Nitrogen concentration prediction method, nitrogen concentration control method, nitrogen concentration prediction device, and nitrogen concentration control device

Abstract

A nitrogen concentration prediction method according to the present invention includes: a prediction step for calculating a predicted value of the nitrogen concentration in molten steel under an operation condition of a prediction object by using a nitrogen concentration prediction model which is a physical model in which a condition of vacuum degassing treatment is used as an input variable and a predicted value of the nitrogen concentration in the molten steel under the condition is used as an output variable; and a correction step for correcting the predicted value of the nitrogen concentration in the molten steel, which has been calculated in the prediction step, by using a machine learning correction model which is a machine learning model in which a condition of the vacuum degassing treatment is used as an explanatory variable, and a prediction error which is a difference value between an actual value and the predicted value of the nitrogen concentration in the molten steel calculated by using the nitrogen concentration prediction model for the condition is used as an objective variable.

Description

Nitrogen concentration prediction method, nitrogen concentration control method, nitrogen concentration prediction device, and nitrogen concentration control device

[0001] The present invention relates to a nitrogen concentration prediction method, a nitrogen concentration control method, a nitrogen concentration prediction device, and a nitrogen concentration control device.

[0002] Generally, molten steel processed in a converter or the like is transported to a secondary refining process, where its nitrogen concentration is controlled by a vacuum degassing process. Since appropriate control of the nitrogen concentration in molten steel contributes to increased functionality of steel products, control as accurately as possible is required. Against this background, numerous techniques for controlling the nitrogen concentration in molten steel during vacuum degassing have been proposed. Specifically, Patent Document 1 describes a method in which, in the first half of the vacuum degassing process, the nitrogen concentration in molten steel is analyzed while the hydrogen and oxygen concentrations are maintained at or below predetermined values, and in the second half of the vacuum degassing process, processing conditions are determined based on the analyzed nitrogen concentration. In the method described in Patent Document 1, the nitrogen concentration in molten steel is predicted taking into account the nitrogen absorption and denitrification reactions at the surface of the molten steel bath and the bubble interface in a vacuum vessel, the nitrogen absorption reaction due to the intrusion of air through an immersion tube, and mixing due to the reflux of the molten steel.

[0003] Patent No. 5836187

[0004] T. Kuwabara, K. Umezawa, K. Mori, and H. Watanabe: Trans. Iron Steel Inst. Jpn., 28 (1988), 305. 19th Committee of the Japan Society for the Promotion of Science (ed.): Recommended Equilibrium Values ​​for Steelmaking Reactions, Revised and Supplemented: 1984. T. Kitamura, K. Miyamoto, R. Tsujino, S. Mizoguchi, and K. Kato: ISIJ International, 36 (1996), 395. K. Harashima, S. Mizoguchi, H. Kajioka, and M. Sakakura: Tetsu-to-Hagane, 73 (1987), 1559.

[0005] When predicting the nitrogen concentration in molten steel, the method described in Patent Document 1 does not take into consideration the influence of differences in the components and temperature of each molten steel being processed, the influence of auxiliary materials added to the molten steel on the components of the molten steel and the vacuum degassing process, and the influence of disturbances in other processing steps and transportation steps other than the vacuum degassing process. Therefore, with the method described in Patent Document 1, the rate of change of the nitrogen concentration changes due to the above-mentioned influences, which increases the prediction error of the nitrogen concentration, and it may become difficult to accurately control the nitrogen concentration in molten steel.

[0006] The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to provide a nitrogen concentration prediction method and a nitrogen concentration prediction device that can accurately predict the nitrogen concentration in molten steel during vacuum degassing. Another object of the present invention is to provide a nitrogen concentration control method and a nitrogen concentration control device that can accurately control the nitrogen concentration in molten steel during vacuum degassing.

[0007] The nitrogen concentration prediction method according to the present invention is a nitrogen concentration prediction method for predicting the nitrogen concentration in molten steel during vacuum degassing treatment, and includes: a prediction step of calculating a predicted value of the nitrogen concentration in molten steel under operating conditions to be predicted using a nitrogen concentration prediction model, which is a physical model that has vacuum degassing treatment conditions as input variables and a predicted value of the nitrogen concentration in molten steel under those conditions as an output variable; and a correction step of correcting the predicted value of the nitrogen concentration in molten steel calculated in the prediction step using a machine learning correction model, which is a machine learning model that has the vacuum degassing treatment conditions as explanatory variables and a prediction error, which is the difference between the predicted value of the nitrogen concentration in molten steel calculated using the nitrogen concentration prediction model for those conditions, as an objective variable.

[0008] The explanatory variables may include the number of times the vacuum degassing equipment is used.

[0009] The correction step may be performed at the timing when the actual value of the nitrogen concentration in the molten steel is acquired.

[0010] The nitrogen concentration control method according to the present invention includes a step of controlling the conditions of a vacuum degassing treatment in accordance with the nitrogen concentration in molten steel predicted by the nitrogen concentration prediction method according to the present invention.

[0011] This method is preferably applied to a vacuum degassing process in which the reflux gas flow rate per ton of molten steel is in the range of 4 to 20 NL / min·ton and the degree of vacuum in the vacuum chamber is 12 / 76 (atm) or less.

[0012] This method is preferably applied to a vacuum degassing process in which the target value of the nitrogen concentration in the molten steel is in the range of 30 to 100 ppm.

[0013] The nitrogen concentration prediction device according to the present invention is a nitrogen concentration prediction device that predicts the nitrogen concentration in molten steel during vacuum degassing treatment, and includes: a prediction means that calculates a predicted value of the nitrogen concentration in molten steel under operating conditions to be predicted using a nitrogen concentration prediction model, which is a physical model that has vacuum degassing treatment conditions as input variables and a predicted value of the nitrogen concentration in molten steel under those conditions as an output variable; and a correction means that corrects the predicted value of the nitrogen concentration in molten steel calculated by the prediction means using a machine learning correction model, which is a machine learning model that has vacuum degassing treatment conditions as explanatory variables and a prediction error, which is the difference between the predicted value of the nitrogen concentration in molten steel calculated using the nitrogen concentration prediction model for those conditions, as an objective variable.

[0014] The nitrogen concentration control device according to the present invention includes control means for controlling the conditions of the vacuum degassing process in accordance with the nitrogen concentration in the molten steel predicted by the nitrogen concentration prediction device according to the present invention.

[0015] The nitrogen concentration prediction method and nitrogen concentration prediction device according to the present invention can accurately predict the nitrogen concentration in molten steel during vacuum degassing treatment, and the nitrogen concentration control method and nitrogen concentration control device according to the present invention can accurately control the nitrogen concentration in molten steel during vacuum degassing treatment.

[0016] Fig. 1 is a schematic diagram showing the configuration of a vacuum degassing apparatus according to one embodiment of the present invention. Fig. 2 is a diagram showing an image of nitrogen concentration prediction by combining a nitrogen concentration prediction model and a machine learning correction model. Fig. 3 is a flowchart showing the flow of nitrogen concentration prediction processing according to one embodiment of the present invention.

[0017] Hereinafter, a nitrogen concentration prediction method, a nitrogen concentration control method, a nitrogen concentration prediction device, and a nitrogen concentration control device according to an embodiment of the present invention will be described in detail with reference to the drawings.

[0018] [Vacuum Degassing Apparatus] First, with reference to FIG. 1, the configuration of a vacuum degassing apparatus according to one embodiment of the present invention will be described.

[0019] Fig. 1 is a schematic diagram showing the configuration of a vacuum degassing apparatus according to one embodiment of the present invention. As shown in Fig. 1, the vacuum degassing apparatus 1 according to one embodiment of the present invention includes a ladle 2 for accommodating molten steel S and a vacuum vessel 3 whose interior can be controlled to a vacuum state. Two immersion pipes 4 are provided at the bottom of the vacuum vessel 3 and communicate with the interior of the vacuum vessel 3. One of the immersion pipes 4 is formed with an inlet 5 for injecting a reflux gas G into the vacuum vessel 3. An exhaust port 6 is formed at the top of the vacuum vessel 3 for discharging gas generated by the vacuum degassing process to the outside.

[0020] When performing vacuum degassing using such a vacuum degassing apparatus, first, the lower ends of the two immersion tubes 4 are immersed in the molten steel S contained in the ladle 2. Next, the vacuum vessel 3 is depressurized while a reflux gas G, such as argon gas or nitrogen gas, is blown in through the blowing port 5, thereby circulating the molten steel S between the ladle 2 and the vacuum vessel 3, thereby performing the degassing treatment. During this process, the molten steel S in the vacuum vessel 3 is exposed to a reduced-pressure atmosphere, and degassing reactions actively occur, primarily resulting in decarburization, denitrification, dehydrogenation, and the like. The gases generated by the degassing reactions are discharged to the outside through the exhaust port 6. Additionally, alloys, auxiliary materials, and the like are added to the molten steel S to adjust the composition of the molten steel S.

[0021] When it is required to control the nitrogen concentration in the molten steel S, the nitrogen concentration in the molten steel S is controlled by changing the operating conditions, such as switching the type of reflux gas G, adjusting the flow rate of the reflux gas G, and adjusting the degree of vacuum in the vacuum vessel 3. For example, using nitrogen gas as the reflux gas G increases the nitrogen absorption rate. Also, increasing the degree of vacuum in the vacuum vessel 3 and lowering the pressure in the vacuum vessel 3 increases the denitrification rate in the vacuum vessel 3. The nitrogen concentration in the molten steel S can be controlled by such operations.

[0022] The control of the operating conditions of the vacuum degassing apparatus 1 is performed by a control device 10 configured with an information processing device such as a computer. The control device 10 controls the operating conditions of the vacuum degassing apparatus 1 based on information indicating the operating state of the vacuum degassing apparatus 1 obtained from sensors and the like provided in the vacuum degassing apparatus 1. In the present embodiment, the control device 10 functions as a nitrogen concentration prediction unit 11 and a correction unit 12 by an arithmetic processing unit within the information processing device executing a computer program. The nitrogen concentration prediction unit 11 predicts the nitrogen concentration in the molten steel S using a nitrogen concentration prediction model, and controls the operating conditions of the vacuum degassing apparatus 1 based on the predicted nitrogen concentration, thereby controlling the nitrogen concentration in the molten steel S. Here, the nitrogen concentration prediction model is a physical model that uses the operating conditions of the vacuum degassing apparatus 1 as input variables and the predicted value of the nitrogen concentration in the molten steel S as an output variable. The correction unit 12 corrects the predicted value of the nitrogen concentration calculated by the nitrogen concentration prediction unit 11 using a machine learning correction model. Here, the machine learning correction model is a machine learning model that uses the operating conditions of the vacuum degassing apparatus 1 as explanatory variables and the prediction error of the nitrogen concentration, which is the difference between the predicted value of the nitrogen concentration calculated by the nitrogen concentration prediction unit 11 using those operating conditions and the actual value, as a response variable. Details of the nitrogen concentration prediction model and the machine learning correction model will be described later.

[0023] [Nitrogen Concentration Prediction Model] Next, the nitrogen concentration prediction model will be described.

[0024] Rate of change d [N] of nitrogen concentration in molten steel S in ladle 2 and vacuum vessel 3 Ｌ / dt, d[N] Ｖ Assuming complete mixing of the molten steel S, / dt can be expressed by the following formulas (1) and (2): In formulas (1) and (2), t is time (sec) [N] Ｖ is the nitrogen concentration (mass%) in the molten steel S in the vacuum vessel 3, V Ｖ is the volume of molten steel S in the vacuum vessel 3 (m ３ ), Q is the amount of molten steel S circulating (m ３ / sec), [N] Ｌ is the nitrogen concentration (mass%) in the molten steel S in the ladle 2, V Ｌ is the volume of molten steel S in the ladle 2 (m ３) is shown.

[0025]

[0026]

[0027] In Equation (1), the terms on the right side, starting from the first term, represent the mixing of molten steel S between the vacuum vessel 3 and the ladle 2 due to reflux, the rate of change in nitrogen concentration at the surface of the molten steel bath in the vacuum vessel 3, and the rate of change in nitrogen concentration at the bubble interface of the reflux gas. Equation (2) also takes into account the mixing of molten steel S between the vacuum vessel 3 and the ladle 2 due to reflux.

[0028] The amount of molten steel S returned, Q, in the formulas (1) and (2) can be calculated using the formula (3) shown below, with reference to the description in Non-Patent Document 1. In the formula (3), Q ｇ is the reflux gas flow rate (NM ３ / min), D is the diameter of the immersion tube 4 (m), P Ｖ indicates the pressure (atm) inside the vacuum chamber 3.

[0029]

[0030] In this embodiment, the rate of change in the nitrogen concentration at each reaction site of the nitrogen absorption and denitrification reactions is assumed to be controlled by a mixture of mass transfer on the molten steel side and chemical reaction at the gas-liquid interface, as shown in the following formula (4). By rearranging formula (4), the following formula (5) is obtained, which shows the reaction rate of nitrogen in the molten steel S in the vacuum vessel 3. Therefore, the nitrogen concentration prediction model calculates the predicted value of the nitrogen concentration in the molten steel S by inputting the necessary values ​​into formula (5) shown below. In formulas (4) and (5), A is the gas-liquid reaction interface area (m ２ ), k ｍ is the mass transfer coefficient of nitrogen on the molten steel side (m / sec), k ｒ is the chemical reaction rate constant of nitrogen (m / sec mass%), [N] ｉ is the nitrogen concentration (mass%) on the molten steel side of the gas-liquid interface, [N] ｅｑ indicates the nitrogen concentration in molten steel in equilibrium with nitrogen gas in the gas phase (equilibrium nitrogen concentration) (mass %).

[0031]

[0032]

[0033] The reaction sites for calculating the rate of change of the nitrogen concentration using equation (5) are the bath surface and the bubble interface in the vacuum chamber 3. The difference between these is the gas-liquid reaction interface area A and the equilibrium nitrogen concentration [N]. ｅｑ Among these, the equilibrium nitrogen concentration [N] ｅｑ can be calculated using the following formula (6) with reference to the description in Non-Patent Document 2. In formula (6), f Ｎ is the activity coefficient of nitrogen, P Ｎ２ is the nitrogen partial pressure, R is the gas constant (J / K·mol), and T is the temperature of the molten steel S (K).

[0034]

[0035] The nitrogen activity coefficient f in formula (6) Ｎ is calculated taking into consideration the influence of the main components C, Si, Mn, and O. Specifically, the activity coefficient f of nitrogen is calculated by the following formula (7) using the values ​​shown in Table 1 below as the first-order interaction coefficients between these components and nitrogen described in Non-Patent Document 2. Ｎ In the formula (7), e Ｎ ｊ , and the coefficient of first-order interaction with nitrogen (j=C, Si, Mn, O). However, the above-mentioned components are only examples, and it goes without saying that the number of components to be considered will increase or decrease depending on the component system of the steel type to be processed.

[0036]

[0037]

[0038] In addition, the nitrogen mass transfer coefficient k ｍ can be calculated using the following formulas (8) to (12) with reference to the description in Non-Patent Document 3. In formulas (8) to (12), the molten steel side mass transfer coefficient k of nitrogen is calculated by taking into account the time it takes for the molten steel S around the bubbles to be replaced. ｍ In the formulas (8) to (12), θ is the contact time (min) between the molten steel S and the bubbles, D Ｎ is the diffusion coefficient of nitrogen (m ２ / min), d Ｂis the bubble diameter (m), U Ｂ is the bubble rising speed (m / min), U Ｌ is the rising speed of the molten steel (m / min), g is the gravitational acceleration (m ２ / s), and L indicates the bath surface height (m) in the vacuum chamber 3 from the position where the reflux gas is blown in.

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] In addition, the nitrogen chemical reaction rate constant k ｒ can be calculated using the following formula (13) with reference to the description in Non-Patent Document 4. In formula (13), [O] represents the oxygen concentration (mass%) in the molten steel S, and [S] represents the sulfur concentration (mass%) in the molten steel S.

[0045]

[0046] Regarding the gas-liquid reaction interface area A in Equation (5), when considering the reaction rate at the surface of the molten steel bath and the reaction rate at the bubble interface, the molten steel bath surface area and the bubble interface area become the gas-liquid reaction interface area, respectively. Here, the molten steel bath surface area is proportional to 2 / 3 of the reflux gas flow rate Q, and the bubble interface area is proportional to the reflux gas flow rate Q. ｇ If the values ​​are proportional to the 2 / 3 power of α, they can be estimated by the following equations (14) and (15). Ｓ ,α Ｂ As another method for estimating the gas-liquid reaction interface area, a calculation method that takes into account the physical behavior of the bubble or molten steel bath surface area using the reflux gas flow rate, reflux gas species, degree of vacuum, etc. can also be considered.

[0047]

[0048]

[0049] On the other hand, the reaction that occurs at the interface of the nitrogen bubbles when nitrogen gas is injected will be either a nitriding reaction or a denitriding reaction, depending on the magnitude of the nitrogen concentration in equilibrium between the nitrogen concentration in the molten steel S and the nitrogen partial pressure in the nitrogen bubbles. Furthermore, as the injected nitrogen bubbles rise in the molten steel, the static pressure from the molten steel changes, causing the nitrogen partial pressure in the nitrogen bubbles to decrease. As a result, the reaction speed and direction of the rising nitrogen bubbles are not constant, and the nitrogen concentration prediction model takes this into account and calculates the rate of change in the nitrogen concentration at the interface of the nitrogen bubbles using the following equation (16). [N] in equation (16) ｅｑ (x) is the equilibrium nitrogen concentration (mass %) in the nitrogen bubbles at the floating distance x from the position where the reflux gas is injected, and can be calculated using the above-mentioned formula (6).

[0050]

[0051] [Machine Learning Correction Model] Next, the machine learning correction model will be described.

[0052] The reason for using a machine learning model to correct the predicted nitrogen concentration is to take into account differences in operating conditions and disturbance effects that are difficult to account for using physical model predictions alone, which are the cause of low nitrogen concentration prediction accuracy. When examining the processing results of vacuum degassing processes, there are cases where the rate of change of nitrogen concentration varies even with roughly the same reflux gas pattern, vacuum level, and nitrogen concentration. In other words, there are cases where the change in the rate of change of nitrogen concentration cannot be explained by simple vacuum degassing processing conditions alone, and in such cases, the prediction accuracy of nitrogen concentration calculations using physical models is low. There are several possible reasons why denitrification behavior varies under roughly the same processing conditions, and some of these reasons are listed below.

[0053] The first cause is the influence of molten steel components. Although the nitrogen concentration prediction model described above takes into account the influences of C, Si, P, Mn, O, and S, it only considers the direct influence on changes in nitrogen concentration and does not take into account differences in these components or changes in temperature and slag composition due to changes in components during vacuum degassing.

[0054] The second cause is the influence of added alloys and auxiliary materials. The addition of alloys and auxiliary materials not only changes the composition of the molten steel, but can also cause the intrusion of impurities, change the temperature of the molten steel, affect the flow of the molten steel, and change the degree of vacuum or atmosphere during addition, which in some cases can affect the rate of change in the nitrogen concentration.

[0055] The third cause is influences outside the vacuum degassing process. The processing conditions in the converter, which is a process that precedes the vacuum degassing process, and the time and environment until the molten steel is transported to the vacuum degassing equipment can change the state of the molten steel at the start of processing. Due to processing time and process constraints, there are cases where the molten steel is not analyzed at the start of the vacuum degassing process, and analytical values ​​immediately after tapping from the converter are used as the initial conditions of the molten steel, including the nitrogen concentration. In this case, the molten steel is susceptible to influences outside the vacuum degassing process.

[0056] Although the above factors are thought to cause changes in denitrification behavior, it is difficult to evaluate how these factors affect the denitrification rate, and it is also extremely difficult to separate the effects of multiple factors. In addition, there may be other factors besides the above factors that are difficult to even conceive. Naturally, it is not easy to formulate these effects, reflect them in a physical model, and improve prediction accuracy.

[0057] After careful consideration of countermeasures for these issues, we found that using machine learning with operational conditions as feature values ​​was effective. Machine learning is useful as a means of resolving the above issues because it can simultaneously evaluate the complex effects of input feature values ​​without requiring logic to explain how they affect something.

[0058] The machine learning model created in this invention is a regression prediction model that predicts a dependent variable from explanatory variables called feature quantities. Therefore, when creating a machine learning model, it is necessary to set a dependent variable and select appropriate feature quantities. When examining an effective dependent variable in this invention, we determined that it was the discrepancy in the predicted value of nitrogen concentration obtained by a physical model. In predictive models of physical quantities using machine learning, fitting parameters in a reaction equation are often used as dependent variables. This is used when, as a result of examining a reaction, it is clear that unknown parameters that cannot be derived from existing numerical values ​​exist. In other words, this method can be applied when there is a certain understanding of how the feature quantities used in machine learning contribute to the physical reaction. However, in the aforementioned problem, it is difficult to understand how the effects occur, and there is also the possibility that multiple influencing factors exist and interfere with each other, making this approach unacceptable.

[0059] Therefore, in this embodiment, instead of using the parameters in the physical model as the objective variable, the error between the predicted result of the physical model and the actual value is used as the objective variable. This separates the objective variable from the calculation of the physical model and makes the machine learning model independent of predictions that take into account the reaction theory of the physical model. This makes it possible to apply machine learning to the above-mentioned problem. Furthermore, a secondary benefit of this method is that it can be widely applied regardless of the calculation method of the physical model. Because the calculation results of the physical model are required as learning data for machine learning, accuracy improvement through machine learning can be expected regardless of the accuracy or theory of the physical model itself. This means that machine learning is easy to introduce and can be applied to many processes, not just vacuum degassing processing.

[0060] Another advantage is that the physical model and the machine learning model are largely independent, making it easy to verify accuracy and modify them. In methods where parameters in a physical model are determined by a machine learning model, only one prediction result is output, so if the accuracy is low, it is not necessarily easy to determine whether the accuracy of the physical model or the machine learning model is the problem. With this method, the prediction result from the physical model and the correction result from the machine learning model that corrects it are obtained separately, making it easy to evaluate the accuracy of each model individually. In addition, it is easy to verify and modify each model to improve its prediction accuracy.

[0061] To select features, we investigated operational variables that could affect denitrification behavior and created a machine learning model to evaluate the contribution of each feature to accuracy. Based on this, we selected features that are generally considered to have an impact, such as molten steel composition, reflux gas pattern, vacuum level, alloy and auxiliary material addition amounts, and transport time from the previous process. Other than these, we found that the number of times equipment was used, such as the number of times the ladle was used, the number of times the submerged tube was used, and the number of times the vacuum vessel was used, significantly affected accuracy. While the steel refining process primarily focuses on thermodynamic reactions, the physical behavior of molten steel, particularly the gas-liquid reaction interfacial area, is important in vacuum degassing. Therefore, as the refractory shape changes with the number of times the equipment was used, changes in the molten steel flow during vacuum degassing are thought to affect denitrification behavior. For these reasons, the significant impact of the number of times the equipment was used is a distinctive feature of vacuum degassing and is considered to be important information that significantly affects prediction accuracy.

[0062] The importance of the number of equipment uses as a feature does not depend on the type of objective variable. In other words, when using machine learning to predict and control the denitrification reaction in vacuum degassing processes, selecting the number of equipment uses as a feature is highly effective in most cases. However, the above considerations are considered to be applicable to processes where the reflux gas flow rate per ton of molten steel is 4 to 20 (NL / min·Tton) and the vacuum level in the vacuum chamber is less than 12 / 76 (atm). When simulating the molten steel flow using numerical analysis, it was found that outside this range, the molten steel flow becomes peculiar, making it impossible to effectively consider the influence of differences in refractory shape due to the number of equipment uses. Specifically, when the reflux gas flow rate is extremely low or the vacuum level in the vacuum chamber is low, the amount of reflux molten steel becomes very small. When the reflux gas flow rate is extremely high, the stirring by gas bubbles increases, resulting in significant turbulence on the molten steel bath surface. In these cases, the influence of refractory at a location different from that of molten steel during normal processing becomes significant, and it may not be possible to simply determine the influence by the number of times the equipment is used. Therefore, the accuracy of machine learning deteriorates outside the above-mentioned range. However, by modifying the physical model and optimizing the machine learning model to accommodate such a range of processing conditions, it is highly possible that the method of the present invention will be effective in improving control accuracy even outside the range of numerical conditions shown above.

[0063] There are no particular restrictions on the algorithm used for machine learning, and an appropriate algorithm is selected according to the different conditions of each facility and the characteristics of the amount of training data. In the examples of the present invention, a model was created using gradient boosting based on a decision tree algorithm, but it has been confirmed that similar accuracy can be achieved with many other algorithms, such as random forests and neural networks.

[0064] Next, a method for correcting the nitrogen concentration using the machine learning correction model will be described. In this embodiment, the objective variable of the machine learning correction model is the error [N] between the predicted value and the actual value of the nitrogen concentration prediction model, which is expressed by the following formula (17): Ｅｒｒｏｒ (mass%). In formula (17), [N] Ｒｅｓｕｌｔis the actual value of nitrogen concentration (mass%), [N] Ｐｈｙｓ－Ｃａｌｃ indicates the predicted value (mass %) of the nitrogen concentration using the nitrogen concentration prediction model.

[0065]

[0066] Error [N] Ｅｒｒｏｒ In other words, the machine learning correction model using [N] as the objective variable can predict how the predicted value of the nitrogen concentration prediction model will deviate under given operating conditions. Therefore, as shown in the following formula (18), by using this predicted error, the predicted value of the nitrogen concentration prediction model can be corrected to a value closer to the actual value. In formula (18), [N]Prediction is the predicted value (mass%) of the final nitrogen concentration, and [N]Prediction is the predicted value (mass%) of the final nitrogen concentration. Ｅｒｒｏｒ indicates the prediction error (mass %) obtained from the machine learning correction model.

[0067]

[0068] FIG. 2 shows an image of nitrogen concentration prediction using a combination of a nitrogen concentration prediction model and a machine learning correction model. As shown in FIG. 2, prediction using the nitrogen concentration prediction model progresses from the start of processing, and correction is performed at processing times when correction using the machine learning correction model is possible (in this example, processing times of 10, 20, and 30 minutes). Prediction using the nitrogen concentration prediction model progresses again based on the corrected predicted nitrogen concentration value, and this process is repeated thereafter to predict the nitrogen concentration at the end of processing. A condition for applying the machine learning correction model is that the processing time must be one for which past actual values ​​that serve as learning data are available. In this case, the above correction can be performed only at processing times for which analysis has been performed and actual nitrogen concentration data has been accumulated. For example, in a process in which sampling is performed at 10 and 20 minutes into processing and at the end of processing, and actual value data is available, the predicted values ​​of the nitrogen concentration prediction model at these processing times can be corrected.

[0069] The more frequently machine learning correction is performed, the more likely prediction accuracy is to improve. Since machine learning correction can essentially be performed at the timing of intermediate analysis, the more frequently intermediate analysis is performed, the more likely control accuracy is to be improved by the present invention. Specifically, higher control accuracy can be expected for processes in which machine learning correction is performed at least once per 15 minutes of processing time, or in which machine learning correction is performed at least once between the scheduled processing end time and 15 minutes before that time. This value of 15 minutes was found by taking into account the accuracy of the nitrogen concentration prediction model, the frequency of adding alloys and auxiliary materials, the time required to melt and homogenize the alloys and auxiliary materials, the frequency of sample collection and the time required for analysis, etc.

[0070] When creating a machine learning correction model, it is desirable to have as much training data as possible. Therefore, when there are many past records of nitrogen concentration control for similar steel types, the application of the present invention significantly improves prediction accuracy. For many steel types requiring nitrogen concentration control to which the present invention is applied, the target nitrogen concentration is in the range of 30 to 100 ppm, and improvement in prediction accuracy has been confirmed, particularly within this range. When aiming for a target nitrogen concentration of less than 30 ppm, denitrification treatment will be performed for a long time, although this will depend on the nitrogen concentration value before treatment. Conversely, when the target nitrogen concentration is greater than 100 ppm, denitrification treatment is often performed for a long time.

[0071] In such cases, the nature of the nitrogen concentration control differs from that of the present invention, which switches between denitrification and absorption rate processes with the aim of keeping the nitrogen concentration within a range of upper and lower limits centered around the target nitrogen concentration value. Furthermore, the actual data obtained will be biased toward either denitrification or absorption and lack diversity. Including such extreme data in the training data group may result in a decrease in prediction accuracy. Therefore, to maintain high prediction accuracy, it is desirable to apply this method to steel types with a target nitrogen concentration range of 30 to 100 ppm. However, predictions with higher accuracy than conventional techniques are possible for steel types outside this range. Additionally, it is entirely conceivable to improve prediction accuracy within different target nitrogen concentration ranges by switching the nitrogen concentration prediction model and machine learning correction model used for each target nitrogen concentration.

[0072] [Nitrogen Concentration Prediction Processing] Finally, with reference to FIG. 3, the flow of nitrogen concentration prediction processing according to one embodiment of the present invention will be described.

[0073] Fig. 3 is a flowchart showing the flow of a nitrogen concentration prediction process according to one embodiment of the present invention. The flowchart shown in Fig. 3 starts when initial conditions are input to the control device 10 and a command to execute the nitrogen concentration prediction process is input to the control device 10 (step S1), and the nitrogen concentration prediction process proceeds to step S2. Examples of the initial conditions input to the control device 10 include the amount of molten steel, the initial composition of the molten steel, and the conditions of the vacuum degassing process (degree of vacuum, reflux gas flow rate, reflux gas type, etc.).

[0074] In the process of step S2, the control device 10 determines parameters used in calculations of the nitrogen molten steel side mass transfer coefficient, chemical reaction rate constant, activity coefficient, etc. in the reaction equation included in the nitrogen concentration prediction model, as well as the reflux amount and reaction interface area. This completes the process of step S2, and the nitrogen concentration prediction process proceeds to the processes of steps S3 and S4.

[0075] In the process of step S3, the control device 10 calculates the rate of change of the nitrogen concentration at the bubble interface using the nitrogen concentration prediction model. This completes the process of step S3, and the nitrogen concentration prediction process proceeds to the process of step S5.

[0076] In the process of step S4, the control device 10 calculates the rate of change of the nitrogen concentration at the surface of the molten steel bath in the vacuum vessel using the nitrogen concentration prediction model. This completes the process of step S4, and the nitrogen concentration prediction process proceeds to the process of step S5.

[0077] In the process of step S5, the control device 10 calculates the change amount Δ[N] of the nitrogen concentration in the molten steel in the vacuum vessel and the ladle, using the change rate of the nitrogen concentration at the bubble interface calculated in the process of step S3 and the change rate of the nitrogen concentration at the bath surface in the vacuum vessel calculated in the process of step S4. This completes the process of step S5, and the nitrogen concentration prediction process proceeds to the process of step S6.

[0078] In the process of step S6, the control device 10 calculates the nitrogen concentration [N] of the molten steel in the vacuum vessel and the ladle using the change amount Δ[N] of the nitrogen concentration in the molten steel in the vacuum vessel and the ladle calculated in the process of step S5. Ｖ , [N] Ｌ This completes the process of step S6, and the nitrogen concentration prediction process proceeds to step S7.

[0079] In the process of step S7, the control device 10 determines whether the current time is a time when machine learning correction is possible. If the result of the determination is that the current time is not a time when machine learning correction is possible (step S7: No), the control device 10 returns the nitrogen concentration prediction process to the process of step S2. On the other hand, if the current time is a time when machine learning correction is possible (step S7: Yes), the control device 10 advances the nitrogen concentration prediction process to the process of step S8.

[0080] In the process of step S8, the control device 10 calculates the prediction error of the nitrogen concentration prediction model by inputting the operating conditions of the vacuum degassing process into the machine learning correction model. Then, the control device 10 uses the calculated prediction error to calculate the nitrogen concentration [N] of the molten steel in the vacuum vessel and the ladle calculated in the process of step S6. Ｖ , [N] Ｌ This completes the process of step S8, and the nitrogen concentration prediction process proceeds to step S9.

[0081] In step S9, the control device 10 determines whether the calculated vacuum degassing process time exceeds the set processing time. If the calculated vacuum degassing process time does not exceed the set processing time (step S9: No), the control device 10 increments the vacuum degassing process time by a predetermined time and returns the nitrogen concentration prediction process to step S2. On the other hand, if the calculated vacuum degassing process time exceeds the set processing time (step S9: Yes), the control device 10 terminates the nitrogen concentration prediction process. Thereafter, the operator compares the predicted nitrogen concentration value with the target value and modifies the processing conditions for adjusting the nitrogen concentration based on the comparison results. Examples of processing conditions include the recirculation gas type, recirculation gas flow rate, vacuum degassing process time, and vacuum level.

[0082] In this example, nitrogen concentration control was performed on multiple steel grades. The molten steel composition before vacuum degassing was in the range of [C] < 0.1%, [P] < 0.01%, [S] < 0.01%, [Mn] = 0.02-1.2%, [Si] = 0.1-4.0%, and [Al] < 0.1%. The target nitrogen concentration was 20-120 ppm, the molten steel temperature was 1480-1630°C, and the processed molten steel volume was 180-240 tons. Nitrogen gas or argon gas was used as the reflux gas, the reflux gas flow rate was 1000-3000 (NL / min), the degree of vacuum during stable processing was 12 / 76 (atm) or less, and the reflux processing time was 30 minutes. Sampling was performed multiple times during processing, and the nitrogen concentration was analyzed and recorded as the actual value.

[0083] As a control method, initial conditions were input at the start of processing to determine the processing conditions. Nitrogen concentration prediction was performed based on these processing conditions. If the predicted nitrogen concentration at the end of processing differed from the target value by more than a certain value, the processing conditions were modified to control the nitrogen concentration. When actual values ​​were obtained through nitrogen concentration analysis during processing, the nitrogen concentration was predicted again based on the actual values, and if necessary, the processing conditions were similarly modified. The difference between the predicted value and the target value, used as the basis for modifying the processing conditions, should be appropriately determined based on the behavior of nitrogen concentration changes, which differ depending on the equipment, molten steel composition, and molten steel conditions, and the accuracy of the nitrogen concentration prediction model. In the example, it was set to 2 ppm. The above-described control was performed, and the control accuracy was evaluated from the results. The results of the control accuracy evaluation are shown in Table 2. Approximately 100 charges were evaluated for each example. The evaluation method was the standard deviation σ of the difference between the target value and the actual value at the end of processing, and the percentage of deviations from the target nitrogen concentration range of ±5 ppm.

[0084]

[0085] As shown in Table 2, significant improvements in control accuracy were observed in Examples 1 to 4 compared to the comparative example using conventional technology. Comparing Example 1 and Example 2, Example 2 also had higher accuracy than the comparative example, but the improvement in control accuracy was even more significant when the number of times the equipment was used as a feature was used. Furthermore, Examples 1, 3, and 4 differ in the timing of the analysis and machine learning correction, but a decrease in accuracy was observed when the correction frequency was low, as in Example 3. In Example 4, the number of corrections was the same as in Example 1, but the corrections were biased toward the first half of the processing, with no correction near the end of the processing, resulting in inferior control accuracy. As a result, it was confirmed that high control accuracy could be obtained by ensuring the number of corrections and performing corrections near the end of the processing, as in Example 1.

[0086] Although the present invention has been described above as an embodiment, the present invention is not limited to the descriptions and drawings that form part of the disclosure of the present invention. In other words, other embodiments, examples, and operational techniques that can be made by those skilled in the art based on the present invention are all included in the scope of the present invention.

[0087] According to the present invention, it is possible to provide a nitrogen concentration prediction method and a nitrogen concentration prediction device that can accurately predict the nitrogen concentration in molten steel during vacuum degassing treatment. Also, according to the present invention, it is possible to provide a nitrogen concentration control method and a nitrogen concentration control device that can accurately control the nitrogen concentration in molten steel during vacuum degassing treatment.

[0088] REFERENCE SIGNS LIST 1 Vacuum degassing device 2 Ladle 3 Vacuum tank 4 Immersion tube 5 Blowing port 6 Exhaust port 10 Control device 11 Nitrogen concentration prediction unit 12 Correction unit G Circulation gas S Molten steel

Claims

1. A nitrogen concentration prediction method for predicting the nitrogen concentration in molten steel during vacuum degassing treatment, comprising: a prediction step of calculating a predicted value of the nitrogen concentration in molten steel under operating conditions to be predicted using a nitrogen concentration prediction model, which is a physical model that has vacuum degassing treatment conditions as input variables and a predicted value of the nitrogen concentration in molten steel under those conditions as an output variable; and a correction step of correcting the predicted value of the nitrogen concentration in molten steel calculated in the prediction step using a machine learning correction model, which is a machine learning model that has the vacuum degassing treatment conditions as explanatory variables and a prediction error, which is the difference between the predicted value of the nitrogen concentration in molten steel calculated using the nitrogen concentration prediction model for those conditions, as an objective variable.

2. The nitrogen concentration prediction method according to claim 1, wherein the explanatory variables include the number of times the vacuum degassing equipment is used.

3. The nitrogen concentration prediction method according to claim 1 or 2, wherein the correction step is executed at the timing when an actual value of the nitrogen concentration in the molten steel is acquired.

4. A nitrogen concentration control method comprising a step of controlling the conditions of a vacuum degassing process in accordance with the nitrogen concentration in molten steel predicted by the nitrogen concentration prediction method according to any one of claims 1 to 3.

5. A nitrogen concentration prediction device that predicts the nitrogen concentration in molten steel during vacuum degassing treatment, comprising: a prediction means that calculates a predicted value of the nitrogen concentration in molten steel under operating conditions to be predicted using a nitrogen concentration prediction model that is a physical model whose input variables are vacuum degassing treatment conditions and whose output variables are predicted values ​​of the nitrogen concentration in molten steel under those conditions; and a correction means that corrects the predicted value of the nitrogen concentration in molten steel calculated by the prediction means using a machine learning correction model that is a machine learning model whose output variables are vacuum degassing treatment conditions and whose explanatory variables are prediction errors that are the difference between the predicted value of the nitrogen concentration in molten steel calculated using the nitrogen concentration prediction model for those conditions.

6. A nitrogen concentration control device comprising control means for controlling the conditions of vacuum degassing treatment in accordance with the nitrogen concentration in molten steel predicted by the nitrogen concentration prediction device according to claim 5.

Images