediction model; and calculate the refining result including a plurality of prediction values among a predicted value of a molten metal component after the refining process, a predicted value of the molten metal temperature, and a predicted value of a slag component by inputting the refining conditions before the refining process and the planned value of the operation amount into the prediction model, wherein the state prediction device comprises the control unit.

[0010] [2] The state prediction device according to [1] above, wherein the control unit calculates the similarity by applying a monotonic decreasing function to a distance between the set of the refining conditions before the refining process and the planned value of the operation amount, and the set of the actual value of the past refining conditions and the actual value of the past operation amount.

[0011] [3] The state prediction device according to [2] above, wherein the control unit applies a function that decreases exponentially according to the distance as the monotonic decreasing function, and calculates the similarity such that newer actual refining data has a relatively larger similarity due to a forgetting factor α.

[0012] [4] The state prediction device according to [2] or [3] above, wherein the control unit calculates a difference between the planned value and the actual value for each item, and calculates the distance based on a sum of the differences or a sum of squares of the differences.

[0013] [5] The state prediction device according to any one of [2] to [4] above, wherein prior to calculating the distance, the control unit standardizes scales of each item of the planned value and the actual value, and calculates the distance for items after linear transformation and dimension reduction using a loading matrix P obtained by principal component analysis.

[0014] [6] The state prediction device according to any one of [1] to [5] above, wherein the control unit: as prediction models for predicting a refining result, generates a plurality of prediction models having at least two or more of the molten metal component, the molten metal temperature and the slag component as objective variables; and calculates a predicted value of the objective variable of the prediction model by inputting, into the prediction model, a predicted value or a target value of an objective variable other than the objective variable of the prediction model, in addition to refining conditions before the refining treatment and a planned value of the operation amount.

[0015] [7] The state prediction device according to any one of [1] to [6] above, wherein the control unit repeatedly calculates the predicted value of each objective variable using the plurality of prediction models until a difference or a change rate between an input value used for prediction of an objective variable of one prediction model among the plurality of prediction models and a predicted value of another prediction model having the input value as an objective variable falls below a threshold value.

[0016] [8] The state prediction device according to any one of the above items [1] to [7], wherein the control unit calculates the refining result by inputting the refining conditions before the refining process and the planned values ​​of the manipulated amount into a separate prediction model prepared in advance, if the number of past performance data for which the similarity is less than or equal to the first threshold is less than or equal to a predetermined number, or if the value of the evaluation function is greater than or equal to the second threshold as a result of solving the optimization problem.

[0017] [9] A refining control system comprising a state prediction device as described in any one of [1] to [8] above, and a control terminal for controlling various operational quantities in a refining process, wherein the state prediction device further comprises a communication unit capable of communicating the refining results to the control terminal, and the control terminal calculates a set value for an operational quantity, including at least one of the acid supply amount, acid supply rate, stirring gas flow rate, lance height, and auxiliary material input amount, based on the refining results, such that at least one of the molten metal components, molten metal temperature, or slag components after the refining process is within a desired range, and generates a control signal based on the set value to control the actuator of the refining equipment.

[0018]

[10] A state prediction method for predicting the refining result in a refining facility, comprising: a step of obtaining refining conditions before the refining process, including measured values ​​of the molten metal components and measured values ​​of the molten metal temperature before the refining process to be predicted, and planned values ​​of the operating quantities, including the amount of auxiliary materials added and the amount of acid supplied in the refining process to be predicted; a step of calculating the similarity between the planned values ​​of the refining conditions before the refining process and the operating quantities and the actual values ​​of the past refining conditions and the actual values ​​of the past operating quantities in past refining results; a step of generating a prediction model that relates the actual values ​​of the past refining conditions and the actual values ​​of the past operating quantities to the actual values ​​of past refining results; a step of determining the parameters of the prediction model by solving an optimization problem using an evaluation function that uses the similarity as a weight as an evaluation function that evaluates the prediction error of the prediction model; and a step of calculating the refining result, including a plurality of predicted values ​​among the predicted values ​​of the molten metal components, the predicted value of the molten metal temperature and the predicted value of the slag components after the refining process, by inputting the planned values ​​of the refining conditions before the refining process and the operating quantities into the prediction model.

[0019]

[11] A refining control method that, based on the refining result calculated by the state prediction method described in

[10] above, calculates a set value for an operating variable including at least one of the acid supply amount, acid supply rate, stirring gas flow rate, lance height, and auxiliary material input amount, such that at least one of the molten metal components, molten metal temperature, or slag components after the refining process is within a desired range, and generates a control signal based on the set value to control the actuator of the refining equipment.

[0020] According to the state prediction device, refining control system, state prediction method, and refining control method described herein, the refining results of multiple objects in a refining facility can be predicted with high accuracy.

[0021] This is a schematic diagram showing an example of a refining facility according to one embodiment of the present disclosure. This is a block diagram showing an example of the configuration of a state prediction device according to one embodiment of the present disclosure. This is a flowchart showing an example of the operation of a state prediction device according to one embodiment of the present disclosure. This is a flowchart showing an example of the operation of a state prediction device according to one embodiment of the present disclosure. This is a diagram showing the relationship between the actual value of the P concentration after treatment and the predicted value of the P concentration after treatment for Comparative Example 1, Example 1 and Example 2.

[0022] The embodiments of this disclosure will be described below with reference to the drawings.

[0023] Figure 1 is a schematic diagram showing an example of a refining facility 1 according to one embodiment of the present disclosure. The refining facility 1 comprises a state prediction device 10, a control terminal 20, a converter 30, a lance 40, a duct 50, and a flow meter 60. The state prediction device 10 and the control terminal 20 constitute a refining control system.

[0024] The state prediction device 10 predicts the refining results in the refining facility 1. For example, the state prediction device 10 predicts the refining results after the refining process for the molten metal 101 and slag 102 in the converter 30. The refining results predicted by the state prediction device 10 may include multiple predicted values, such as predicted values ​​of the molten metal components of the molten metal 101 after the refining process, predicted values ​​of the molten metal temperature of the molten metal 101 after the refining process, and predicted values ​​of the slag components of the slag 102 after the refining process. Here, in this specification, "refining results" refers to actual values ​​(measured values) for at least one of the molten metal components, molten metal temperature, and slag components after the refining process. The predicted values ​​of the refining results calculated by the state prediction device 10 refer to the predicted values ​​(or information based on these) for each item corresponding to the refining results. In this embodiment, the predicted value of the molten metal components of the molten metal 101 means the predicted value of the concentration of the molten metal components of the molten metal 101. Furthermore, in this embodiment, the predicted value of the slag component of slag 102 refers to the predicted value of the concentration of the slag component of slag 102.

[0025] Hereafter, the "molten metal components of molten metal 101" may be simply referred to as "molten metal components." Similarly, the "molten metal temperature of molten metal 101" may be simply referred to as "molten metal temperature." Furthermore, the "slag components of slag 102" may be simply referred to as "slag components."

[0026] The state prediction device 10 may be a general-purpose computer such as a workstation or personal computer, or it may be a dedicated computer configured to function as a state prediction device 10.

[0027] Figure 2 is a block diagram showing an example of the configuration of a state prediction device 10 according to one embodiment of the present disclosure.

[0028] The state prediction device 10 comprises a control unit 11, an input unit 12, an output unit 13, a storage unit 14, and a communication unit 15.

[0029] The control unit 11 includes at least one processor, at least one dedicated circuit, or a combination thereof. The processor is a general-purpose processor such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit), or a dedicated processor specialized for a specific process. The dedicated circuit is, for example, an FPGA (Field-Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit).

[0030] The control unit 11 reads programs, data, etc., stored in the storage unit 14 and executes various functions.

[0031] The input unit 12 includes one or more input interfaces that detect user input and acquire input information based on user operations. The input unit 12 includes, for example, physical keys, capacitive keys, a touchscreen integrated with the display of the output unit 13, or a microphone that accepts voice input.

[0032] The output unit 13 includes one or more output interfaces that output information and notify the user. The output unit 13 includes, for example, a display that outputs information as an image, a speaker that outputs information as sound, etc. The display included in the output unit 13 may be, for example, an LCD (Liquid Crystal Display), a CRT (Cathode Ray Tube) display, etc.

[0033] The storage unit 14 is, for example, a flash memory, a hard disk, or an optical memory. Part of the storage unit 14 may be located outside the state prediction device 10. In this case, part of the storage unit 14 may be a hard disk, memory card, or the like, connected to the state prediction device 10 via an arbitrary interface.

[0034] The memory unit 14 stores programs for the control unit 11 to execute various functions, data used by those programs, and so on.

[0035] The communication unit 15 includes at least one of a communication module that supports wired communication and a communication module that supports wireless communication. The state prediction device 10 can communicate with the control terminal 20 via the communication unit 15.

[0036] Details of the operation of the state prediction device 10 will be described later.

[0037] The control terminal 20 may be a general-purpose computer such as a workstation or personal computer, or it may be a dedicated computer configured to function as a control terminal 20.

[0038] The control terminal 20 can control various parameters in the refining process so that the molten metal composition and temperature of the molten metal 101 after refining are within a desired range. These parameters may include, for example, the amount of auxiliary materials added to the converter 30 and the amount of high-pressure oxygen supplied from the lance 40 to the molten metal 101 (acid supply amount). The parameters may further include, for example, the supply rate of high-pressure oxygen (acid supply rate), the flow rate of the stirring gas, and the height of the lance 40 (lance height). In this specification, the acid supply amount may include the integrated amount obtained by integrating the acid supply rate over time.

[0039] The control terminal 20 may control various operations in the refining process in response to commands from the state prediction device 10.

[0040] Furthermore, the control terminal 20 can collect actual values ​​of various operations in the refining process, measured values ​​such as molten metal components, molten metal temperature, and slag components, and other information related to the refining equipment 1. The control terminal 20 transmits the collected information to the state prediction device 10.

[0041] The converter 30 contains the molten metal 101 that is to be refined.

[0042] The lance 40 is positioned above the molten metal 101 contained in the converter 30. The lance 40 can eject high-pressure oxygen from its tip toward the molten metal 101 contained in the converter 30.

[0043] When high-pressure oxygen is ejected from the tip of the lance 40 into the molten metal 101, impurities in the molten metal 101 are oxidized by the high-pressure oxygen. The oxidized impurities are incorporated into the slag 102. This type of process is called refining.

[0044] The duct 50 is installed above the converter 30. The duct 50 can discharge exhaust gases and other substances generated during the refining process.

[0045] A ventilation hole 31 is formed at the bottom of the converter 30. A stirring gas is blown into the molten metal 101 inside the converter 30 through the ventilation hole 31. The stirring gas is an inert gas such as argon (Ar).

[0046] When the stirring gas is blown into the molten metal 101, the stirring gas agitates the molten metal 101. This agitation promotes the reaction between the high-pressure oxygen ejected from the lance 40 and the molten metal 101.

[0047] The flow meter 60 measures the flow rate of the stirring gas blown into the molten metal 101 through the vent hole 31.

[0048] Auxiliary materials are fed into the converter 30 from the top of the furnace. The auxiliary materials may include a heat-raising material containing carbon (C), silicon (Si), etc., a coolant containing iron oxide, etc., and a solvent containing calcium (Ca), manganese (Mn), magnesium (Mg), etc.

[0049] (Operation of the state prediction device) Next, the operation of the state prediction device 10 will be explained.

[0050] The memory unit 14 stores performance data of refining processes performed in the past. For each refining process, the past performance data is stored in association with the actual values ​​of the refining conditions, the actual values ​​of the manipulated quantities, the actual values ​​of the refining results, the predicted values ​​of the refining results, and the prediction model used to predict the refining results.

[0051] The actual values ​​for the refining conditions include the measured values ​​of the molten metal components and the measured value of the molten metal temperature before the refining process. The control terminal 20 collects the measured values ​​of the molten metal components and the molten metal temperature before the refining process from the measuring device and transmits them to the state prediction device 10. The control unit 11 of the state prediction device 10 acquires the measured values ​​of the molten metal components and the molten metal temperature before the refining process transmitted by the control terminal 20 via the communication unit 15 and stores them in the storage unit 14.

[0052] The actual values ​​of the manipulated quantities include the actual values ​​of the amount of auxiliary materials added and the amount of acid supplied during the refining process. The control unit 11 of the state prediction device 10 stores the amount of auxiliary materials added and the amount of acid supplied instructed to the control terminal 20 in the storage unit 14 as the actual values ​​of the amount of auxiliary materials added and the actual values ​​of the amount of acid supplied.

[0053] The actual refining results include measured values ​​of the molten metal components after refining, measured values ​​of the molten metal temperature after refining, and measured values ​​of the slag components after refining. The control terminal 20 collects the measured values ​​of the molten metal components, molten metal temperature after refining, and slag components after refining from the measuring device and transmits them to the state prediction device 10. The control unit 11 of the state prediction device 10 acquires the measured values ​​of the molten metal components, molten metal temperature after refining, and slag components after refining transmitted by the control terminal 20 via the communication unit 15 and stores them in the storage unit 14.

[0054] The predicted refining result is the refining result predicted by the state prediction device 10 in past refining processes. The predicted refining result includes the predicted values ​​of the molten metal components, molten metal temperature, and slag components after refining, which were calculated by the state prediction device 10 in past refining processes. The control unit 11 of the state prediction device 10 stores the calculated predicted values ​​of the molten metal components, molten metal temperature, and slag components after refining in the storage unit 14.

[0055] The prediction model used to predict the refining results is the same prediction model that the state prediction device 10 used to predict the refining results in past refining processes. The control unit 11 of the state prediction device 10 stores the prediction model that the state prediction device 10 used to predict the refining results in the storage unit 14.

[0056] Before executing the refining process for which the refining result is to be predicted, the control unit 11 acquires the refining conditions before executing the refining process to be predicted and the planned values ​​of the manipulated quantities in the refining process to be predicted.Hereafter, "refining conditions before executing the refining process to be predicted" may be simply referred to as "refining conditions before the refining process."Hereafter, "planned values ​​of the manipulated quantities in the refining process to be predicted" may be simply referred to as "planned values ​​of the manipulated quantities."

[0057] The refining conditions before performing the refining process to be predicted may include, for example, measured values ​​of the molten metal components and the molten metal temperature before the refining process. The control unit 11 may obtain the measured values ​​of the molten metal components and the molten metal temperature before the refining process from the control terminal 20.

[0058] Furthermore, the control unit 11 may acquire information as refining conditions before executing the refining process to be predicted, such as the state of the refining equipment 1 and the results of the refining process performed immediately before. The information on the state of the refining equipment 1 may include, for example, information on the number of times the refining equipment 1 has been used. In addition, the control unit 11 may include information as refining conditions such as the mixing ratio of molten metal and cold iron, information on the pretreatment process of the molten metal, measurement results including the shape of the refining equipment 1, and actual values ​​of the molten metal components, molten metal temperature, and slag components after processing in the refining process performed immediately before at the refining equipment 1 to be predicted.

[0059] The parameters in the smelting process to be predicted may include, for example, the amount of auxiliary materials added and the amount of acid supplied. The control unit 11 may obtain the parameters in the smelting process to be predicted by, for example, an operator inputting data to the input unit 12. The parameters in the smelting process to be predicted may further include the supply rate of high-pressure oxygen (acid supply rate), the flow rate of the agitated gas, the height of the lance 40, and so on.

[0060] The control unit 11, after obtaining the planned values ​​for the refining conditions and manipulated quantities before the refining process, which is the target of the refining result prediction, refers to the past performance data stored in the memory unit 14 and calculates the similarity between the planned values ​​for the refining conditions and manipulated quantities before the refining process and the past actual values ​​for the refining conditions and manipulated quantities in past refining results. A specific example of how the control unit 11 calculates the similarity will be described later.

[0061] Furthermore, the control unit 11 refers to past performance data stored in the memory unit 14 and generates a predictive model that relates past performance values ​​of refining conditions and past performance values ​​of manipulation quantities with past performance values ​​of refining results.

[0062] When the control unit 11 calculates the similarity and generates the prediction model, it determines the parameters of the prediction model by solving an optimization problem using an evaluation function that uses the similarity as a weight to evaluate the prediction error of the prediction model. A specific example of how the control unit 11 determines the parameters will be described later.

[0063] When the control unit 11 determines the parameters of the prediction model, it inputs the planned values ​​of the refining conditions and manipulated quantities before the refining process into the prediction model with the determined parameters, and calculates a refining result that includes multiple predicted values ​​among the predicted values ​​of the molten metal components, the predicted molten metal temperature, and the predicted values ​​of the slag components after the refining process.

[0064] In this way, by solving an optimization problem for an evaluation function weighted by similarity and determining the parameters of the prediction model, the control unit 11 can generate a prediction model capable of predicting the refining results with high accuracy. Furthermore, because weighting based on similarity increases the contribution of past performance close to the required point (i.e., the set of planned values ​​for refining conditions and manipulative quantities) relatively, and suppresses the contribution of past performance that deviates from the required point, a prediction model with high stability of predicted values ​​near the required point can be obtained. Moreover, using the generated prediction model, the control unit 11 can predict the refining results, including multiple predicted values ​​among the predicted values ​​of molten metal components, molten metal temperature, and slag components after the refining process, with high accuracy. In addition, when predicting multiple items (molten metal components, molten metal temperature, and slag components) simultaneously, the contribution of past performance close to the required point can be reflected with a unified weighting for each item, thus ensuring consistency of each predicted value for the same operating conditions.

[0065] (Calculation of Similarity and Determination of Parameter Values) Next, an example of the calculation of similarity and determination of the parameters of the prediction model by the control unit 11 will be explained in detail with specific examples.

[0066] The control unit 11 starts the process of predicting the refining result when it is instructed to execute the refining result prediction process by an operator's operation to the input unit 12. Alternatively, the control unit 11 starts the process of predicting the refining result each time it obtains the planned values ​​of the refining conditions and manipulated quantities before the refining process.

[0067] When the control unit 11 acquires the scheduled values of the refining conditions and operation amounts before refining for the refining process that is an object for which a refining result is to be predicted, the control unit 11 refers to past performance data stored in the storage unit 14 to calculate the similarity between the scheduled values of the refining conditions and operation amounts before refining and the actual values of past refining conditions and past operation amounts in past refining performance. At this time, the control unit 11 calculates the similarity for each piece of past performance data. In the following description, the scheduled values of refining conditions and operation amounts used before the refining process are treated as a set of each item (set of scheduled values), and the actual values of refining conditions and operation amounts in past refining performance are treated as a set of each item (set of actual values). Here, each item includes items related to refining conditions and operation amounts such as temperature, concentration, flow rate, and position. In this specification, the term "request point x" refers to a representation of the refining conditions and operation amounts (set of scheduled values) related to the refining process to be predicted as a point on an input variable space.

[0068] Specifically, for example, the control unit 11 first sets a point in the input variable space corresponding to the refining process conditions and operation amounts as the request point x (≡ [x １ , x ２ , …, x Ｍ Ｔ ), and for the actual value x ｎ of each input variable stored in the storage unit 14, calculates the distance L ｎ from the request point x using mathematical formula (1) shown below. The distance used here is positioned as an index indicating the degree of proximity between the set of scheduled values and the set of actual values. The distance is a comprehensive index based on the difference between the scheduled value and the actual value for each item, and a smaller distance means a higher proximity between the scheduled value and the actual value. If necessary, the magnitude of the distance may be determined after standardizing scale differences (unit differences and digit differences) for the values of each item. However, in mathematical formula (1), λ ｍ is a weighting factor for scaling input variables measured on different scales such as chemical components and temperature.

[0069] Subsequently, for the actual value x ｎ of each input variable stored in the storage unit 14, the control unit 11 calculates the distance L from the request point x using mathematical formula (2) shown below ｎ ​The similarity W of two points ｎ The similarity is calculated. In this embodiment, the similarity is an index whose value is determined according to distance, with a higher similarity value set for actual data at shorter distances and a lower similarity value set for actual data at longer distances. This allows for greater emphasis on actual data under conditions close to the planned values, and relatively reduces the contribution of actual data under conditions that deviate from the planned values. The similarity is used as a weight in the evaluation function described later. However, in equation (2), p is an adjustment parameter, and σ Ｌ This is the distance L in equation (1). ｎ This is the standard deviation.

[0070] Furthermore, the control unit 11 uses the following formula (3) instead of the above formula (2) to move from the required point x to a distance L ｎ The similarity W of two points ｎ You may calculate this. However, in equation (3), α is the forgetting element, and is a value greater than 0 and less than 1. If we substitute the forgetting element α into equation (2) to get equation (3), we get the new actual value x ｎ The similarity increases, and the old performance value x ｎ The similarity will decrease. By setting the similarity to include the forgetting element α, it is possible to update the neighbor parameters with an emphasis on the new data even when new performance data is added.

[0071] Note that the distance L in the above formula (1) ｎ This is a weighted distance based on the difference between each item, but this is just one example. The control unit 11 may use distance indicators between sets of values ​​for each item, such as Euclidean distance, Cityblock distance, Minkowski distance, or Mahalanobis distance. Furthermore, it may be set to use a similarity indicator, such as cosine similarity, as a substitute for the distance between k-dimensional vectors. A short distance (or high similarity) between the calculated k-dimensional vectors means that the planned value and the actual value are close. In addition, when calculating similarity, it may be set to refer only to actual data whose distance is below a predetermined first threshold, and to treat the contribution of actual data whose distance exceeds the first threshold as zero or reduced.

[0072] When the control unit 11 calculates the similarity, it uses the N actual data (actual value x of the input variable) stored in the memory unit 14. ｎ ) and the similarity W between it and the required point x ｎ Using this, a local prediction model is generated that emphasizes past performance data similar to the required point x.

[0073] Specifically, for example, the control unit 11 may generate a predictive model represented by the following formula (4). The following formula (5) represents the parameters of formula (4). The parameter θ in formula (5) is the similarity W represented by the following formulas (6) to (9). ｎ This can be calculated by solving an optimization problem that minimizes the value of the evaluation function J, which is the sum of squared errors between the measured value and the predicted value, weighted by .

[0074]

[0075] However, in formula (7), y ｎ (n = 1, 2, ..., N) are the values ​​of the output variable corresponding to the nth actual data point. Also, in equation (8), diag(s) represents a diagonal matrix whose primary diagonal elements are the elements of s.

[0076] The control unit 11 determines the parameters of the prediction model by calculating the model parameters that minimize the weighted sum of squares of the predicted and actual values. This allows the control unit 11 to generate a local prediction model that better fits the actual data with high similarity, i.e., data close to the required point x.

[0077] If there are multiple items to predict as refining results, the control unit 11 may calculate the above formulas (4) to (9) as many times as there are items to predict and generate prediction models for each item. For example, when predicting three items—molten metal components after refining, molten metal temperature after refining, and slag components after refining—the control unit 11 may generate a prediction model for predicting the molten metal components after refining, a prediction model for predicting the molten metal temperature after refining, and a prediction model for predicting the slag components after refining.

[0078] Furthermore, when the control unit 11 generates a prediction model for one prediction item y, it may use the predicted or target values ​​of other prediction items as input variables x.

[0079] For example, when the control unit 11 generates multiple prediction models for predicting refining results, with at least two of the following as target variables: molten metal components, molten metal temperature, and slag components, it may generate the prediction model in such a way that it calculates the predicted value of the target variable of the prediction model by inputting the predicted or target value of the target variable other than the target variable of the prediction model, in addition to the planned values ​​of the refining conditions and manipulated quantities before the refining process.

[0080] More specifically, for example, when the control unit 11 generates a prediction model with molten metal components as the objective variable and a prediction model with molten metal temperature as the objective variable, when generating the prediction model with molten metal components as the objective variable, it may add a predicted or target value of the molten metal temperature to the input variable x, in addition to the planned values ​​of the refining conditions and manipulated amounts before the refining process.

[0081] In this case, when the control unit 11 calculates the similarity W, it uses the past molten metal temperature data x ｎ The target value x of the molten metal temperature to be predicted is used.

[0082] Furthermore, for example, when the control unit 11 generates a prediction model with molten metal components as the objective variable and a prediction model with molten metal temperature as the objective variable, when generating the prediction model with molten metal temperature as the objective variable, the predicted or target value of the molten metal components may be added to the input variable x in addition to the planned values ​​of the refining conditions and manipulated amounts before the refining process.

[0083] In this case, when the control unit 11 calculates the similarity W, it uses the past performance molten metal concentration x ｎ The target value x of the molten metal concentration to be predicted is used.

[0084] Furthermore, if the difference between the target molten metal temperature used to predict the molten metal components and the predicted molten metal temperature calculated using the predicted molten metal components is greater than or equal to a threshold, a prediction model may be generated to predict the molten metal components again, with the predicted molten metal temperature being x. In addition, the process of calculating the predicted molten metal temperature using the prediction model generated with the newly calculated predicted molten metal components being x, and then calculating the predicted molten metal components using the prediction model generated with the newly calculated predicted molten metal temperature being x, may be repeated until the change in the predicted values ​​of multiple prediction items falls below a threshold.

[0085] Figure 4 shows a specific flowchart. Here, F Ｔ This is a predictive model with molten metal temperature as the target variable, F Ｍ This is a predictive model with molten metal components as the target variable, F Ｓ This represents a predictive model with the slug component as the dependent variable. Ｍ and F Ｓ In this case, a different prediction model is generated for each type of component. T represents the value used to predict the molten metal temperature, M represents the molten metal component (subscript i represents the type of component), and S represents the value used to predict the slag component (subscript j represents the type of component). T' represents the predicted value calculated by the respective prediction models for molten metal temperature, M' for molten metal component, and S' for slag component. In the flowchart shown in Figure 4, multiple prediction models F have molten metal component, molten metal temperature, and slag component as the target variables. Ｔ F Ｍ F Ｓ The predicted values ​​for each dependent variable are repeatedly calculated using multiple prediction models until the difference or rate of change between the input value used to predict the dependent variable of one prediction model and the predicted value of another prediction model using the same input value as the dependent variable falls below a threshold. This improves the prediction accuracy of molten metal temperature, molten metal components, and slag components.

[0086] As described above, when there are multiple prediction items, by repeating the process of calculating similarity and generating a prediction model, it becomes possible to use the predicted values ​​of other prediction items as input variables x when generating a prediction model for one prediction item y. In other words, by repeatedly updating while mutually referencing the predicted or target values ​​of each target variable until the difference or rate of change falls below a predetermined threshold, the predicted values ​​of each target variable are brought closer to mutually consistent values, and the value of the evaluation function using those predicted values ​​is reduced. This makes it possible to obtain mutually consistent and highly accurate prediction results for multiple items all at once.

[0087] Furthermore, when solving the optimization problem, the following constraints may be imposed. Specifically, as constraints, the range of the partial regression coefficient φ of the input variable in the model parameters represented by equation (10) may be restricted by the following equations (11) to (13). Here, the lower and upper limits represented by equations (12) and (13) shall provide physical foresight information between the input and output variables.

[0088] Specifically, as the molten metal temperature given as an input variable increases, the P concentration in the molten metal after processing also increases. Therefore, the lower and upper limits of the model parameter corresponding to the molten metal temperature are set to 0 and ∞, respectively. By adding constraints on the information obtained from the physical model, it is possible to better fit actual data close to the required point and generate a local prediction model that has partial regression coefficients that match the physical characteristics of the target to be predicted.

[0089] Furthermore, the control unit 11 may determine whether or not to adopt the generated prediction model, and if it cannot adopt it, it may adopt another prediction model prepared in advance and input the planned values ​​of the refining conditions and manipulated amounts before the refining process into the other prediction model prepared in advance, thereby calculating a refining result that includes multiple predicted values ​​among the predicted values ​​of the molten metal components, the predicted molten metal temperature, and the predicted values ​​of the slag components after the refining process.

[0090] The other models prepared in advance may include, for example, models based on physical reaction models, empirical rule models, machine learning models, or statistical models that do not use calculated similarity scores.

[0091] The control unit 11 may, for example, decide not to adopt the generated prediction model if the number of historical data points (out of N historical data points) whose calculated similarity is less than or equal to a predetermined first threshold is less than or equal to a predetermined number. Alternatively, the control unit 11 may, for example, decide not to adopt the generated prediction model if, as a result of solving the optimization problem to minimize the value of the evaluation function J in the above formula (6), the value of the evaluation function J is greater than or equal to a predetermined second threshold.

[0092] For example, if the number of historical data points whose calculated similarity is below a predetermined first threshold is less than or equal to a predetermined number, or if the value of the evaluation function obtained after solving the optimization problem is above a predetermined second threshold, the prediction accuracy may decrease. However, by predicting the refining results using a pre-prepared alternative model, the control unit 11 can predict the refining results with high accuracy. Furthermore, since the pre-prepared alternative model is automatically selected according to the threshold determination, the control unit 11 can continue generating control signals based on the predicted values ​​without interrupting the updating of the predicted values, and can maintain the control of the acid supply amount, auxiliary material input amount, stirring gas flow rate, and lance height in refining equipment such as converters to a safe level. The above model switching enables the continuous supply of predictions even when unknown conditions and disturbances occur, and functions as a fail-safe to ensure the continuity and safety of operations.

[0093] The operation of the state prediction device 10 according to this embodiment will be explained with reference to the flowchart shown in Figure 3.

[0094] Step S101: Before executing the refining process for which the refining result is to be predicted, the control unit 11 of the state prediction device 10 acquires the refining conditions before executing the refining process to be predicted and the planned values ​​of the manipulated quantities in the refining process to be predicted.

[0095] Step S102: The control unit 11 obtains the planned values ​​of the refining conditions and manipulated quantities before the refining process, which is the target of predicting the refining result. Then, it refers to the past performance data stored in the memory unit 14 and calculates the similarity between the planned values ​​of the refining conditions and manipulated quantities before the refining process and the past actual values ​​of the refining conditions and manipulated quantities in past refining performance.

[0096] Step S103: The control unit 11 refers to the past performance data stored in the memory unit 14 and generates a predictive model that relates the past performance values ​​of refining conditions and past performance values ​​of manipulation quantities with the past performance values ​​of refining results.

[0097] Step S104: After calculating the similarity and generating the prediction model, the control unit 11 determines the parameters of the prediction model by solving an optimization problem using an evaluation function that uses the similarity as a weight to evaluate the prediction error of the prediction model.

[0098] Step S105: Once the control unit 11 determines the parameters of the prediction model, it inputs the planned values ​​of the refining conditions and manipulated quantities before the refining process into the prediction model with the determined parameters, thereby calculating a refining result that includes multiple predicted values ​​among the predicted values ​​of the molten metal components, the predicted molten metal temperature, and the predicted values ​​of the slag components after the refining process.

[0099] Step S106: The control unit 11 transmits the calculated refining result to the control terminal 20 via the communication unit 15. Based on the refining result, the control terminal 20 controls various operations in the refining process so that the molten metal components and the molten metal temperature of the molten metal 101 after the refining process are within a desired range, thereby producing molten steel.

[0100] The control terminal 20 receives the refining results (predicted values ​​of molten metal components, predicted molten metal temperature, and predicted values ​​of slag components (FeO, etc.)) transmitted from the control unit 11, and refers to the desired operational range for each item from the storage unit within the control terminal 20 or an external storage device.

[0101] The control terminal 20 calculates the set values ​​for the manipulated quantities based on the refining results received from the control unit 11 (predicted values ​​of molten metal components, predicted molten metal temperature, and predicted values ​​of slag components (FeO, etc.)) so that each item falls within the desired range. The manipulated quantities may include, for example, the amount of acid supplied, the acid supply rate, the stirring gas flow rate, the lance height, and the amount of auxiliary materials added. Specifically, the control terminal 20 derives the direction of the type and required amount of input materials according to the operating conditions such as molten metal temperature, components, and slag amount through thermodynamic equilibrium calculations and physical calculations based on heat balance, and determines the initial settings for auxiliary material input amounts, etc., based on the results. Furthermore, the control terminal 20 refers to coefficient tables (for example, the amount of temperature reduction per 1 kg of coolant used for temperature adjustment (cooling coefficient), the coefficient of concentration reduction relative to the amount of lime used for component adjustment (dephosphorization coefficient, etc.)) stored in the memory area of ​​the control terminal 20 or in the memory unit 14 for each condition (temperature, amount of slag, slag composition, etc.) and calculates the required amount of operation according to the deviation between the predicted value and the target value.

[0102] On the other hand, regarding the adjustment of slag components (FeO, etc.), the blowing pattern may be determined by referring to a table of operational rules related to the blowing pattern (acid supply rate, stirring gas flow rate, lance height) stored in the memory area of ​​the control terminal 20. For example, if there is a lot of FeO, the setting values ​​may be changed in the direction of lowering the lance height or increasing the acid supply rate, and if there is little FeO, the setting values ​​may be changed in the direction of raising the lance height or lowering the acid supply rate, thus determining the direction and amount of change from the current setting values.

[0103] The calculated set values ​​are used in subsequent fine-tuning processes, referencing the predicted values ​​for molten metal temperature, molten metal composition, and slag composition. In semi-automatic mode, the fine-tuning process and subsequent control signal generation may be performed by the control terminal 20 in response to operator confirmation via an output interface (display, etc.) displaying operation guidance. In automatic mode, the control terminal 20 may perform these processes automatically based on the set values.

[0104] The control terminal 20 uses the set values ​​as initial values, cross-references the predicted values ​​of the molten metal temperature, molten metal components, and slag components, and repeatedly fine-tunes the set value group (acid supply amount, acid supply rate, stirring gas flow rate, lance height, auxiliary material input amount) until the difference or rate of change between each predicted item falls below a predetermined threshold. In semi-automatic mode, the control terminal 20 may display the fine-tuning results as operation guidance and proceed to the next stage upon operator approval. In automatic mode, the control terminal 20 may automatically perform the fine-tuning. When the termination conditions are met, the control terminal 20 terminates the calculation and confirms the set value group in which the predicted values ​​for the same operating conditions are consistent.

[0105] After the set of values ​​is determined, the control terminal 20 generates control signals for the acid supply amount, acid supply rate, stirring gas flow rate, lance height, and auxiliary material input amount based on the set of values. In semi-automatic mode, the control terminal 20 may display the contents of the control signals to notify the operator and execute actuator control according to the approved operation. In automatic mode, the control terminal 20 may directly control the actuator based on the control signals.

[0106] (Dimensionality Reduction) Before executing the process of calculating similarity, the control unit 11 may perform a linear transformation and dimensionality reduction on the refining conditions and manipulated quantities using principal component analysis. By performing such dimensionality reduction, robustness against noise and reduction of computational resources can be achieved simultaneously, and the convergence of similarity calculation and parameter optimization can be improved. Therefore, dimensionality reduction can contribute to shortening the response time of online prediction. The control unit 11 can predict the refining results with higher accuracy. Dimensionality reduction will be described below.

[0107] Specifically, the control unit 11 sets the actual values ​​of the refining conditions and the manipulated amount, which are input variables, to x ｎ (=[x １ ｎ , x ２ ｎ , ..., x Ｌ ｎ ] Ｔ) (where n = 1, 2, ..., N, L are the number of input variables) First, use the formula (14) shown below to set each actual value x such that the mean is 0 and the standard deviation is 1. ｎ Standardize it.

[0108] Note that in formula (14), x Ｌ，ａｖ is the actual value x ｎ This is the average value, and the denominator represents the standard deviation. Actual values ​​of molten metal state and blowing conditions after standardization x ｎ to z ｎ (=[z １ ｎ , z ２ ｎ , ..., z Ｌ ｎ ] Ｔ ) or as shown in the following formula (15).

[0109]

[0110] Next, the control unit 11 calculates a covariance matrix V defined by the following formula (16), and calculates the eigenvalues ​​of this covariance matrix V and their corresponding eigenvectors. The covariance matrix V has multiple non-negative eigenvalues ​​and multiple corresponding eigenvectors. Therefore, the eigenvectors are sorted in descending order of their corresponding eigenvalues, and M eigenvectors are selected in descending order of their corresponding eigenvalues ​​to form a matrix P (= [w １ ,w ２ ... lol Ｍ ] Ｔ This is expressed as follows: where M is a natural number smaller than the number of input variables L, and matrix P is a matrix called the loading matrix. The control unit 11 stores in the storage unit 14 the result of linearly transforming the refining conditions and manipulated quantity z using the loading matrix P as shown in the following formula (17).

[0111]

[0112] Furthermore, the control unit 11 determines the molten metal state and blowing conditions x (= [x １ , x ２ , ..., x Ｌ ] ＴSimilarly, for each element of ), first, standardize using the formula (18) shown below, and then predict the molten metal state and blowing conditions z (= [z １ , z ２ , ..., z Ｌ ] Ｔ The control unit 11 then calculates the following. The control unit 11 uses the loading matrix P to perform a linear transformation as shown in equation (19) below and uses the result as the required point.

[0113]

[0114] As described above, in the state prediction device 10 according to this embodiment, the control unit 11 acquires the refining conditions before the refining process, including the measured values ​​of the molten metal components and molten metal temperature before the refining process to be predicted, and the planned values ​​of the operating quantities, including the amount of auxiliary materials added and the amount of acid supplied in the refining process to be predicted. The control unit 11 calculates the similarity between the planned values ​​of the refining conditions and operating quantities before the refining process and the actual values ​​of the refining conditions and operating quantities in past refining results. The control unit 11 generates a prediction model that relates the actual values ​​of the refining conditions and operating quantities in the past to the actual values ​​of the refining results in the past. The control unit 11 determines the parameters of the prediction model by solving an optimization problem using an evaluation function that uses the calculated similarity as a weight to evaluate the prediction error of the prediction model. The control unit 11 then inputs the planned values ​​of the refining conditions and operating quantities before the refining process into the prediction model with determined parameters, and calculates a refining result that includes multiple predicted values ​​among the predicted values ​​of the molten metal components, the predicted value of the molten metal temperature, and the predicted value of the slag components after the refining process. As described above, the state prediction device 10 according to this embodiment determines the parameters of the prediction model based on the similarity between the planned values ​​of the refining conditions and manipulated quantities before the refining process and the actual values ​​of the past refining conditions and manipulated quantities in past refining results, and calculates a refining result that includes multiple predicted values. Therefore, it can predict the refining results of multiple targets with high accuracy.

[0115] <Examples> Below, we will describe examples of evaluating the prediction accuracy of the state prediction (for example, prediction of the post-treatment P concentration in a converter) based on comparative examples and examples. Figure 5 shows the relationship (for example, a scatter plot) between the actual post-treatment P concentration and the predicted post-treatment P concentration for Comparative Example 1, Example 1, and Example 2.

[0116] (Evaluation Conditions) The evaluation targets are 200 charges in the operation of de-P and de-C treatment by converter refining. For each charge, a predictive model was created based on the most recent 1130 past charge operation data, and the post-treatment P concentration was predicted. As input variables, Comparative Example 1, Example 1, and Example 2 all used molten metal components, molten metal temperature, scrap input amount, and auxiliary material input amount. Example 2 also used the predicted endpoint temperature. Here, the molten metal components include measured values ​​of C, Si, and P [%] before converter blowing. The molten metal temperature is the measured value [°C] before converter blowing. The scrap input amount is the amount of scrap [ton] used as an iron source. The auxiliary material input amount includes the input amounts of solvents (lime, dolomite, etc.), heat risers (C source, Si source), and coolants (iron ore), etc.

[0117] (Evaluation Indicators) For each charge, the difference between the predicted value of the P concentration after treatment and the actual value of the P concentration after treatment (prediction error) was calculated, and the mean and standard deviation of the prediction error were calculated. The absolute values ​​of the normalized values ​​(hereinafter referred to as "normalized error mean" and "normalized error standard deviation") were calculated using the mean and standard deviation of the prediction error of Comparative Example 1 as the baseline (1.0 each).

[0118] (Comparative Example 1) In Comparative Example 1, a multiple regression model was used as the prediction model for the molten metal components (P concentration). In this case, the mean normalized error was set to 1.0 and the standard deviation of the normalized error was set to 1.0.

[0119] (Example 1) In Example 1, the distance L, which represents the "closeness" between the charge to be calculated and each of the past performance data, is calculated based on the measured values ​​of the molten metal components and molten metal temperature before the refining process, the planned values ​​of the scrap input amount and auxiliary material input amount (hereinafter referred to as these as input variables), and the past refining performance data of the most recent 1130 charge. ｎ Calculate the distance L ｎ Applying a monotonically decreasing function to the similarity Wｎ The similarity W was calculated. ｎ Using these as weights, the optimization problem was solved to minimize the evaluation function J, which is defined as the weighted sum of squared prediction errors (e.g., the difference between measured and predicted values), so that historical data similar to the charge under calculation contributes more significantly. This determined the model parameters. As a result, a local regression model was generated that provides a good fit for the local molten metal composition (P concentration) after refining in the vicinity of the charge under calculation. In this case, the mean normalized error was 0.53 and the standard deviation of the normalized error was 0.97, representing a 47% improvement in the mean normalized error and a 3% improvement in the standard deviation of the normalized error compared to the comparative example.

[0120] (Example 2) In Example 2, the similarity W between the charge to be calculated and each past performance data was determined based on the measured values ​​of the molten metal components and molten metal temperature before the refining process, the planned values ​​of the scrap input amount and auxiliary material input amount (input variables), and the past refining performance data of the most recent 1130 charge. ｎ The similarity W is calculated and ｎ A local prediction model was generated using these as weights.

[0121] Specifically, a local regression model was generated with the molten metal composition (P concentration) after refining as the dependent variable, and another local regression model was generated with the molten metal temperature after refining as the dependent variable. The predicted value of the molten metal temperature (predicted molten metal temperature) was used as the input variable for the prediction model of the molten metal composition (P concentration) to calculate the molten metal composition (P concentration) after refining.

[0122] In this embodiment 2, the difference between the predicted molten metal temperature value used to predict the molten metal components (P concentration) and the predicted molten metal temperature value calculated using the predicted molten metal components (P concentration) value fell below a predetermined threshold. Therefore, the process of repeatedly updating the predicted values ​​of each target variable using multiple prediction models (convergence calculation) was not performed. In this case, the mean normalized error was 0.47 and the standard deviation of the normalized error was 0.89, which is an improvement of 53% in the mean normalized error and 11% in the standard deviation of the normalized error compared to Comparative Example 1.

[0123] Based on the above, Example 1, which uses a local regression model, can reduce the bias and variability of the prediction error of the post-treatment P concentration compared to Comparative Example 1, which uses a multiple regression model. Furthermore, Example 2, which adds the endpoint temperature prediction value as an input variable, can further reduce the variability of the prediction error compared to Example 1.

[0124] This disclosure is not limited to the embodiments described above. For example, multiple blocks described in the block diagram may be combined, or a single block may be divided. Instead of executing multiple steps described in the flowchart in chronological order as described, they may be executed in parallel or in a different order, depending on the processing capacity of the device performing each step, or as necessary. Other modifications are possible without departing from the spirit of this disclosure.

[0125] For example, although the above embodiment was described with reference to a converter 30, the state prediction device 10 according to this embodiment can also be applied to other equipment such as secondary refining equipment and pre-treatment equipment.

[0126] For example, in the above embodiment, we showed a case where the refining result after the refining process to be predicted is predicted before the refining process to be predicted is started. However, the state prediction device 10 in this embodiment can also predict the refining result during the refining process.

[0127] 1. Refining equipment 10. State prediction device 11. Control unit 12. Input unit 13. Output unit 14. Memory unit 15. Communication unit 20. Control terminal 30. Converter 31. Ventilation hole 40. Lance 50. Duct 60. Flow meter 101. Molten metal 102. Slag

Claims

1. A state prediction device for predicting the refining results in a refining facility, comprising: a control unit; which acquires pre-refining conditions including measured values ​​of molten metal components and molten metal temperature before the refining process to be predicted, and planned values ​​of operational quantities including the amount of auxiliary materials added and the amount of acid supplied during the refining process to be predicted; calculates the similarity between the pre-refining conditions and planned values ​​of operational quantities and the actual values ​​of past refining conditions and operational quantities in past refining results; generates a prediction model relating the actual values ​​of past refining conditions and operational quantities and the actual values ​​of past refining results; determines the parameters of the prediction model by solving an optimization problem using an evaluation function weighted by the similarity as an evaluation function for evaluating the prediction error of the prediction model; and calculates the refining results including a plurality of predicted values ​​among predicted values ​​of molten metal components, predicted values ​​of molten metal temperature and predicted values ​​of slag components after the refining process by inputting the pre-refining conditions and planned values ​​of operational quantities into the prediction model.

2. The state prediction device according to claim 1, wherein the control unit applies a monotonically decreasing function to the distance between the set of planned values ​​for the refining conditions before the refining process and the manipulated amount, and the set of actual values ​​for the past refining conditions and the actual values ​​for the past manipulated amount, to calculate the similarity.

3. The state prediction device according to claim 2, wherein the control unit applies a function that decreases exponentially with respect to the distance as the monotonically decreasing function, and calculates the similarity such that the similarity becomes relatively larger for newer refining performance data due to the forgetting element α.

4. The state prediction device according to claim 2 or 3, wherein the control unit calculates the difference between the planned value and the actual value for each item, and calculates the distance based on the sum of the differences or the sum of the squares of the differences.

5. The state prediction device according to any one of claims 2 to 4, wherein the control unit, prior to calculating the distance, standardizes the scale of each item of the planned value and the actual value, and calculates the distance for the items after linear transformation and dimensionality reduction using the loading matrix P obtained by principal component analysis.

6. The state prediction device according to any one of claims 1 to 5, wherein the control unit generates a plurality of prediction models for predicting the refining result, with at least two of the molten metal components, the molten metal temperature, and the slag components as target variables, and calculates the predicted value of the target variable of the prediction model by inputting the predicted value or target value of the target variable other than the target variable of the prediction model, in addition to the planned values ​​of the refining conditions before the refining process and the manipulated amount into the prediction model.

7. The state prediction device according to claim 6, wherein the control unit repeatedly calculates the predicted value of each target variable using the plurality of prediction models until the difference or rate of change between the input value used to predict the target variable of one of the plurality of prediction models and the predicted value of another prediction model that uses the input value as the target variable falls below a threshold.

8. The state prediction device according to any one of claims 1 to 7, wherein the control unit calculates the refining result by inputting the refining conditions before the refining process and the planned values ​​of the manipulated quantities into a separate prediction model prepared in advance, if the number of past performance data whose similarity is less than or equal to a first threshold is less than or equal to a predetermined number, or if, as a result of solving the optimization problem, the value of the evaluation function is greater than or equal to a second threshold.

9. A refining control system comprising a state prediction device according to any one of claims 1 to 8, and a control terminal for controlling various operational quantities in a refining process, wherein the state prediction device further comprises a communication unit capable of communicating the refining results to the control terminal, and the control terminal calculates a set value for an operational quantity, including at least one of the acid supply amount, acid supply rate, stirring gas flow rate, lance height, and auxiliary material input amount, based on the refining results, such that at least one of the molten metal components, molten metal temperature, or slag components after the refining process is within a desired range, and generates a control signal based on the set value to control the actuator of the refining equipment.

10. A state prediction method for predicting the refining results in a refining facility, comprising: a step of obtaining pre-refining conditions including measured values ​​of molten metal components and molten metal temperature before the refining process to be predicted, and planned values ​​of operational quantities including the amount of auxiliary materials added and the amount of acid supplied in the refining process to be predicted; a step of calculating the similarity between the pre-refining conditions and the planned values ​​of the operational quantities and the actual values ​​of past refining conditions and past operational quantities in past refining results; a step of generating a prediction model that relates the actual values ​​of past refining conditions and past operational quantities and the actual values ​​of past refining results; a step of determining the parameters of the prediction model by solving an optimization problem using an evaluation function that uses the similarity as a weight as an evaluation function to evaluate the prediction error of the prediction model; and a step of calculating the refining result including a plurality of predicted values ​​among predicted values ​​of molten metal components, predicted values ​​of molten metal temperature and predicted values ​​of slag components after the refining process by inputting the pre-refining conditions and the planned values ​​of the operational quantities into the prediction model.

11. A refining control method comprising: calculating a set value for an operating variable, including at least one of the acid supply amount, acid supply rate, stirring gas flow rate, lance height, and auxiliary material input amount, based on the refining result calculated by the state prediction method described in claim 10, such that at least one of the molten metal components, molten metal temperature, or slag components after the refining process is within a desired range; and generating a control signal based on the set value to control the actuator of the refining equipment.

Images