A kmeans-bo-rf-based rh refining molten steel endpoint alloy composition prediction method and system

Abstract

The present application belongs to the technical field of steel metallurgy, and specifically relates to a RH refining molten steel end-point alloy composition prediction method and system based on Kmeans-BO-RF, which realizes fine division of production furnace data according to RH refining treatment modes and process targets and combined clustering analysis of a Kmeans algorithm, adopts a random forest (RF) algorithm for modeling, and iteratively optimizes hyperparameters in the random forest (RF) algorithm through a Bayesian optimization (BO) algorithm to fully exert the maximum prediction effect of the random forest (RF) algorithm, and finally realize high-precision prediction of RH refining molten steel end-point alloy composition under multiple working conditions.

Description

Technical Field

[0001] This invention relates to the field of iron and steel metallurgy technology, specifically to a method and system for predicting the final alloy composition of RH refining molten steel based on Kmeans-BO-RF. Background Technology

[0002] RH refining is a key process in the smelting of ultra-low carbon steel. Its main task is to achieve deep decarburization of molten steel while simultaneously ensuring that the alloy element content of the molten steel meets the required composition at the outlet through precise alloying, thereby ensuring that the steel quality meets the expected standards. To achieve precise alloying, it is necessary to accurately predict the relationship between the content of each alloy element at the endpoint and each smelting operation. This ensures that while meeting the steel composition requirements, operations such as minimizing the amount of alloy material added and shortening the smelting time are minimized, which helps to reduce production costs and environmental impact.

[0003] For the alloying problem in RH refining, several scholars have conducted research on different modeling methods. These include: using multiple linear regression on process parameters that may affect the content of acid-soluble aluminum in historical data to obtain the relationship between each factor and the acid-soluble aluminum content; using the random forest algorithm to determine the contribution of each process parameter to the alloy yield, and then using a weighted Euclidean distance-based real-time learning method to find similar furnaces to predict the yield of each alloying element; establishing a dynamic library of alloy yields to address the problem of alloy yield variation during RH refining, followed by optimizing the model using linear regression and nonlinear programming to achieve accurate prediction and cost control; and research on four forms of manganese loss during RH vacuum smelting, finding that the manganese yield during RH smelting can be improved by reducing the RH vacuum level, the manganese content and temperature at the inlet, and reducing the RH vacuum treatment time. A manganese alloying model with a hit rate close to 100% was constructed, which can control the manganese content error at the end of RH to ±0.01% of the target manganese mass fraction. The aforementioned studies mostly aim to achieve their goals by constructing regression models between various data points during the operation and the content of alloying elements at the endpoint, based on the analysis of metallurgical mechanisms. However, current models for the content of alloying elements at the smelting endpoint face the following unresolved issues: 1) The data composition is extensive, and the same regression function is prone to local overfitting, making it difficult to reflect the global situation; 2) The relationship between each feature value and the prediction target is complex, making it difficult to express using simple multivariate nonlinear regression; 3) Intelligent algorithms often use multiple different types of parameters to optimize the algorithm, and existing models rarely solve for the optimal combination of parameters. Summary of the Invention

[0004] To address the problems existing in the prior art, the main objective of this invention is to propose a method and system for predicting the alloy composition at the endpoint of RH refining molten steel based on Kmeans-BO-RF. By combining Kmeans clustering analysis, random forest (RF), and Bayesian optimization (BO), the method achieves higher accuracy in predicting the alloy element content at the endpoint of RH refining.

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

[0006] A method for predicting the final alloy composition of RH-refined molten steel based on Kmeans-BO-RF includes the following steps:

[0007] S1. Collect historical furnace data of RH refining, and divide the historical furnace data into N datasets based on the processing modes such as light treatment, decarburization treatment, and basic treatment, as well as the process objectives such as degassing, composition fine-tuning, and temperature adjustment, and combine them with the Kmeans algorithm for cluster analysis.

[0008] S2. Based on the functions of different alloying elements, especially aluminum, silicon and other alloying elements which have both deoxidation and alloying functions, determine the parameter variables that affect the alloy composition at the RH refining endpoint, and use them as model input variables.

[0009] S3. For each dataset, the Random Forest (RF) algorithm is used to model the model. The hyperparameters in the Random Forest (RF) algorithm are iteratively optimized using the Bayesian Optimization (BO) algorithm. The candidate models constructed with different hyperparameters are compared, and the candidate model with the closest fitting coefficient to 1 is selected as the RH refining steel endpoint alloy composition prediction model suitable for this dataset.

[0010] S4. Collect real-time heat data of RH refining, determine the processing mode and process target of the heat, and divide the real-time heat data into the dataset with the highest similarity according to the distance similarity analysis of the model input variables. Substitute the data into the prediction model applicable to the dataset and output the prediction result of the alloy composition of the molten steel at the end of the real-time heat.

[0011] According to another aspect of the present invention, the present invention provides the following technical solution:

[0012] A Kmeans-BO-RF-based system for predicting the final alloy composition of RH-refined molten steel includes:

[0013] Dataset partitioning module: Collect historical furnace data of RH refining, and divide the historical furnace data into N datasets based on the processing mode and process objectives, combined with Kmeans algorithm for cluster analysis.

[0014] Model input variable determination module: Determines the parameter variables that affect the final alloy composition of RH refined steel as model input variables;

[0015] Prediction model building module: For each dataset, a prediction model for the final alloy composition of RH refined steel is constructed by coupling random forest and Bayesian optimization algorithms.

[0016] Endpoint alloy composition prediction module: Collects real-time heat data of RH refining, determines the processing mode and process target of the heat, inputs it into the corresponding prediction model, and outputs the real-time heat steel endpoint alloy composition prediction results.

[0017] The beneficial effects of this invention are as follows:

[0018] This invention proposes a method and system for predicting the final alloy composition of RH refining steel based on Kmeans-BO-RF. According to the RH refining treatment mode and process objectives, and combined with Kmeans algorithm for cluster analysis, it achieves fine division of production furnace data. Random Forest (RF) algorithm is used for modeling, and the hyperparameters in the Random Forest (RF) algorithm are iteratively optimized using Bayesian Optimization (BO) algorithm to fully leverage the maximum predictive effect of the Random Forest (RF) algorithm, ultimately achieving high-precision prediction of the final alloy composition of RH refining steel under multiple operating conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the prediction method of the present invention.

[0021] Figure 2 This is a schematic diagram of the random forest prediction model in an embodiment of the present invention.

[0022] Figure 3 It is the sum of squared errors of each group in the embodiments of the present invention.

[0023] Figure 4 This is a schematic diagram of the Bayesian optimization process according to an embodiment of the present invention.

[0024] Figure 5 The accuracy of different models in predicting the acid-soluble aluminum content at the RH refining endpoint.

[0025] Figure 6 The accuracy of different models in predicting silicon content at the RH refining endpoint.

[0026] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] The technical solutions described below in conjunction with the embodiments will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention proposes a method and system for predicting the final alloy composition of RH refining molten steel based on Kmeans-BO-RF. According to the RH refining treatment mode and process objectives, and combined with Kmeans algorithm for cluster analysis, the system achieves fine division of production furnace data. By combining Kmeans cluster analysis, random forest (RF) and Bayesian optimization (BO) method, the system achieves higher accuracy in predicting the alloy element content at the final RH refining point.

[0029] According to one aspect of the present invention, the present invention provides the following technical solution:

[0030] like Figure 1 As shown, a method for predicting the final alloy composition of RH-refined molten steel based on Kmeans-BO-RF includes the following steps:

[0031] S1. Collect historical furnace data of RH refining, and divide the historical furnace data into N datasets based on the processing modes such as light treatment, decarburization treatment, and basic treatment, as well as the process objectives such as degassing, composition fine-tuning, and temperature adjustment, and combine them with the Kmeans algorithm for cluster analysis.

[0032] S2. Based on the functions of different alloying elements, especially aluminum, silicon and other alloying elements which have both deoxidation and alloying functions, determine the parameter variables that affect the alloy composition at the RH refining endpoint, and use them as model input variables.

[0033] S3. For each dataset, the Random Forest (RF) algorithm is used to model the model. The hyperparameters in the Random Forest (RF) algorithm are iteratively optimized using the Bayesian Optimization (BO) algorithm. The candidate models constructed with different hyperparameters are compared, and the candidate model with the closest fitting coefficient to 1 is selected as the RH refining steel endpoint alloy composition prediction model suitable for this dataset.

[0034] S4. Collect real-time heat data of RH refining, determine the processing mode and process target of the heat, and divide the real-time heat data into the dataset with the highest similarity according to the distance similarity analysis of the model input variables. Substitute the data into the prediction model applicable to the dataset and output the prediction result of the alloy composition of the molten steel at the end of the real-time heat.

[0035] Preferably, in step S1, RH refining can be divided into degassing light treatment, decarburization light treatment, temperature adjustment light treatment, composition fine-tuning light treatment, conventional vacuum decarburization treatment, oxygen blowing enhanced decarburization treatment, deep degassing treatment, deep decarburization treatment, and composition precise control treatment according to the processing mode and process objectives.

[0036] Preferably, in step S1, the K-means algorithm is used for cluster analysis to divide the historical furnace data into N datasets. The mean of the data in each dataset is used as the cluster center. The goal is to minimize the sum of the distances between the data points in the dataset and their respective cluster centers.

[0037] Preferably, in step S2, based on the functions of different alloying elements, and combined with metallurgical mechanisms and data correlation analysis methods, the correlation between the parameter variables of the RH refining process and the final alloy composition of the molten steel is analyzed to determine the input variables of the model.

[0038] Preferably, in step S3, the prediction model constructed using the random forest algorithm is expressed as follows:

[0039]

[0040] In the formula, x represents the influencing factor, {θ} n Let} be the nth regression tree, and f(x,{θ} be the regression tree. n}) represents the prediction result, f n (x,{θ n}) represents the preliminary result for the nth tree, where N is the number of trees.

[0041] Preferably, in step S3, within the hyperparameter search space χ, the Bayesian optimization hyperparameters are expressed as:

[0042] x * =argmin x∈χ f(x)

[0043] In the formula, x* is the optimal combination of hyperparameters; f(x) is the objective optimization function, which is also the acquisition function in the Bayesian optimization algorithm.

[0044] Preferably, in step S3, the hyperparameters of the random forest algorithm include the number of decision trees, the maximum number of tree depths, the minimum number of sample split trees, and the minimum number of sample leaves.

[0045] Preferably, in step S3, the formula for calculating the model fitting coefficient is:

[0046]

[0047] In the formula, R 2y is the model fit coefficient; m is the number of test samples; i This represents the true value of the i-th sample. Let i be the predicted values ​​for i samples; This is the average of the true values.

[0048] According to another aspect of the present invention, the present invention provides the following technical solution:

[0049] A Kmeans-BO-RF-based system for predicting the final alloy composition of RH-refined molten steel includes:

[0050] Dataset partitioning module: Collect historical furnace data of RH refining, and divide the historical furnace data into N datasets based on the processing mode and process objectives, combined with Kmeans algorithm for cluster analysis.

[0051] Model input variable determination module: Determines the parameter variables that affect the final alloy composition of RH refined steel as model input variables;

[0052] Prediction model building module: For each dataset, a prediction model for the final alloy composition of RH refined steel is constructed by coupling random forest and Bayesian optimization algorithms.

[0053] Endpoint alloy composition prediction module: Collects real-time heat data of RH refining, determines the processing mode and process target of the heat, inputs it into the corresponding prediction model, and outputs the real-time heat steel endpoint alloy composition prediction results.

[0054] Example

[0055] Taking the content of Als and Si elements in the acid-soluble aluminum at the RH refining endpoint of JBXS30 steel in a certain steel plant as an example, the requirements for the composition of the molten steel leaving the station are shown in Table 1.

[0056] Table 1. Composition Requirements for JBXS30 Steel RH Refining and Finishing Process

[0057] Exit requirements / % ≤0.0020 0.18-0.25 ≤0.004 ≤0.004 0.95-0.98 0.225-0.3 ≤0.0015

[0058] The alloying operation for this steel grade is carried out after decarburization. After confirming the amount of alloy to be added on-site, the alloy material is weighed from the silo, conveyed to the hopper for storage, and then added to the molten steel in a prescribed order. The order of alloy addition is as follows: first, aluminum shot is added for deoxidation, while controlling the acid-soluble aluminum content in the molten steel; then, high-purity ferrosilicon, metallic manganese, and desulfurizing agents are added to ensure that the content of other elements meets the requirements for leaving the station.

[0059] For the prediction of the final alloy element content, based on field experience, the final acid-soluble aluminum content and the final silicon content are mainly affected by the following characteristic conditions:

[0060] X1 Vacuum tank age, X2 Slag thickness, X3 Oxygen concentration at inlet, X4 Inlet temperature, X5 Total oxygen blowing, X6 [O] before alloying, X7 Outlet temperature, X8 Smelting time after decarburization, X9 Amount of metallic manganese added, X 10 High-purity ferrosilicon addition amount, X 11 Aluminum particle addition amount, X 12 molten steel quantity, X 13 C content and X content upon entry 14 S content and X content upon entry 15 Incoming Si content, X 16 Mn content and X at the station 17 The content of acid-soluble aluminum entering the station. X1 and X2 represent the overall operating status of the ladle; X3 to X7 represent the impact of oxygen-temperature changes in the molten steel on aluminum deoxidation and heating processes, also reflecting some operations in the smelting process; X8 represents the smelting status after decarburization; X9 to X... 11 This represents the addition of the main alloying materials, X 12 ~X 17 This represents the composition and condition of the molten steel entering the station.

[0061] By processing the production data of 1051 heats of JBXS30 steel smelted by a steel plant, and removing outliers, 950 heats remained. Multiple linear regression analysis was used to analyze the impact of each eigenvalue on the endpoint content of the target element, yielding the following results:

[0062]

[0063]

[0064] The prediction model constructed using the random forest algorithm is expressed as follows:

[0065]

[0066] In the formula, x represents the influencing factor, {θ} n Let} be the nth regression tree, and f(x,{θ} be the regression tree. n}) represents the prediction result, f n (x,{θ n}) represents the preliminary result for the nth tree, where N is the number of trees.

[0067] This prediction model uses 800 furnaces as the training set and utilizes on-site production characteristic values ​​X1 to X2. 17 As input, the number of decision trees is 150, the maximum tree depth is 30, the minimum number of splits is 5, and the minimum number of leaves is 2. The random forest structure of the endpoint element content prediction model is as follows. Figure 2 As shown.

[0068] Cluster analysis of smelting furnaces allows for local modeling of target furnaces, improving model accuracy. Based on the characteristics of K-means clustering analysis, the only parameter that needs adjustment is the K value, i.e., the total number of clusters. This embodiment uses the elbow method to determine the K value. A key criterion for the elbow method is the sum of squared errors (SSE), which reflects the clustering effect of a set of samples. As the K value increases, the SSE value decreases, indicating that the elbow method's advantage lies in its ability to perform more in-depth analysis of samples, thus achieving the best clustering effect. If the K value is lower than the expected clustering value, the SSE value will significantly increase. However, if the K value is higher than the expected clustering value, the SSE value will drop sharply. Furthermore, as the K value increases further, the SSE value will show a gradual upward trend. Therefore, the k value corresponding to the elbow in the image represents the number of clusters in the data. The SSE formula is shown below, where μ... (j) This represents the center of cluster j.

[0069]

[0070] Cluster analysis was performed on the databases for the final acid-soluble aluminum content and the final silicon content. Based on field experience, the classification criteria for acid-soluble aluminum were the oxygen content before alloying, the steel liquid composition, the aluminum content, and the aluminum content upon arrival at the treatment plant. The classification criteria for the final silicon content were the amount of high-purity ferrosilicon added, the silicon content upon arrival at the treatment plant, and the steel quality. After normalization, these characteristics were used as clustering features, and the sum of squared errors for clusters with a number of clusters ranging from 2 to 10 was calculated using an enumeration method. Figure 3 As shown, Figure 3 (a) represents the final acid-soluble aluminum content; Figure 3 (b) represents the final silicon content.

[0071] It can be observed that the endpoint of acid-soluble aluminum content shows a clear inflection point when the number of clusters is 2, so the number of clusters is selected as 2; the endpoint of silicon content shows a clear inflection point when the number of clusters is 3, so the number of clusters is selected as 3.

[0072] Within the hyperparameter search space χ, the Bayesian optimization hyperparameters are expressed as:

[0073] x * =argmin x∈χ f(x) (5)

[0074] In the formula, x* is the optimal combination of hyperparameters; f(x) is the objective optimization function, which is also the acquisition function in the Bayesian optimization algorithm.

[0075] The core of the Bayesian optimization algorithm consists of two parts: a prior function and a sampling function. The prior function uses Gaussian process regression to calculate the function mean and covariance at each point, thus obtaining the posterior probability. Then, the sampling function is constructed by refining the probability. This function is used to select the next set of hyperparameters, aiming to find a function value smaller than the current minimum. Commonly used sampling functions include PI, EI, and UCB. Since the BP neural network requires fewer parameters, this scheme adopts the EI strategy based on a boosting strategy.

[0076] The endpoint parameters, trained using feature values ​​from smelting data within different clusters, are optimized using Bayesian methods. The fitting coefficients (R²) obtained by comparing the parameter combinations within the maximum number of iterations are then compared. 2 This process obtains the optimal parameter combination within the given cluster. The Bayesian optimization process is as follows: Figure 4 As shown.

[0077] For different clusters, random forest-based prediction models for the final content of alloy elements were built. The optimal parameter combination was obtained by Bayesian optimization. Test data were selected in each cluster to test and evaluate the accuracy of the model.

[0078] Using the smelting data of JBXS30 steel from a steel company from June to October 2023, a prediction model for the endpoint acid-soluble aluminum and endpoint silicon content based on Kmeans-BO-RF was established after screening. At the same time, prediction models using existing algorithms and RF random forests were established and compared to analyze the advantages of the Kmeans-BO-RF method in terms of prediction accuracy.

[0079] The RH smelting process of this enterprise is as follows: ladle arrival → ladle lifting → vacuuming → oxygen blowing for decarburization → deoxidation and alloying → vacuum breaking → ladle descent and departure. The deoxidation and alloying process involves first adding aluminum shot with an aluminum content greater than 99.9% as a deoxidizer to remove as much oxygen as possible from the molten steel while controlling the acid-soluble aluminum content. Then, high-purity ferrosilicon with a silicon content of 76wt% is added to control the silicon content in the molten steel. This scheme uses 950 heats of smelting data from June to October 2023. After removing obviously abnormal heats, 800 heats were randomly selected as the training set, and the remaining 150 heats were used as the test sets for the endpoint acid-soluble aluminum and endpoint silicon content, respectively.

[0080] Through statistical analysis of actual production data, the distribution of data on the influencing factors of the acid-soluble aluminum content and the silicon content at the RH endpoint was obtained, as shown in Table 2.

[0081] Table 2 Data distribution of the influence of the endpoint content of each element on the characteristic values

[0082]

[0083]

[0084] Comparative analysis of the effects of various prediction methods

[0085] By combining Matlab and Python programming languages, prediction models for the content of acid-soluble aluminum and silicon at the endpoint were established based on multiple linear regression, random forest algorithm and Kmeans-BO-RF method, respectively.

[0086] 150 sets of data each for the endpoint acid-soluble aluminum content and silicon content were selected as test sets. The characteristic values ​​of both were input into three different models to predict the endpoint content of each element. The relative errors between the predicted and actual endpoint contents were statistically analyzed, and the mean and distribution of the prediction errors were determined to characterize the prediction performance of different models under different conditions. The prediction error distributions for the three models and the two element endpoint contents are shown below. Figure 5 , Figure 6 As shown.

[0087] The error distribution of the endpoint content prediction models established by the multiple linear regression (MLR), random forest and Kmeans-BO-RF methods was summarized into the mean error, as shown in Table 3.

[0088] Table 3 Summary of prediction accuracy of alloy element content at the RH smelting endpoint by different models

[0089]

[0090] As shown in Table 3, for the prediction model of the final acid-soluble aluminum content, the error is improved by 10.67% compared with the MLR model and 5.33% compared with the random forest model in the range of [-%1, 1%]; by 20.56% compared with the MLR model and 12.00% compared with the random forest model in the range of [-%2, 2%]; by 19.9% ​​compared with the MLR model and 15.34% compared with the random forest model in the range of [-%3, 3%]; and by 17.23% compared with the MLR model and 8.00% compared with the random forest model in the range of [-%4, 4%], with a mean absolute value of 2.19%.

[0091] For the endpoint silicon content prediction model, the error improvement was 16% compared to the MLR model and 2.00% compared to the random forest model in the range of [-%1, 1%]; 7.99% compared to the MLR model and 6.00% compared to the random forest model in the range of [-%2, 2%]; 9.34% compared to the MLR model and 1.33% compared to the random forest model in the range of [-%3, 3%]; and 6.67% compared to the MLR model and 4.00% compared to the random forest model in the range of [-%4, 4%], with a mean absolute error of 2.05%.

[0092] Therefore, it can be inferred that compared to traditional multiple linear regression models, random forests possess a greater ability to fit multiple nonlinear data for predicting the final content of alloying elements. Furthermore, in actual production, the relationship between various eigenvalues ​​and the final content of alloying elements is not a simple linear one, which to some extent explains why random forest models are more suitable for constructing prediction models for the final content of alloying elements. In actual production, especially for alloying elements such as aluminum that serve both deoxidation and alloying functions, using cluster analysis to classify elements and then training each group with the random forest algorithm can effectively improve model accuracy. In addition, Bayesian optimization can also improve the accuracy of the model for predicting the final alloying element content to some extent. Moreover, during model operation, the training time of the Kmeans-BO-RF model does not exceed 10 minutes, and the average time for predicting the final alloying element content does not exceed 5 seconds, meeting the requirements of on-site operation and providing strong support for precise control of the RH alloying process.

[0093] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for predicting the final alloy composition of RH-refined molten steel based on Kmeans-BO-RF, characterized in that, Includes the following steps: S1. Collect historical furnace data for RH refining. Based on the processing modes including light treatment, decarburization treatment, and basic treatment, and the process objectives including degassing, composition fine-tuning, and temperature adjustment, and using the K-means algorithm for cluster analysis, divide the historical furnace data into N datasets. According to the processing mode and process objectives, RH refining can be divided into degassing light treatment, decarburization light treatment, temperature adjustment light treatment, composition fine-tuning light treatment, conventional vacuum decarburization treatment, oxygen blowing enhanced decarburization treatment, deep degassing basic treatment, deep decarburization basic treatment, and composition precise control basic treatment. S2. Based on the deoxidation and alloying functions of aluminum and silicon alloying elements, and combining metallurgical mechanisms and data correlation analysis methods, the correlation between RH refining process parameters and the final alloy composition of molten steel is analyzed to determine the input variables of the model. The final acid-soluble aluminum content and the final silicon content are affected by the following RH refining process parameters: X1 Vacuum tank age, X2 Slag thickness, X3 Oxygen concentration at the inlet, X4 Inlet temperature, X5 Total oxygen blowing, X6 [O] before alloying, X7 Outlet temperature, X8 Smelting time after decarburization, X9 Amount of metallic manganese added, X 10 High-purity ferrosilicon addition amount, X 11 Aluminum particle addition amount, X 12 molten steel quantity, X 13 C content and X content upon entry 14 S content and X content upon entry 15 Incoming Si content, X 16 Mn content and X at the station 17 The content of acid-soluble aluminum entering the station; where X1 and X2 represent the overall operating status of the ladle, X3~X7 represent the influence of oxygen temperature changes in the molten steel on the aluminum deoxidation and aluminum heating process, and also reflect some operations in the smelting process, X8 represents the smelting status after decarburization, X9~X 11 This represents the addition of the main alloying materials, X 12 ~X 17 This represents the composition and condition of the molten steel entering the station; S3. For each dataset, a random forest algorithm is used to model the model. The hyperparameters in the random forest algorithm are iteratively optimized using the Bayesian optimization algorithm. The candidate models constructed with different hyperparameters are compared, and the candidate model with the closest fitting coefficient to 1 is selected as the RH refining steel endpoint alloy composition prediction model suitable for this dataset. S4. Collect real-time heat data of RH refining, determine the processing mode and process target of the heat, and divide the real-time heat data into the dataset with the highest similarity according to the distance similarity analysis of the model input variables. Substitute the data into the prediction model applicable to the dataset and output the prediction result of the alloy composition of the final steel of the real-time heat. For aluminum and silicon alloying elements that have both deoxidation and alloying functions, using cluster analysis to classify them and then training each group with the random forest algorithm can effectively improve the accuracy of the prediction model for the final alloy composition of RH refined steel. During the model operation, the training time of the prediction model for the final alloy composition of RH refined steel based on Kmeans-BO-RF does not exceed 10 minutes, and the average prediction time for the content of the final alloying elements does not exceed 5 seconds.

2. The prediction method according to claim 1, characterized in that, In step S1, the K-means algorithm is used for cluster analysis to divide the historical furnace data into N datasets. The mean of the data in each dataset is used as the cluster center. The goal is to minimize the sum of the distances between the data points in the dataset and their respective cluster centers.

3. The prediction method according to claim 1, characterized in that, In step S3, the prediction model constructed using the random forest algorithm is expressed as follows: In the formula, x represents the influencing factors. For the nth regression tree, For the predicted results, This is the preliminary result for the nth tree, where N is the number of trees.

4. The prediction method according to claim 1, characterized in that, In step S3, within the hyperparameter search space χ, the Bayesian optimization hyperparameters are expressed as: In the formula, x* is the optimal combination of hyperparameters; f(x) is the objective optimization function, which is also the acquisition function in the Bayesian optimization algorithm.

5. The prediction method according to claim 1, characterized in that, In step S3, the hyperparameters of the random forest algorithm include the number of decision trees, the maximum tree depth, the minimum sample split tree, and the minimum number of sample leaves.

6. The prediction method according to claim 1, characterized in that, In step S3, the formula for calculating the model fitting coefficient is as follows: In the formula, These are the model fitting coefficients; This represents the number of test samples; This represents the true value of the i-th sample. Let i be the predicted values ​​for i samples; This is the average of the true values.

7. A K-means-BO-RF-based system for predicting the final alloy composition of RH-refined molten steel, used to implement the prediction method according to any one of claims 1-6, comprising: Dataset partitioning module: Collect historical furnace data of RH refining, and divide the historical furnace data into N datasets based on the processing mode and process objectives, combined with Kmeans algorithm for cluster analysis. Model input variable determination module: Determines the parameter variables that affect the final alloy composition of RH refined steel as model input variables; Prediction model building module: For each dataset, a prediction model for the final alloy composition of RH refined steel is constructed by coupling random forest and Bayesian optimization algorithms. Endpoint alloy composition prediction module: Collects real-time heat data of RH refining, determines the processing mode and process target of the heat, inputs it into the corresponding prediction model, and outputs the real-time heat steel endpoint alloy composition prediction results.

Images