A real-time prediction method and system for the carbon content of molten steel in RH refining based on machine learning

Abstract

The present invention belongs to the technical field of iron and steel metallurgy, and specifically relates to a real-time prediction method and system for the carbon content of molten steel in RH refining based on machine learning. By analyzing and mining the time-series data of RH refining flue gas, a flue gas correction coefficient and representative characteristic variables are proposed around the RH decarburization process, and a prediction model is constructed using machine learning algorithms. It can fully mine the characteristic parameters in the flue gas data, describe the deviation between the flue gas data and the actual carbon content, and realize the high-precision real-time prediction of the carbon content of molten steel in the RH refining process.

Description

Technical Field

[0001] The present invention relates to the technical field of iron and steel metallurgy, and particularly to a real-time prediction method and system for the carbon content of molten steel in RH refining based on machine learning. Background Art

[0002] Decarburization is a key step in the RH refining process. Through vacuum circulation degassing, carbon and oxygen in the molten steel react to form carbon monoxide or carbon dioxide and are discharged through a vacuum pump, thereby achieving the purpose of decarburization. In order to achieve precise control of the decarburization process, by real-time predicting the carbon content of the molten steel, operators can timely understand the current refining state, and thus adjust parameters such as oxygen blowing amount and temperature to optimize the refining process, which helps to shorten the refining cycle and improve production efficiency.

[0003] At present, real-time prediction of the carbon content of molten steel in RH refining in iron and steel enterprises is rare, and the refining process still mainly relies on manual experience for process control. Some scholars have carried out research on real-time prediction methods for the carbon content of molten steel in RH refining, such as: combining the carbon and oxygen balance equations of the ladle and the vacuum chamber to construct a decarburization mechanism model for RH refining; on the basis of the decarburization mechanism model for RH refining, introducing a decarburization factor, and determining the decarburization factor under different production scenarios through data driving to improve the adaptability of the mechanism model to the complex production environment; and also real-time predicting the carbon content of the molten steel based on carbon balance calculation through the collected real-time flue gas data of RH refining. However, the above-mentioned methods proposed have not been well applied in the production site, mainly due to the following problems: 1) The black-box model constructed based on the decarburization mechanism has many assumed parameters and is difficult to accurately predict the real-time carbon content of the molten steel due to the complex changes in different real-time sensed on-site production conditions; 2) Although flue gas data collection is an important means to improve the perception of the RH refining production site, it is difficult to obtain satisfactory prediction results for the carbon content of the molten steel only by carbon balance calculation; 3) The current research methods still do not fully exploit the RH refining flue gas data, resulting in poor application effects of machine learning in flue gas data. Summary of the Invention

[0004] To solve the problems existing in the prior art, the main object of the present invention is to propose a real-time prediction method and system for the carbon content of molten steel in RH refining based on machine learning. By analyzing and mining the time series data of RH refining flue gas, a flue gas correction coefficient and representative characteristic variables are proposed around the RH decarburization process, and a prediction model is constructed using a machine learning algorithm to achieve higher-precision real-time prediction of the carbon content of molten steel in the RH refining process.

[0005] According to one aspect of the present invention, the following technical solution is provided:

[0006] A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning, comprising the following steps:

[0007] S1. Collect the production data of RH refining historical heats, screen and preprocess abnormal heats, and process the flue gas time-series data of each historical heat as follows:

[0008] S11. Divide the flue gas time-series data into multiple segments according to the set time step, and calculate the cumulative decarburization amount, cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content within each segment through flue gas integration.

[0009] S12. Superimpose the cumulative decarburization amounts within all segments to obtain the cumulative decarburization amount calculated based on flue gas analysis during the smelting cycle.

[0010] S13. Calculate the actual decarburization amount during the smelting cycle according to the initial carbon content and end-point carbon content data of this historical heat.

[0011] S14. Define the flue gas correction coefficient to represent the deviation between flue gas analysis and actual detection, and obtain it by calculating the ratio of the actual decarburization amount to the cumulative decarburization amount.

[0012] S15. Combine the flue gas correction coefficient to calculate the actual decarburization amount within each time-step segment.

[0013] S2. Store the process parameters related to the decarburization process in the RH refining production data and the parameters generated after processing the flue gas time-series data in step S1 into the model database for subsequent modeling.

[0014] S3. Take the cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, average CO2 content, and process parameters within each time-step segment as model input variables, take the actual decarburization amount within each time-step segment as the model output variable, use machine learning algorithms for weight extraction and feature selection, and construct a prediction model.

[0015] S4. Real-time collect the production data of the current heat in RH refining. After each time step ends, calculate the cumulative decarburization amount, cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content within this time step; substitute the model input variables into the prediction model to output the prediction result of the actual decarburization amount within each time-step segment; combine the initial carbon content of the current heat to realize the real-time prediction of the molten steel carbon content.

[0016] According to another aspect of the present invention, the present invention provides the following technical solution:

[0017] A real-time prediction system for molten steel carbon content in RH refining based on machine learning, comprising:

[0018] Model database construction module: Collect the production data of historical RH refining heats, screen and preprocess abnormal heats, process the flue gas time-series data of each historical heat, and construct a model database for modeling.

[0019] Prediction model construction module: Determine the input variables and output variables of the model, and construct a prediction model for the actual decarburization amount within each time step through machine learning algorithms.

[0020] Real-time molten steel carbon content prediction module: According to the collected real-time heat production data, output the prediction results of the actual decarburization amount within each time step through the model, and realize the real-time prediction of the molten steel carbon content.

[0021] The beneficial effects of the present invention are as follows:

[0022] The present invention proposes a real-time molten steel carbon content prediction method and system based on machine learning. By analyzing and mining the RH refining flue gas time-series data, a flue gas correction coefficient and representative characteristic variables are proposed around the RH decarburization process, and a prediction model is constructed using machine learning algorithms. It can fully mine the characteristic parameters in the flue gas data, describe the deviation between the flue gas data and the actual carbon content, and realize the high-precision real-time prediction of the molten steel carbon content in the RH refining process. Brief description of the drawings

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the structures shown in these drawings without creative efforts.

[0024] Figure 1 It is a framework schematic diagram of the prediction method of the present invention.

[0025] Figure 2 It is the real-time prediction of the molten steel carbon content for heats with an incoming carbon content of 200 - 300 ppm in the embodiments of the present invention.

[0026] Figure 3 It is the real-time prediction of the molten steel carbon content for heats with an incoming carbon content of 300 - 400 ppm in the embodiments of the present invention.

[0027] Figure 4 It is the real-time prediction of the molten steel carbon content for heats with an incoming carbon content of 400 - 500 ppm in the embodiments of the present invention.

[0028] Figure 5 It is the real-time prediction of the molten steel carbon content for heats with an incoming carbon content of more than 500 ppm in the embodiments of the present invention.

[0029] The realization, functional features, and advantages of the present invention will be further described in conjunction with embodiments and with reference to the accompanying drawings. Specific Embodiments

[0030] The technical solutions in the embodiments will be clearly and completely described below. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0031] The present invention provides a real-time prediction method and system for the carbon content of molten steel in RH refining based on machine learning. By analyzing and mining the time-series data of RH refining flue gas, a flue gas correction coefficient and representative characteristic variables are proposed for the RH decarburization process, and a prediction model is constructed using machine learning algorithms to achieve higher-precision real-time prediction of the carbon content of molten steel in the RH refining process.

[0032] According to one aspect of the present invention, the following technical solutions are provided:

[0033] As Figure 1 shown, a real-time prediction method for the carbon content of molten steel in RH refining based on machine learning includes the following steps:

[0034] S1. Collect the production data of historical RH refining heats, screen and preprocess abnormal heats, and process the time-series data of the flue gas for each historical heat as follows:

[0035] S11. Divide the time-series data of the flue gas into multiple segments according to the set time step, and calculate the cumulative decarburization amount, cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content within each segment through flue gas integration.

[0036] S12. Superimpose the cumulative decarburization amounts within all segments to obtain the cumulative decarburization amount calculated based on flue gas analysis during the smelting cycle.

[0037] S13. Calculate the actual decarburization amount during the smelting cycle based on the initial carbon content and end carbon content data of this historical heat.

[0038] S14. Define the flue gas correction coefficient to represent the deviation between flue gas analysis and actual detection, and obtain it by calculating the ratio of the actual decarburization amount to the cumulative decarburization amount.

[0039] S15. Combine the flue gas correction coefficient to calculate the actual decarburization amount within each time-step segment.

[0040] S2. Store the process parameters related to the decarbonization process in the RH refining production data and the parameters generated after processing the flue gas time series data in step S1 into the model database for subsequent modeling;

[0041] S3. Use the cumulative flue gas volume and average flue gas content, cumulative CO volume and average CO content, cumulative CO2 volume and average CO2 content, and process parameters within each time step segment as model input variables, and use the actual decarbonization amount within each time step segment as the model output variable. Adopt a machine learning algorithm for weight extraction and feature selection, and construct a prediction model;

[0042] S4. Real-time collect the production data of the current heat in RH refining. After each time step ends, calculate the cumulative decarbonization amount, cumulative flue gas volume and average flue gas content, cumulative CO volume and average CO content, cumulative CO2 volume and average CO2 content within the time step; substitute the model input variables into the prediction model to output the predicted result of the actual decarbonization amount within each time step segment; combine with the initial carbon content of the current heat to realize the real-time prediction of the molten steel carbon content.

[0043] Preferably, in step S1, the following situations are determined as abnormal heats: 1) Adding carbon powder during the production process; 2) Adjusting the temperature by adding scrap steel; 3) Abnormal signal acquisition; 4) Abnormal detection results.

[0044] Preferably, in step S1, the lower limit of the time step setting should be greater than the flue gas data acquisition time interval.

[0045] Preferably, in step S1, the calculation method of the cumulative decarbonization amount within each time step is

[0046]

[0047] In the formula, ΔC i represents the cumulative decarbonization amount within the time step, ppm; Q gas represents the flue gas flow rate, kg / h; represents the percentage content of CO in the flue gas, %; represents the percentage content of CO2 in the flue gas, %; M C represents the relative atomic mass of C; M CO represents the relative atomic mass of CO; represents the relative atomic mass of CO2; t represents time, s; W steel represents the weight of the molten steel, kg.

[0048] Preferably, in step S1, the calculation methods of the cumulative flue gas volume and average flue gas content, cumulative CO volume and average CO content, cumulative CO2 volume and average CO2 content within each time step segment are

[0049] m 烟气累积 = ∫Q gas dt

[0050] m 烟气平均 = ∫Q gas dt / t

[0051]

[0052]

[0053]

[0054]

[0055] Preferably, in the step S1, the calculation method of the flue gas correction coefficient is

[0056]

[0057] In the formula, W represents the flue gas correction coefficient; C represents the carbon content at the inlet station, ppm; C ′ represents the carbon content at the outlet station, ppm; ∫ΔC i represents the cumulative decarbonization amount of the flue gas during the smelting time.

[0058] Preferably, in the step S1, the calculation method of the actual decarbonization amount in each time step is

[0059] ΔC i ′ = W·ΔC i

[0060] Preferably, in the step S2, the process parameters related to decarbonization in the RH refining production data include the amount of molten steel, the carbon content at arrival, the oxygen activity at arrival, the vacuum chamber life, the immersion tube life, the temperature at arrival, the slag layer thickness, the decarbonization time, etc.

[0061] Preferably, in the step S3, if there is a delay in the on-site flue gas data, the real-time prediction result of the molten steel carbon content should consider the delay time.

[0062] According to another aspect of the present invention, the present invention provides the following technical solution:

[0063] A real-time prediction system for the carbon content of molten steel in RH refining based on machine learning, comprising:

[0064] Model database construction module: Collect the production data of historical RH refining heats, screen and preprocess abnormal heats, and process the flue gas time series data of each historical heat to construct a model database for modeling;

[0065] Prediction model construction module: Determine the input variables and output variables of the model, and construct a prediction model for the actual decarbonization amount within each time step through machine learning algorithms;

[0066] Real-time prediction module for molten steel carbon content: According to the collected real-time furnace production data, the prediction results of the actual decarbonization amount within each time step are output through the model to achieve real-time prediction of the molten steel carbon content.

[0067] Embodiment

[0068] Taking the RH refining production data of low-carbon and extra-low-carbon steel in a certain steel plant from August 2023 to October 2024 as an example, after the data extraction is completed, abnormal furnace runs are screened. If a furnace run has the following five situations during production, it is identified as an abnormal furnace run:

[0069] 1) Adding carbon powder during production, which is determined as an abnormal furnace run during production;

[0070] 2) Adding scrap steel to adjust the temperature. Since the end-point carbon content is relatively low and the addition amount of scrap steel is large, it has a greater impact on the end-point carbon content. Therefore, the furnace runs with added scrap steel are screened out to exclude their influence on the real-time carbon content calculation;

[0071] 3) The off-station carbon content is greater than 30 ppm. The off-station requirement for low-carbon and extra-low-carbon steel is that the carbon content is less than 30 ppm. If the carbon content at off-station is too high, it is determined as abnormal production;

[0072] 4) Abnormal signal acquisition, without parameters such as flue gas composition, vacuum degree, and oxygen blowing amount in the trend software;

[0073] 5) Abnormal detection results, without recording the in-station carbon content, off-station carbon content, or the off-station carbon content detection result is 0.

[0074] After screening according to the above criteria, 1401 furnaces meet the requirements. On-site, an energy spectrometer is used to detect the flue gas at the cold end, and at the same time, the real-time flue gas flow is obtained through a flue gas flow detector.

[0075] Set the time step to 10 s. For the flue gas time series data of 1401 furnace runs, calculate the cumulative decarbonization amount, cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content at intervals of 10 s; by calculating the ratio of the actual decarbonization amount to the cumulative decarbonization amount within the smelting cycle of the furnace run, obtain the flue gas correction coefficient, and then calculate the actual decarbonization amount within each time step; combined with the above analysis and mining of the flue gas time series data, a total of 1,734,553 data are formed and stored in the model database.

[0076] Use the XGBoost algorithm to construct a prediction model, and the model expression is:

[0077]

[0078] where T represents the number of regression trees; x i represents the feature value of the i-th sample; represents the predicted value of the i-th sample after T iterations of the model; f j (x i ) represents the prediction result of the j-th tree. As an additive model, it can be transformed into the sum of the results of the previous T - 1 trees plus the last tree.

[0079] The objective function obj of the XGBoost algorithm consists of the loss function L and the regularization term complexity function Ω:

[0080]

[0081] where L represents the mean squared error of the loss function; Ω represents the regularization term for the loss caused by the model complexity, which is obtained by summing the regularization terms of T classification and regression trees.

[0082] The complexity of the regression tree represented by the regularization term complexity function Ω is affected by two points. One is the weight w on the leaf nodes, and the other is the number of leaf nodes T1 on the tree. The smaller the parameters of these two parts, the lower the complexity of the model. The expression is:

[0083]

[0084] where γ represents the difficulty of node splitting; λ represents the regularization coefficient.

[0085] The XGBoost algorithm sets four types of hyperparameters, namely general hyperparameters, learning task parameters, Booster parameters, and regularization parameters, as shown in Table 1.

[0086] Table 1 Hyperparameter Settings of the XGBoost Algorithm

[0087]

[0088]

[0089] Predictions were made for four different furnace heats with different inlet carbon contents on-site. The real-time prediction results of the molten steel carbon content are as Figures 2 to 5 shown. The inlet carbon contents are in four intervals of 200 - 300 ppm, 300 - 400 ppm, 400 - 500 ppm, and above 500 ppm respectively. It can be found from the figure that this method realizes the real-time and accurate prediction of the molten steel carbon content during the RH refining process under various inlet carbon content conditions.

[0090] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structural transformation made by using the content of the specification of the present invention under the inventive concept of the present invention, or direct / indirect application in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning, characterized in that, It includes the following steps: S1. Collect the production data of historical heats in RH refining, screen and preprocess abnormal heats, and process the flue gas time-series data of each historical heat as follows: S11. Divide the flue gas time-series data into multiple segments according to the set time step. Calculate the cumulative decarburization amount, cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content within each segment through flue gas integration; S12. Superimpose the cumulative decarburization amounts within all segments to obtain the cumulative decarburization amount calculated based on flue gas analysis during the smelting cycle; S13. Calculate the actual decarburization amount during the smelting cycle according to the initial carbon content and end-point carbon content data of this historical heat; S14. Define the flue gas correction coefficient to represent the deviation between flue gas analysis and actual detection, and obtain it by calculating the ratio of the actual decarburization amount to the cumulative decarburization amount; S15. Combine the flue gas correction coefficient to calculate the actual decarburization amount within each time-step segment; S2. Store the process parameters related to the decarburization process in the RH refining production data and the parameters generated after processing the flue gas time-series data in step S1 into the model database for subsequent modeling; S3. Take the cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, average CO2 content, and process parameters within each time-step segment as model input variables, take the actual decarburization amount within each time-step segment as the model output variable, use machine learning algorithms for weight extraction and feature selection, and construct a prediction model; S4. Real-time collect the production data of the current heat in RH refining. After each time step ends, calculate the cumulative decarburization amount, cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content within this time step; substitute the model input variables into the prediction model to output the predicted result of the actual decarburization amount within each time-step segment; combine with the initial carbon content of the current heat to achieve real-time prediction of the molten steel carbon content; In step S1, the calculation method of the cumulative decarburization amount within each time step is In the formula, represents the cumulative decarbonization amount within the time step, in ppm; represents the flue gas flow rate, in kg / h; represents the percentage content of CO in the flue gas, in %; represents the percentage content of CO2 in the flue gas, in %; represents the relative atomic mass of C; represents the relative atomic mass of CO; represents the relative atomic mass of CO2; t represents time, in s; represents the weight of molten steel, in kg; In step S1, the calculation methods of the cumulative flue gas amount, average flue gas content, cumulative CO amount, average CO content, cumulative CO2 amount, and average CO2 content within each time-step segment are 。 2. The real-time prediction method for the carbon content of molten steel in RH refining based on machine learning according to claim 1, characterized in that, In step S1, the following situations are judged as abnormal heats: 1) Adding carbon powder during the production process; 2) Adjusting the temperature by adding scrap steel; 3) Abnormal signal collection; 4) Abnormal detection results.

3. A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning according to claim 1, characterized in that, In step S1, the lower limit of the set time step should be greater than the flue gas data collection time interval.

4. A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning according to claim 1, characterized in that, In step S1, the calculation method of the flue gas correction coefficient is Wherein, represents the flue gas correction coefficient; represents the carbon content at the inlet, ppm; represents the carbon content at the outlet, ppm; represents the cumulative decarbonization amount of the flue gas during the smelting time.

5. A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning according to claim 1, characterized in that, In step S1, the calculation method of the actual decarburization amount within each time step is 。 6. A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning according to claim 1, characterized in that, In step S2, the process parameters related to decarburization in the RH refining production data include the molten steel amount, incoming station carbon content, incoming station oxygen activity, vacuum chamber life, dip tube life, incoming station temperature, slag layer thickness, and decarburization time.

7. A real-time prediction method for the carbon content of molten steel in RH refining based on machine learning according to claim 1, characterized in that, In step S3, if there is a delay in the generation of on-site flue gas data, the real-time prediction result of the molten steel carbon content should consider the delay time.

8. A real-time prediction system for the carbon content of molten steel in RH refining based on machine learning, which is used to implement the prediction method described in any one of claims 1-7, includes: Model database construction module: Collect the production data of historical RH refining heats, screen and preprocess abnormal heats, and process the flue gas time-series data of each historical heat to construct a model database for modeling; Prediction model construction module: Determine the input variables and output variables of the model, and construct a prediction model for the actual decarburization amount within each time step through machine learning algorithms; Real-time prediction module for the carbon content of molten steel: According to the collected real-time heat production data, the model outputs the prediction results of the actual decarburization amount within each time step to achieve real-time prediction of the carbon content of molten steel.

Images