Converter steelmaking endpoint carbon temperature real-time online prediction method and system based on instant learning
Through real-time learning methods and trust domain reflection and multivariate linear regression algorithm, a dynamic prediction model of the carbon temperature at the end point of converter steelmaking is constructed, which solves the problems of low prediction accuracy and parameter update delay in the existing technology, and achieves efficient real-time carbon temperature Online prediction.
Patent Information
- Application Number
- CN202411110990.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-29
AI Technical Summary
The carbon temperature prediction accuracy of the end point of the existing converter steelmaking is not high, and the model parameter update delays, making it difficult to meet the accuracy requirements of the steel industry.
By using the instant learning method, the converter production database is constructed, the data is synchronized in real time, and the decarbonization and heating coefficients are reversely calculated using trust domain reflection and multivariate linear regression algorithms to build a dynamic prediction model to realize real-time online prediction of carbon temperature.
The accuracy of carbon temperature prediction of the end point of converter steelmaking is improved, the limitations of traditional global models are avoided, and real-time updates and efficient matching of model parameters are achieved.
Smart Images

Figure CN119049607A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of iron and steel metallurgy, and specifically relates to a real-time online prediction method and system for converter steelmaking endpoint carbon temperature based on instant learning. Background Art
[0002] At present, large-scale domestic converters mainly use auxiliary gun detection technology to predict the converter endpoint carbon temperature. The auxiliary gun detection technology is to lower the auxiliary gun for sampling and analysis about 3 minutes before the end of blowing to determine whether the current molten steel temperature and composition meet the current steel grade requirements, and to modify the smelting parameters in time. This technology does not require furnace pouring sampling and has high detection accuracy. However, the accuracy of auxiliary gun model prediction is still not ideal, and there is a delay in the update of model parameters. With the increasingly stringent requirements of the steel industry and products, the prediction accuracy of the converter endpoint needs to be improved urgently. Therefore, it is necessary to design a real-time online prediction method for the converter steelmaking endpoint carbon temperature with instant learning to solve the above technical problems. Summary of the invention
[0003] In order to solve the problems existing in the prior art, the main purpose of the present invention is to propose a real-time online prediction method and system for converter steelmaking endpoint carbon temperature based on just-in-time learning.
[0004] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:
[0005] A real-time online prediction method for converter steelmaking endpoint carbon temperature based on instant learning comprises the following steps:
[0006] S1. Build a converter production database to store historical furnace process parameters and test result data sets, and synchronize the current furnace process parameters and test result data sets in real time;
[0007] S2. After the TSC measurement of the converter auxiliary gun of the current heat is completed, the TSC test results of the current heat and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of just-in-time learning (JITL) is used to find the heat with high similarity with the current heat in the historical heats as the sample set for local modeling;
[0008] S3. Based on the converter decarburization mechanism equation and the heating mechanism equation, local modeling is performed using historical furnaces with high similarity as local sample sets. The trust region reflective (TRF) algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The multiple linear regression (MLR) algorithm is used to reversely solve the heating coefficient of the local sample set and use it as the heating coefficient of the current furnace.
[0009] S4, substituting the solved decarburization coefficient and temperature rise coefficient into the converter decarburization mechanism equation and the temperature rise mechanism equation to construct a dynamic prediction model of the carbon content and temperature at the end point of the current furnace;
[0010] S5. Using the TSC test results of the current furnace as the initial value for predicting carbon content and temperature, input the real-time process parameters into the dynamic prediction model, and use the model to calculate and output the predicted values of carbon content and temperature at different times in real time online until the end point of converter steelmaking is reached.
[0011] As a preferred embodiment of the real-time online prediction method of the carbon temperature at the end point of converter steelmaking based on real-time learning described in the present invention, wherein: in the step S1, the converter production database includes a real-time database and a relational database, and the furnace process parameters and test result data sets include: actual molten iron quantity, actual scrap steel quantity, molten iron [C], molten iron [Si], molten iron [Mn], molten iron [P], molten iron [S], molten iron temperature, slag weight, molten steel quantity, actual ore value, actual lime value, actual dolomite value, oxygen supply in the main blowing stage, actual dynamic oxygen supply, actual dynamic pellets, actual dynamic raw dolomite, target [C], target temperature, TSC detection carbon content, TSC detection temperature, TSO detection carbon content, TSO detection temperature. The above data are obtained from the primary database and secondary database of the steel enterprise.
[0012] As a preferred solution of the real-time online prediction method of converter steelmaking endpoint carbon temperature based on just-in-time learning described in the present invention, wherein: in the step S2, the TSC detection results include: molten steel carbon content, molten steel temperature; the main blowing stage process parameters include: slag weight.
[0013] As a preferred solution of the real-time online prediction method of converter steelmaking endpoint carbon temperature based on real-time learning described in the present invention, wherein: in the step S2, the similarity measurement method based on real-time learning is specifically described as follows: the real-time learning algorithm adopts the operation mode of "modeling and predicting at the same time", selects the historical samples most relevant to the sample to be tested through the similarity measurement function, builds a local model online, and then predicts. This solves the problem that the traditional global model is easily disturbed by multiple working conditions and changes in the operating environment.
[0014] As a preferred embodiment of the real-time online prediction method of converter steelmaking endpoint carbon temperature based on instant learning described in the present invention, wherein: in step S3, the converter decarburization mechanism equation is:
[0015]
[0016] Where: ω[C] t is the real-time carbon content, 10 -2wt%; ω[C]0 is the carbon content of the converter when the decarburization rate is 0, 10 - 2 wt%;ω[C] M Measuring carbon content for the secondary gun TSC, 10 -2 wt%;V(O2) X is the actual oxygen blowing amount, m 3 ; V(O2) XM is the oxygen blowing amount when the auxiliary gun starts to measure, m 3 ; B i is the oxygen content in the converter auxiliary material, m 3 / t; R i is the weight of auxiliary materials added after the auxiliary gun is measured, t; α, β are decarburization coefficients; W ST is the weight of molten steel, t.
[0017] As a preferred embodiment of the real-time online prediction method of converter steelmaking endpoint carbon temperature based on instant learning described in the present invention, wherein: in step S3, the converter heating mechanism equation is:
[0018]
[0019] Where: T(t) is the predicted temperature of the molten steel at time t, °C; T M is the temperature of the molten steel at the time of auxiliary gun detection, ℃; γ is the oxygen blowing temperature rise coefficient, ℃ / (m 3 / t); δ is the temperature rise constant, ℃; K i is the cooling capacity of the coolant, ℃ / t.
[0020] As a preferred embodiment of the real-time online prediction method of converter steelmaking endpoint carbon temperature based on instant learning described in the present invention, wherein: in step S3, the multiple linear regression algorithm is specifically described as: the multiple linear regression model studies the relationship between a dependent variable and multiple independent variables. The general multiple linear regression formula is as follows
[0021] Y=w0+w1X1+w2X2+…+w n X n +ε
[0022] Where: Y is the dependent variable, w1~w n is the regression coefficient, w0 is the intercept, X1~X n is the independent variable, ε is the error term;
[0023] When the error between the predicted result and the true result is the smallest, the corresponding regression parameter is the desired result.
[0024] As a preferred solution of the real-time online prediction method of converter steelmaking endpoint carbon temperature based on just-in-time learning described in the present invention, wherein: in step S3, the trust region reflective algorithm is specifically described as: the trust region reflective algorithm can be used to solve nonlinear least squares problems with constraints. Its basic principle is: take the approximate quadratic model m of the objective function in the neighborhood of the current iteration point k (s), the extreme value model m in this neighborhood k (s), solve the next iteration point, and adjust the trust region radius according to the similarity between the quadratic model and the objective function until the optimal solution is found.
[0025] To solve the above technical problem, according to another aspect of the present invention, the present invention provides the following technical solution:
[0026] A real-time online prediction system for converter steelmaking endpoint carbon temperature based on instant learning, comprising:
[0027] The converter production database construction module stores historical furnace process parameters and test result data sets, and synchronizes the current furnace process parameters and test result data sets in real time;
[0028] The sample set search module for local modeling, after the TSC measurement of the converter auxiliary gun of the current heat is completed, the TSC test results of the current heat and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of just-in-time learning (JITL) is used to find the heat with high similarity with the current heat in the historical heats as the sample set for local modeling;
[0029] The module for determining the decarburization coefficient and the temperature rise coefficient is based on the converter decarburization mechanism equation and the temperature rise mechanism equation. The module uses the historical furnaces with high similarity as the local sample set for local modeling. The trust region reflective (TRF) algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The multiple linear regression (MLR) algorithm is used to reversely solve the temperature rise coefficient of the local sample set and use it as the temperature rise coefficient of the current furnace.
[0030] The module for constructing a dynamic prediction model for the final carbon temperature is used to substitute the solved decarburization coefficient and temperature rise coefficient into the converter decarburization mechanism equation and temperature rise mechanism equation to construct a dynamic prediction model for the final carbon content and temperature of the current furnace.
[0031] The dynamic prediction model calculation module uses the TSC test results of the current furnace as the initial value for predicting carbon content and temperature, inputs real-time process parameters into the dynamic prediction model, and uses the model to calculate and output the predicted values of carbon content and temperature at different times in real time online until the end point of converter steelmaking is reached.
[0032] The beneficial effects of the present invention are as follows:
[0033] The present invention proposes a real-time online prediction method and system for the endpoint carbon temperature of converter steelmaking based on just-in-time learning. The just-in-time learning algorithm is used to query similar samples in a historical database in real time, and the optimal similar samples are selected for modeling. The model parameters that best match the current heat can be calculated based on the information of the current heat, thereby avoiding the limitations of the offline global model and effectively improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.
[0035] Figure 1 This is a schematic diagram of the architecture of converter endpoint carbon content prediction based on real-time learning according to Example 1 of the present invention.
[0036] Figure 2 Schematic diagram of the architecture of converter endpoint temperature prediction based on real-time learning according to Example 2 of the present invention.
[0037] Figure 3 This is a comparison chart of the results of predicting the converter endpoint carbon content using the real-time learning model of Example 1 of the present invention and the traditional auxiliary lance model.
[0038] Figure 4 This is a comparison chart of the results of predicting the converter endpoint temperature using the real-time learning model of Example 2 of the present invention and the traditional auxiliary lance model.
[0039] The realization of the effects, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0040] The following will be described clearly and completely in conjunction with the technical solutions in the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0041] The present invention aims to solve the problem that the traditional auxiliary gun model has poor adaptive ability and is difficult to be applied stably in the converter production process with complex scenes and changeable working conditions for a long time. It provides a real-time online prediction method and system for the converter end carbon temperature based on just-in-time learning. That is, the similarity measurement method of just-in-time learning (JITL) is used to find the furnace with high similarity with the current furnace in the historical furnaces as the sample for local modeling, and the trust region reflective (TRF) algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The multivariable linear regression (MLR) algorithm is used to reversely solve the heating coefficient of the local sample set and use it as the heating coefficient of the current furnace. The solved decarburization coefficient and heating coefficient are substituted into the converter decarburization mechanism equation and heating mechanism equation to realize the prediction of the converter end carbon temperature. The detailed steps of the prediction method include:
[0042] S1. Build a converter production database to store historical furnace process parameters and test result data sets, and synchronize the current furnace process parameters and test result data sets in real time;
[0043] S2. After the TSC measurement of the converter auxiliary gun of the current heat is completed, the TSC test results of the current heat and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of just-in-time learning (JITL) is used to find the heat with high similarity with the current heat in the historical heats as the sample set for local modeling;
[0044] S3. Based on the converter decarburization mechanism equation and the heating mechanism equation, the historical furnaces with high similarity are used as the local sample set for local modeling. The trust region reflective (TRF) algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The multivariable linear regression (MLR) algorithm is used to reversely solve the heating coefficient of the local sample set and use it as the heating coefficient of the current furnace.
[0045] S4, substituting the solved decarburization coefficient and temperature rise coefficient into the converter decarburization mechanism equation and the temperature rise mechanism equation to construct a dynamic prediction model of the carbon content and temperature at the end point of the current furnace;
[0046] S5. Using the TSC test results of the current furnace as the initial value for predicting carbon content and temperature, input the real-time process parameters into the dynamic prediction model, and use the model to calculate and output the predicted values of carbon content and temperature at different times in real time online until the end point of converter steelmaking is reached.
[0047] Preferably, in step S1, the converter production database includes a real-time database and a relational database, and the furnace process parameters and test result data sets include: actual molten iron quantity, actual scrap steel quantity, molten iron [C], molten iron [Si], molten iron [Mn], molten iron [P], molten iron [S], molten iron temperature, slag weight, molten steel quantity, actual value of ore, actual value of lime, actual value of dolomite, oxygen supply in the main blowing stage, actual dynamic oxygen supply, actual dynamic pellets, actual dynamic raw dolomite, target [C], target temperature, TSC detection carbon content, TSC detection temperature, TSO detection carbon content, TSO detection temperature. The above data are obtained from the primary database and secondary database of the steel enterprise. When constructing the historical data set, it is necessary to perform data preprocessing on the historical data to prevent the influence of abnormal data on the model prediction results of the model, and the data preprocessing includes the deletion of abnormal values and missing values.
[0048] Preferably, in step S2, the TSC test results include: carbon content of molten steel, temperature of molten steel; process parameters of the main blowing stage include: slag weight. The similarity measurement method based on real-time learning is specifically described as follows: the real-time learning algorithm adopts the "modeling and prediction" operation mode, selects the historical samples most relevant to the sample to be tested through the similarity measurement function, builds a local model online, and then makes a prediction. This solves the problem that the traditional global model is easily disturbed by multiple working conditions and changes in the operating environment. Before using the real-time learning method, the variable X needs to be normalized, and the min-max method is used for normalization. The formula is X norm =(XX min ) / (X max -X min ), where X norm is the normalized result, X max is the maximum value of variable X, X min is the minimum value of variable X. The similarity between samples is measured by the Euclidean distance method. The Euclidean distance formula is: d = || XX i ||, where: d is the distance between two samples; X is the feature vector of the target sample; X i is the feature vector of sample i in the training set, then the similarity between the two samples is
[0049] Preferably, in step S3, the converter decarburization mechanism equation is:
[0050]
[0051] Where: ω[C] t is the real-time carbon content, 10 -2 wt%; ω[C]0 is the carbon content of the converter when the decarburization rate is 0, 10- 2 wt%;ω[C] M Measuring carbon content for the secondary gun TSC, 10 -2 wt%;V(O2) X is the actual oxygen blowing amount, m 3 ; V(O2) XM is the oxygen blowing amount when the auxiliary gun starts to measure, m 3 ; B i is the oxygen content in the converter auxiliary material, m 3 / t; R i is the weight of auxiliary materials added after the auxiliary gun is measured, t; α, β are decarburization coefficients; W ST is the weight of molten steel, t.
[0052] Preferably, in step S3, the converter heating mechanism equation is:
[0053]
[0054] Where: T(t) is the predicted temperature of the molten steel at time t, °C; T M is the temperature of the molten steel at the time of auxiliary gun detection, ℃; γ is the oxygen blowing temperature rise coefficient, ℃ / (m 3 / t); δ is the temperature rise constant, ℃; K i is the cooling capacity of the coolant, ℃ / t.
[0055] Preferably, in step S3, the multiple linear regression algorithm is specifically described as follows: The multiple linear regression model studies the relationship between a dependent variable and multiple independent variables. The general multiple linear regression formula is as follows:
[0056] Y=w0+w1X1+w2X2+…+w n X n +ε
[0057] Where: Y is the dependent variable, w1~w n is the regression coefficient, w0 is the intercept, X1~X n is the independent variable, ε is the error term;
[0058] When the error between the predicted result and the true result is the smallest, the corresponding regression parameter is the desired result.
[0059] Preferably, in step S3, the trust region reflective algorithm is specifically described as: the trust region reflective algorithm can be used to solve nonlinear least squares problems with constraints. Its basic principle is: take an approximate quadratic model m of the objective function in the neighborhood of the current iteration point k (s), the extreme value model m in this neighborhood k(s), solve the next iteration point, and adjust the trust region radius according to the similarity between the quadratic model and the objective function until the optimal solution is found. The form of the constrained nonlinear least squares problem is as follows Where: l is the lower limit of the variable x, u is the upper limit of the variable x, and the solution steps of the trust region reflective algorithm are as follows:
[0060] (1) Set 0<λ<ξ<1, λ and ξ are constants, the initial iteration accuracy ε>0, the initial iteration point x0, and the trust region radius Δ0.
[0061] (2) Calculate the function value f at step k k , and calculate the gradient G k 、Hessian matrix H k .
[0062] (3) Solve the trust region model for displacement s Find the displacement s k , and calculate the ratio of the actual drop to the predicted drop ρ k =[f(x k )-f(x k +s k )] / [f(x k )-m k (s k )].
[0063] (4) Update the iteration point position x k+1 :If k >ξ, then x k+1 =x k +Δ k , otherwise, x k+1 =x k .
[0064] (5) Update the trust region radius Δ k+1 :If k >ξ, then Δ k+1 =2Δ k ; if λ≤ρ k ≤ξ, then Δ k+1 =Δ k If k <λ, then
[0065] (6) Return to step (2) and repeat until ||G k ||≤ε, find the optimal solution s*.
[0066] According to another aspect of the present invention, the present invention provides the following technical solution:
[0067] A real-time online prediction system for converter steelmaking endpoint carbon temperature based on instant learning, comprising:
[0068] The converter production database construction module stores historical furnace process parameters and test result data sets, and synchronizes the current furnace process parameters and test result data sets in real time;
[0069] The sample set search module for local modeling, after the TSC measurement of the converter auxiliary gun of the current heat is completed, the TSC test results of the current heat and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of just-in-time learning (JITL) is used to find the heat with high similarity with the current heat in the historical heats as the sample set for local modeling;
[0070] The module for determining the decarburization coefficient and the temperature rise coefficient is based on the converter decarburization mechanism equation and the temperature rise mechanism equation. The module uses the historical furnaces with high similarity as the local sample set for local modeling. The trust region reflective (TRF) algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The multivariable linear regression (MLR) algorithm is used to reversely solve the temperature rise coefficient of the local sample set and use it as the temperature rise coefficient of the current furnace.
[0071] The module for constructing a dynamic prediction model for the final carbon temperature is used to substitute the solved decarburization coefficient and temperature rise coefficient into the converter decarburization mechanism equation and temperature rise mechanism equation to construct a dynamic prediction model for the final carbon content and temperature of the current furnace.
[0072] The dynamic prediction model calculation module uses the TSC test results of the current furnace as the initial value for predicting carbon content and temperature, inputs real-time process parameters into the dynamic prediction model, and uses the model to calculate and output the predicted values of carbon content and temperature at different times in real time online until the end point of converter steelmaking is reached.
[0073] The technical solution of the present invention is further described below in conjunction with specific embodiments.
[0074] Example 1
[0075] The embodiment provides a method for predicting the carbon content of a converter end point based on real-time learning, and its architecture is as follows: Figure 1 The specific steps of the prediction method include:
[0076] S1. The actual production data of 2620 furnaces of silicon steel converters in a steel plant is used to build a converter production database, store historical furnace process parameters and test result data sets, and synchronize the current furnace process parameters and test result data sets in real time. The statistical results of the converter production database are shown in Table 1. Table 1 Statistical results of the converter production database
[0077] Variable Name Variable Symbols Maximum Minimum Mean Standard Deviation Molten iron weight <![CDATA[X1]]> 217.3 192.7 205.8 3.97 Scrap weight <![CDATA[X2]]> 49.38 23.53 37.14 3.77 Molten iron[C] <![CDATA[X3]]> 4.85 4.12 4.39 0.12 Molten iron[Si] <![CDATA[X4]]> 0.92 0.19 0.39 0.08 Molten iron[Mn] <![CDATA[X5]]> 0.63 0.2 0.29 0.04 Hot metal [P] <![CDATA[X6]]> 0.177 0.111 0.127 0.006 Hot metal[S] <![CDATA[X7]]> 0.1055 0.0001 0.007 0.013 Hot metal[T] <![CDATA[X8]]> 1425 1200 1321.44 36.75 Slag volume <![CDATA[X9]]> 35.35 8.513 21.931 4.817 Molten steel volume <![CDATA[X 10 ]]> 230.72 216.12 223.53 2.42 Ore addition amount <![CDATA[X 11 ]]> 7483 0 1710.33 1623.75 Lime addition amount <![CDATA[X 12 ]]> 8855 1113 4435.95 985.25 Dolomite addition <![CDATA[X 13 ]]> 3279 0 1527.09 555.49 Raw dolomite addition <![CDATA[X 14 ]]> 13919 0 2221.87 2479.09 Oxygen blowing amount <![CDATA[X 15 ]]> 10219 7731 9016.19 311.16 TSC[C] <![CDATA[X 16 ]]> 0.7511 0.0375 0.3673 0.1374 TSC[T] <![CDATA[X 17 ]]> 1659 1539 1600.46 20.50 TSO[C] <![CDATA[Y1]]> 0.0852 0.0244 0.0509 0.0118 TSO[T] <![CDATA[Y2]]> 1694 1619 1657.11 13.78
[0078] In step S1, the converter production database includes a real-time database and a relational database, and the furnace process parameters and test result data sets include: actual molten iron quantity, actual scrap steel quantity, molten iron [C], molten iron [Si], molten iron [Mn], molten iron [P], molten iron [S], molten iron temperature, slag weight, molten steel quantity, actual ore value, actual lime value, actual dolomite value, oxygen supply in the main blowing stage, actual dynamic oxygen supply, actual dynamic pelletizing, actual dynamic raw white, target [C], target temperature, TSC detected carbon content, TSC detected temperature, TSO detected carbon content, and TSO detected temperature.
[0079] S2. After the TSC measurement of the converter auxiliary gun of the current furnace is completed, the TSC test results of the current furnace and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of just-in-time learning (JITL) is used to find furnaces with high similarity with the current furnace in historical furnaces as samples for local modeling. The similarity measurement method based on just-in-time learning is specifically described as follows: the just-in-time learning algorithm adopts the "modeling and prediction" operation mode, selects the historical samples most relevant to the sample to be tested through the similarity measurement function, builds a local model online, and then predicts. It solves the problem that the traditional global model is easily disturbed by multiple working conditions and changes in the operating environment. Furthermore, before using the just-in-time learning method, the variable X needs to be normalized, and the min-max method is used for normalization. The formula is X norm =(XX min ) / (X max -X min ), where X norm is the normalized result, X max is the maximum value of variable X, X min is the minimum value of variable X. The similarity between samples is measured by the Euclidean distance method. The Euclidean distance formula is: d = || XX i ||, where: d is the distance between two samples; X is the feature vector of the target sample; X i is the feature vector of sample i in the training set, then the similarity between the two samples is TSC test results include: molten steel temperature; main blowing stage process parameters include: slag weight.
[0080] S3. Based on the converter decarburization mechanism equation, the historical furnaces with high similarity are used as local sample sets for local modeling. The trust region reflective (TRF) algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The trust region reflective algorithm is specifically described as: The trust region reflective algorithm can be used to solve nonlinear least squares problems with constraints. Its basic principle is: at the current iteration point x k Take the approximate quadratic model m of the objective function in the neighborhood of k (s), the extreme value model m in this neighborhood k (s), solve the next iteration point, and adjust the trust region radius according to the similarity between the quadratic model and the objective function until the optimal solution is found. The form of the constrained nonlinear least squares problem is as follows Where: l is the lower limit of the variable x, u is the upper limit of the variable x, and the solution steps of the trust region reflective algorithm are as follows: (1) Set 0<λ<ξ<1, λ and ξ are constants, initialize the iteration accuracy ε>0, the initial iteration point is x0, and the trust region radius is Δ0. (2) Calculate the function value f at the kth step k , and calculate the gradient G k 、Hessian matrix H k (3) Solve the trust region model for displacement s Find the displacement s k , and calculate the ratio of the actual drop to the predicted drop ρ k =[f(x k )-f(x k +s k )] / [f(x k )-m k (s k )]. (4) Update the iteration point position x k+1 :If k >ξ, then x k+1 =x k +Δ k , otherwise, x k+1 =x k (5) Update the trust region radius Δ k+1 :If k >ξ, then Δ k+1 =2Δ k ; if λ≤ρ k ≤ξ, then Δ k+1 =Δ k If k <λ, then (6) Return to step (2) and repeat until ||G k ||≤ε, find the optimal solution s*. The converter decarburization mechanism equation is:
[0081]
[0082] Where: ω[C] t is the real-time carbon content, 10 -2 wt%; ω[C]0 is the carbon content of the converter when the decarburization rate is 0, 10 - 2 wt%;ω[C] M Measuring carbon content for the secondary gun TSC, 10 -2 wt%;V(O2) X is the actual oxygen blowing amount, m 3 ; V(O2) XM is the oxygen blowing amount when the auxiliary gun starts to measure, m 3 ; B i is the oxygen content in the converter auxiliary material, m 3 / t; R i is the weight of auxiliary materials added after the auxiliary gun is measured, t; α, β are decarburization coefficients; W ST is the weight of molten steel, t.
[0083] S4, substituting the solved decarburization coefficient into the converter decarburization mechanism equation to construct a dynamic prediction model for the carbon content at the end point of the current furnace;
[0084] S5. Using the TSC test result of the current furnace as the initial value for carbon content prediction, input the real-time process parameters into the dynamic prediction model, and use the model to calculate and output the predicted carbon content values at different times in real time online until the converter steelmaking endpoint target is reached.
[0085] When the local modeling sample size is 220, the prediction results of the hybrid model and the traditional dynamic model are as follows: Figure 3 As shown. Figure 3 It can be seen that the prediction results of the instant learning model are closer to the true value than those of the secondary gun model. The comparative accuracy of the models is shown in Table 2.
[0086] Table 2 compares the prediction results of the real-time learning model and the auxiliary gun model, that is, the comparison between the auxiliary gun model and the real-time learning modeling method of the present invention. The present invention uses multiple error metrics such as mean absolute error MAE, mean square error RMSE, hit rate HR, etc. as comprehensive evaluation indicators.
[0087]
[0088]
[0089]
[0090] Where: n is the number of samples; is the model prediction value; y i is the true value, n(|真实值-预测值|<ε) is the number of samples within the error range.
[0091] Table 2 Prediction results of converter end point carbon content by real-time learning model and auxiliary lance model
[0092] Model MAE / wt% RMSE / wt% HR / % Just-in-time learning model 0.00989 0.01271 87.82 Secondary gun model 0.01068 0.01382 84.50
[0093] Example 2
[0094] The embodiment provides a method for predicting converter end point temperature based on real-time learning, and its architecture is as follows: Figure 2 The specific steps of the prediction method include:
[0095] S1. The actual production data of 2620 furnaces of silicon steel converters in a steel plant are used to build a converter production database, store historical furnace process parameters and test result data sets, and synchronize the current furnace process parameters and test result data sets in real time. The statistical results of the converter production database are shown in Table 3.
[0096] Table 3 Statistical results of converter production database
[0097] Variable Name Variable Symbols Maximum Minimum Mean Standard Deviation Molten iron weight <![CDATA[X1]]> 217.3 192.7 205.8 3.97 Scrap weight <![CDATA[X2]]> 49.38 23.53 37.14 3.77 Molten iron[C] <![CDATA[X3]]> 4.85 4.12 4.39 0.12 Molten iron[Si] <![CDATA[X4]]> 0.92 0.19 0.39 0.08 Molten iron[Mn] <![CDATA[X5]]> 0.63 0.2 0.29 0.04 Hot metal [P] <![CDATA[X6]]> 0.177 0.111 0.127 0.006 Hot metal[S] <![CDATA[X7]]> 0.1055 0.0001 0.007 0.013 Hot metal[T] <![CDATA[X8]]> 1425 1200 1321.44 36.75 Slag volume <![CDATA[X9]]> 35.35 8.513 21.931 4.817 Molten steel volume <![CDATA[X 10 ]]> 230.72 216.12 223.53 2.42 Ore addition amount <![CDATA[X 11 ]]> 7483 0 1710.33 1623.75 Lime addition amount <![CDATA[X 12 ]]> 8855 1113 4435.95 985.25 Dolomite addition <![CDATA[X 13 ]]> 3279 0 1527.09 555.49 Raw dolomite addition <![CDATA[X 14 ]]> 13919 0 2221.87 2479.09 Oxygen blowing amount <![CDATA[X 15 ]]> 10219 7731 9016.19 311.16 TSC[C] <![CDATA[X 16 ]]> 0.7511 0.0375 0.3673 0.1374 TSC[T] <![CDATA[X 17 ]]> 1659 1539 1600.46 20.50 TSO[C] <![CDATA[Y1]]> 0.0852 0.0244 0.0509 0.0118 TSO[T] <![CDATA[Y2]]> 1694 1619 1657.11 13.78
[0098] In step S1, the converter production database includes a real-time database and a relational database, and the furnace process parameters and test result data sets include: actual molten iron quantity, actual scrap steel quantity, molten iron [C], molten iron [Si], molten iron [Mn], molten iron [P], molten iron [S], molten iron temperature, slag weight, molten steel quantity, actual ore value, actual lime value, actual dolomite value, oxygen supply in the main blowing stage, actual dynamic oxygen supply, actual dynamic pelletizing, actual dynamic raw white, target [C], target temperature, TSC detected carbon content, TSC detected temperature, TSO detected carbon content, and TSO detected temperature.
[0099] S2. After the TSC measurement of the converter auxiliary gun of the current furnace is completed, the TSC test results of the current furnace and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of just-in-time learning (JITL) is used to find furnaces with high similarity with the current furnace in historical furnaces as samples for local modeling. The similarity measurement method based on just-in-time learning is specifically described as follows: the just-in-time learning algorithm adopts the "modeling and prediction" operation mode, selects the historical samples most relevant to the sample to be tested through the similarity measurement function, builds a local model online, and then predicts. It solves the problem that the traditional global model is easily disturbed by multiple working conditions and changes in the operating environment. Furthermore, before using the just-in-time learning method, the variable X needs to be normalized, and the min-max method is used for normalization. The formula is X norm =(XX min ) / (Xmax -X min ), where X norm is the normalized result, X max is the maximum value of variable X, X min is the minimum value of variable X. Furthermore, the similarity between samples is measured by the Euclidean distance method. The Euclidean distance formula is: d = || XX i ||, where: d is the distance between two samples; X is the feature vector of the target sample; X i is the feature vector of sample i in the training set, then the similarity between the two samples is TSC test results include: molten steel temperature, molten steel carbon content; main blowing stage process parameters include: slag weight.
[0100] S3. Based on the converter heating mechanism equation, the historical furnaces with high similarity are used as local sample sets for local modeling. The multiple linear regression (MLR) algorithm is used to reversely solve the heating coefficient of the local sample and use it as the heating coefficient of the current furnace. The multiple linear regression algorithm is specifically described as follows: The multiple linear regression model studies the relationship between a dependent variable and multiple independent variables. The general multiple linear regression formula is as follows:
[0101] Y=w0+w1X1+w2X2+…+w n X n +ε
[0102] Where: Y is the dependent variable, w1~w n is the regression coefficient, w0 is the intercept, X1~X n is the independent variable, ε is the error term;
[0103] When the error between the predicted result and the true result is the smallest, the corresponding regression parameter is the desired result.
[0104] The converter heating mechanism equation is:
[0105]
[0106] Where: T(t) is the predicted temperature of the molten steel at time t, °C; T M is the temperature of the molten steel at the time of auxiliary gun detection, ℃; γ is the oxygen blowing temperature rise coefficient, ℃ / (m 3 / t); δ is the temperature rise constant, ℃; K i is the cooling capacity of the coolant, ℃ / t.
[0107] S4, substituting the solved temperature rise coefficient into the converter temperature rise mechanism equation to construct a dynamic prediction model for the current furnace end temperature;
[0108] S5. Using the TSC test result of the current furnace as the initial value for temperature prediction, input the real-time process parameters into the dynamic prediction model, and use the model to calculate and output the temperature prediction values at different times in real time online until the converter steelmaking endpoint target is reached.
[0109] When the local modeling sample size is 70, the prediction results of the hybrid model and the traditional dynamic model are as follows: Figure 4 As shown. Figure 4 It can be seen that the prediction results of the instant learning model are closer to the true value than the secondary gun model. The comparative accuracy of the models is shown in Table 4.
[0110] Table 4 is a comparison of the prediction results of the converter endpoint temperature between the real-time learning model and the dynamic model, that is, a comparison between the auxiliary gun model and the real-time learning modeling method of the present invention. The present invention uses a variety of error metrics such as mean absolute error MAE, mean square error RMSE, hit rate HR, etc. as comprehensive evaluation indicators.
[0111]
[0112]
[0113]
[0114] Where: n is the number of samples; is the model prediction value; y i is the true value, n (|真实值-预测值|<ε) is the number of samples within the error range.
[0115] Table 4 Prediction results of converter end temperature by real-time learning model and auxiliary lance model
[0116] Model MAE / ℃ RMSE / ℃ HR / % Just-in-time learning model 8.37 10.70 85.15 Secondary gun model 9.22 11.74 80.38
[0117] The embodiments of the present invention use a real-time learning algorithm to query similar samples in a historical database in real time, select the best similar samples for modeling, and can calculate the model parameters that best match the current furnace based on the information of the current furnace, thereby avoiding the limitations of the offline global model and effectively improving the prediction accuracy of the model.
[0118] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present invention.
Claims
1. A real-time online prediction method for converter steelmaking endpoint carbon temperature based on real-time learning, characterized in that: The following steps are involved: S1. Build a converter production database to store historical furnace process parameters and test result data sets, and synchronize the current furnace process parameters and test result data sets in real time; S2. After the TSC measurement of the converter auxiliary gun of the current furnace is completed, the TSC detection results of the current furnace and the process parameters of the main blowing stage are extracted from the converter production database as feature items. The similarity measurement method of real-time learning is used to find furnaces with high similarity with the current furnace in historical furnaces as sample sets for local modeling; S3. Based on the converter decarburization mechanism equation and the heating mechanism equation, local modeling is performed using historical furnaces with high similarity as local sample sets, and the trust region reflection algorithm is used to reversely solve the decarburization coefficient of the local sample and use it as the decarburization coefficient of the current furnace. The multivariate linear regression algorithm is used to reversely solve the heating coefficient of the local sample set and use it as the heating coefficient of the current furnace; S4, substituting the solved decarburization coefficient and temperature rise coefficient into the converter decarburization mechanism equation and the temperature rise mechanism equation to construct a dynamic prediction model of the carbon content and temperature at the end point of the current furnace; S5. Using the TSC test results of the current furnace as the initial value for predicting carbon content and temperature, input the real-time process parameters into the dynamic prediction model, and use the model to calculate and output the predicted values of carbon content and temperature at different times in real time online until the end point of converter steelmaking is reached.
2. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S1, the converter production database includes a real-time database and a relational database, and the furnace process parameters and test result data sets include: actual molten iron quantity, actual scrap steel quantity, molten iron [C], molten iron [Si], molten iron [Mn], molten iron [P], molten iron [S], molten iron temperature, slag weight, molten steel quantity, actual ore value, actual lime value, actual dolomite value, oxygen supply in the main blowing stage, actual dynamic oxygen supply, actual dynamic pellets, actual dynamic raw dolomite, target [C], target temperature, TSC detected carbon content, TSC detected temperature, TSO detected carbon content, and TSO detected temperature.
3. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S2, the TSC test results include: molten steel carbon content, molten steel temperature; the main blowing stage process parameters include: slag weight.
4. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S2, the similarity measurement method based on real-time learning is specifically described as follows: the real-time learning algorithm adopts a modeling and prediction operation mode, selects the historical samples most relevant to the sample to be tested through the similarity measurement function, builds a local model online, and then makes predictions.
5. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S3, the converter decarburization mechanism equation is: Where: ω[C] t is the real-time carbon content, 10 -2 wt%; ω[C]0 is the carbon content of the converter when the decarburization rate is 0, 10 - 2 wt%;ω[C] M Measuring carbon content for the secondary gun TSC, 10 -2 wt%;V(O2) X is the actual oxygen blowing amount, m 3 ; V(O2) XM is the oxygen blowing amount when the auxiliary gun starts to measure, m 3 ; B i is the oxygen content in the converter auxiliary material, m 3 / t; R i is the weight of auxiliary materials added after the auxiliary gun is measured, t; α, β are decarburization coefficients; W ST is the weight of molten steel, t.
6. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S3, the converter heating mechanism equation is: Where: T(t) is the predicted temperature of the molten steel at time t, °C; T M is the temperature of the molten steel at the time of auxiliary gun detection, ℃; γ is the oxygen blowing temperature rise coefficient, ℃ / (m 3 / t); δ is the temperature rise constant, ℃; K i is the cooling capacity of the coolant, ℃ / t.
7. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S3, the multiple linear regression algorithm formula is as follows: Y=w0+w1X1+w2X2+…+w n X n +e Where: Y is the dependent variable, w1~w n is the regression coefficient, w0 is the intercept, X1~X n is the independent variable, ε is the error term; When the error between the predicted result and the true result is the smallest, the corresponding regression parameter is the desired result.
8. The real-time online prediction method for converter steelmaking endpoint carbon temperature based on just-in-time learning according to claim 1 is characterized in that: In step S3, the trust region reflective algorithm is specifically as follows: an approximate quadratic model m of the objective function is taken in the neighborhood of the current iteration point. k (s), the extreme value model m in this neighborhood k (s), solve the next iteration point, and adjust the trust region radius according to the similarity between the quadratic model and the objective function until the optimal solution is found.
9. A real-time online prediction system for converter steelmaking endpoint carbon temperature based on real-time learning, characterized in that: The method for realizing the real-time online prediction of the endpoint carbon temperature of converter steelmaking based on instant learning as described in claim 1 comprises: The converter production database construction module stores historical furnace process parameters and test result data sets, and synchronizes the current furnace process parameters and test result data sets in real time; The sample set search module for local modeling, after the TSC measurement of the converter auxiliary gun of the current furnace is completed, the TSC test results of the current furnace and the process parameters of the main blowing stage are extracted from the converter production database as feature items, and the similarity measurement method of real-time learning is used to find the furnace with high similarity with the current furnace in the historical furnaces as the sample set for local modeling; The module for determining the decarburization coefficient and the temperature rise coefficient is based on the converter decarburization mechanism equation and the temperature rise mechanism equation. The module uses the historical furnaces with high similarity as the local sample set for local modeling. The module uses the trust region reflection algorithm to reversely solve the decarburization coefficient of the local sample and uses it as the decarburization coefficient of the current furnace. The module uses the multivariate linear regression algorithm to reversely solve the temperature rise coefficient of the local sample set and uses it as the temperature rise coefficient of the current furnace. The module for constructing a dynamic prediction model for the final carbon temperature is used to substitute the solved decarburization coefficient and temperature rise coefficient into the converter decarburization mechanism equation and temperature rise mechanism equation to construct a dynamic prediction model for the final carbon content and temperature of the current furnace. The dynamic prediction model calculation module uses the TSC test results of the current furnace as the initial value for predicting carbon content and temperature, inputs real-time process parameters into the dynamic prediction model, and uses the model to calculate and output the predicted values of carbon content and temperature at different times in real time online until the end point of converter steelmaking is reached.
Citation Information
Patent Citations
Method for predicting end-point carbon content and end-point temperature of small and medium-sized converter
CN115456264A
Method for controlling end point of blowing in converter
JP2006233324A
Method of refining molten steel in converter
KR1020140017180A
Cited By
Method for constructing LF ladle refining prediction model based on knowledge graph
CN119721214A
Gas recovery regulation and control method and system based on automatic control and data processing
CN119824165A
Intelligent online optimization method of converter steelmaking process for low-carbon and low-cost smelting
CN119956017A
An intelligent online optimization method for converter steelmaking process oriented to low-carbon and low-cost smelting
CN119956017B
Method and system for predicting end-point alloy components of RH refined molten steel based on Kmean-BO-RF
CN120089224A
