Application method of data model in prediction of converter steelmaking static oxygen supply

Abstract

The invention discloses an application method of a data model in prediction of converter steelmaking static oxygen supply. The method comprises the following steps: S1, steelmaking converter whole-process data preprocessing; s2, making a data model; s3, deploying the data model online; and S4, the converter L2 process control online system calls the data model online. A data set covering the whole converter steelmaking process is manufactured, a data model is trained on the basis of the data set, the trained data model is deployed on a reasoning server in a containerization mode, a converter L2 system transmits heat characteristic data through an API interface, then an oxygen amount calculation model is remotely called, and on the basis of achieving static oxygen supply amount prediction, the oxygen supply amount of the converter is predicted. And the purpose of improving the prediction precision is achieved.

Description

Technical Field

[0001] This invention relates to converter production operation control technology, and more specifically, to a method for using a data model to predict the static oxygen supply in converter steelmaking. Background Technology

[0002] Before the converter begins blowing, the predetermined oxygen supply for that heat must be calculated based on the charge amount, main raw material ratio, target temperature, and target composition. This oxygen supply is a crucial parameter for controlling the converter steelmaking process. Automated controls such as oxygen lance position, oxygen flow rate, bottom blowing gas type switching, auxiliary lance measurement, and auxiliary raw material weighing are all based on the oxygen supply, calculating the setpoints for each control variable at each stage of blowing and then distributing them to the basic automation L1 layer for execution. Therefore, accurately calculating the oxygen supply for each heat is of great significance for the overall converter steelmaking process control.

[0003] Currently, steel companies generally calculate the predetermined oxygen supply for each heat in steelmaking plants by referring to the production performance of similar heats. However, in practical applications, this is constrained by issues such as abnormal performance of the reference heats, resulting in low calculation accuracy. This indirectly leads to poor accuracy of various set control values, increasing the risk of overblowing in converters. Therefore, it is necessary to establish a more accurate converter blowing oxygen supply model.

[0004] Existing patent applications, such as patent application number 202010558406.7, disclose a smelting method for improving the oxygen supply intensity of a converter in stages, including the following steps: calculating the oxygen blowing amount based on the composition of the molten iron and scrap steel in the converter and the amount of charge; starting the converter and smelting with the oxygen lance, adjusting the oxygen lance position in stages according to the calculated oxygen blowing percentage, and simultaneously adjusting the oxygen supply flow rate in stages according to the oxygen blowing percentage; adding auxiliary raw materials for converter smelting in stages according to the oxygen blowing percentage; and ending the smelting process by lifting the lance when the oxygen supply reaches 100%.

[0005] For example, patent application number CN202211163887.7 relates to a method and system for predicting oxygen blowing volume in converter steelmaking. It establishes a converter oxygen blowing volume prediction model by combining a mechanistic model and a statistical model. The part that can be calculated relatively accurately is calculated using a mechanistic model established by oxygen balance, while the remaining oxygen consumption is calculated using a statistical model. Furthermore, by using cluster analysis, the production data is divided into N categories and calculated separately, which effectively improves the prediction accuracy of oxygen blowing volume in converter steelmaking.

[0006] For example, patent application number CN202310812854.9 relates to a method, apparatus, medium, and electronic device for determining the static oxygen content of a converter. The method includes: accessing a reference converter furnace database and calculating the differences in production parameters between each reference converter furnace and a target converter furnace; determining at least one calculation reference furnace based on the differences in production parameters and obtaining the production parameters of each calculation reference furnace; obtaining calculation parameters for static oxygen content; calculating the reference weight of each calculation reference furnace and the static oxygen content of at least one target converter furnace based on the production parameters of each calculation reference furnace and the calculation parameters; and determining the target furnace static oxygen content of the target converter furnace based on the reference weights of each calculation reference furnace and the static oxygen content of each furnace.

[0007] For example, patent application number CN202310272827.7 discloses a method for predicting converter oxygen supply based on segmented oxygen decarburization efficiency, including: using a case reasoning algorithm to find the previous case with the highest similarity to the heat to be solved in the previous case library according to the molten iron conditions and endpoint control target of the heat to be solved, as a reference case; taking into account the different influencing factors of oxygen decarburization efficiency at different stages of converter blowing, dividing the blowing process of the reference case into stages, namely the early blowing stage, the middle blowing stage, and the late blowing stage; based on the preset assumptions, predicting the oxygen supply of the heat to be solved according to the blowing stage division time and the oxygen decarburization efficiency at different stages of the reference case, and obtaining the oxygen supply prediction result of the heat to be solved.

[0008] However, none of the above four patented technologies involve AI model control technology. The methods involved focus on mechanism models and case reasoning, which may be constrained by problems such as abnormal performance of reference furnaces in practical applications, thus affecting the accuracy of prediction. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide a method for using a data model to predict the static oxygen supply in converter steelmaking. This method involves creating a dataset covering the entire converter steelmaking process, training a data model based on this dataset, and then containerizing and deploying the trained data model on an inference server. The converter L2 system can remotely call the oxygen calculation model by inputting furnace feature data through an API interface. This achieves the goal of improving prediction accuracy while realizing the prediction of static oxygen supply.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A method for applying a data model to predict static oxygen supply in converter steelmaking includes the following steps:

[0012] S1, preprocessing of data for the entire process of steelmaking converter;

[0013] S2, Create a data model;

[0014] S3, online deployment of data models;

[0015] S4, the online system for the L2 process control of the converter calls the data model online.

[0016] Preferably, step S1 specifically includes the following process:

[0017] S11, remove data containing null values, or add null values ​​according to business logic;

[0018] S12 uses LabelEncoder to convert the steel tapping mark into a numerical variable;

[0019] S13, Perform statistical analysis on each numerical variable and observe the distribution of each numerical variable based on the histogram;

[0020] S14. The correlation between the characteristic variables and the oxygen supply is analyzed using mutual information, and some weakly correlated characteristics are eliminated.

[0021] S15. The five-number summary method and the isolated forest algorithm are used to find outliers in the data. After exporting the outliers, it is confirmed whether the data needs to be removed based on the process knowledge.

[0022] The better option is S16, standardization and normalization.

[0023] The preferred one is S17, a characteristic structure.

[0024] Preferably, step S2 specifically includes the following process:

[0025] S21. Capture oxygen content characteristic data from a steelmaking database to form a static oxygen content prediction dataset, and process the dataset according to step S1.

[0026] S22, the dataset is divided into training and test sets in a 4:1 ratio;

[0027] S23 uses the Catboost ensemble learning model to fit the data on the training set;

[0028] S24. The hyperparameters of the Catboost ensemble learning model are optimized using a Bayesian optimizer to find the hyperparameters that are most suitable for the training set, thus obtaining the Bo-Catboost model.

[0029] S25. The Bo-Catboost model is tested using the test set data. The test results are visualized, and scatter plots and error analysis plots of the prediction results are drawn.

[0030] S26, Evaluate the performance of the Bo-Catboost model, and calculate the R-value of the model on the test set. 2 RMSE, MAE and prediction errors are within ±300 Nm 3 With ±500Nm 3 Hit rate within the range.

[0031] Preferably, step S3 specifically includes:

[0032] On the model development machine, the model files, model dependencies, and Flask service framework are packaged into a container image and migrated to the inference server for execution. The online server of the converter L2 process control online system reads the feature data of the converter database, calls the API interface of the inference server, obtains the prediction results returned by the inference server, and displays the data model prediction results on the screen terminal of the converter L2 process control online system.

[0033] Preferably, step S4 specifically includes:

[0034] S41, the online process control system of converter L2 obtains the characteristic variables required for oxygen content prediction based on the steel grade HEATNO of the furnace and stores the values ​​in the map dictionary;

[0035] S42, the online process control system for converter L2 uses the C++ nlohmann library to convert the map dictionary into a JSON string, and then transmits the JSON string to the inference server via HTTP protocol using the POST method;

[0036] S43, the inference server parses the received JSON string into a DataFrame;

[0037] S44, the inference server determines whether the DataFrame meets the input conditions of the model. If it does, proceed to step S45. If it does not, return an error code to the converter L2 process control online system and stop the prediction.

[0038] S45, using the pre-loaded model file to predict the DataFrame data frame, calculate the predicted value of static oxygen supply;

[0039] S46, the inference server returns the predicted value of static oxygen supply to the converter L2 process control online system, and displays it on the screen terminal of the converter L2 process control online system.

[0040] The present invention provides a method for predicting the static oxygen supply of converter steelmaking using a data model, achieving accurate prediction of the static oxygen supply of the converter with a prediction deviation of ±300 Nm. 3With a hit rate exceeding 80%, it stably controls the quality of molten steel and reduces the risk of converter over-blowing, improving the control level of converter steelmaking. It is an important component in realizing intelligent manufacturing and one-click steelmaking. It also has the following beneficial effects:

[0041] 1) High prediction accuracy. The oxygen supply model has a prediction error of ±300 Nm. 3 With ±500Nm 3 The hit rates were 80.93% and 95.02%, respectively, which are higher than those of the traditional mechanism plus statistical model;

[0042] 2) Fast reasoning speed. The online system for converter L2 process control remotely calls the oxygen calculation model after receiving furnace characteristic data through the API interface. The response time of the entire process is within 10 milliseconds. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the application method of the present invention;

[0044] Figure 2 This is a schematic diagram illustrating the source of feature data in step S1 of the application method of the present invention;

[0045] Figure 3 This is a flowchart illustrating step S2 in the application method of the present invention;

[0046] Figure 4 This is a schematic diagram of step S3 in the application method of the present invention;

[0047] Figure 5 This is a flowchart illustrating step S4 in the application method of the present invention;

[0048] Figure 6 This is a schematic diagram of the prediction results displayed on the screen in an embodiment of the application method of the present invention;

[0049] Figure 7 This is a schematic diagram comparing the predicted oxygen supply with the actual oxygen supply in an embodiment of the application method of the present invention. Detailed Implementation

[0050] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0051] Combination Figure 1 As shown, the present invention provides a method for using a data model to predict the static oxygen supply in converter steelmaking, comprising the following steps:

[0052] S1, Steelmaking Converter Full-Process Data Preprocessing

[0053] Raw feature data was collected from the converter database, and feature parameters from different tables were linked using the steel grade as the primary key. Some data sources are as follows: Figure 2 As shown.

[0054] In the steelmaking process control system, due to operator errors, equipment maintenance, network connection failures, and other reasons, some null and outlier values ​​may be generated. These values ​​will affect the accuracy of model training and require data preprocessing. The data preprocessing process in step S1 specifically includes the following steps:

[0055] S11, remove a small amount of data containing null values, or add null values ​​according to specific rules based on business logic;

[0056] S12. Since the steel tapping mark variable is a string, the model cannot process it directly. The LabelEncoder encoder is used to convert the steel tapping mark into a numerical variable.

[0057] S13, Perform statistical analysis on each numerical variable and observe the distribution of each numerical variable based on the histogram;

[0058] S14. The correlation between the characteristic variables and the oxygen supply is analyzed using mutual information, and some weakly correlated characteristics are eliminated.

[0059] S15. Use the five-number summary method and the isolated forest algorithm to find outliers in the data. After exporting the outliers, confirm whether the data needs to be removed based on the process knowledge.

[0060] S16, Standardization and Normalization. The models selected in this embodiment are all tree models based on gradient boosting decision trees. Numerical scaling does not affect the shape of the tree, so there is no need to perform standardization and normalization on the data.

[0061] S17, Feature Construction. The distributions of variables such as molten iron temperature, scrap steel quantity, and molten iron carbon content approximate a normal distribution and are highly correlated with oxygen supply. Addition, division, and multiplication operations were performed on these variables to construct new features.

[0062] S2, create a data model, and combine... Figure 3 As shown, the specific process includes the following:

[0063] S21. Capture oxygen content characteristic data from a steelmaking database to form a static oxygen content prediction dataset, and process the dataset according to step S1.

[0064] S22, the dataset is divided into training and test sets in a 4:1 ratio;

[0065] S23 uses the Catboost ensemble learning model to fit the data on the training set;

[0066] S24. The hyperparameters of the Catboost ensemble learning model are optimized using a Bayesian optimizer to find the hyperparameters that are most suitable for the training set, thus obtaining the Bo-Catboost model.

[0067] S25. The Bo-Catboost model is tested using the test set data. The test results are visualized, and scatter plots and error analysis plots of the prediction results are drawn.

[0068] S26. Evaluate the performance of the o-Catboost model, calculating the model's R², RMSE, MAE, and prediction error within ±300 Nm on the test set. 3 With ±500Nm 3 Hit rate within the range.

[0069] In step S2, the model selection is as follows:

[0070] Three models—Random Forest, LightGBM, and Catboost—were selected to evaluate their performance. See Table 1 below for details. The determination coefficient R of the Catboost model is shown. 2 The determination coefficient R0 of the LightGBM model is 0.87. 2 The coefficient of determination R of the random forest model is 0.82. 2 With a coefficient of 0.71, the Catboost model fits the data better than the other two models. 2 It is closer to 1. Furthermore, compared to the other two models, the Catboost model has lower RMSE and MAE, and its predicted values ​​are closer to the actual values, achieving the highest prediction accuracy. Among the three models, the Catboost model also has the highest hit rate, with an error margin of ±300 Nm. 3 With ±500Nm 3 The hit rates were 80.93% and 95.02%, respectively.

[0071] Table 1 Comparison of accuracy of the three models

[0072]

[0073] The Catboost model uses a tree growth algorithm and converges quickly. However, without appropriate hyperparameters, it can grow too fast and cause overfitting. Therefore, hyperparameter optimization is needed to obtain better model performance. A Bayesian optimizer is introduced to search for the best combination of hyperparameters for the Catboost model. The optimization search range is shown in Table 2 below.

[0074] Table 2. Hyperparameter Optimization Settings for the Catboost Model

[0075]

[0076] S3, online deployment of data models

[0077] Combination Figure 4 As shown, the model files, model dependencies, and Flask service framework are packaged into a container image 2 on the model development machine 1 and migrated to the inference server 3 for operation. The online server 4 of the converter L2 process control online system reads the feature data of the converter database 5, calls the API interface of the inference server, obtains the prediction results returned by the inference server, and displays the data model prediction results on the screen terminal 6 of the converter L2 process control online system.

[0078] S4, the online system for converter L2 process control calls the data model online, combined with Figure 5 As shown, the specific steps include:

[0079] S41, the online process control system of converter L2 obtains the characteristic variables required for oxygen content prediction based on the steel grade HEATNO of the furnace and stores the values ​​in the map dictionary;

[0080] S42, the online process control system for converter L2 uses the C++ nlohmann library to convert the map dictionary into a JSON string, and then transmits the JSON string to the inference server via HTTP protocol using the POST method;

[0081] S43, the inference server parses the received JSON string into a DataFrame;

[0082] S44, the inference server determines whether the DataFrame meets the input conditions of the model. If it does, proceed to step S45. If it does not, return an error code to the converter L2 process control online system and stop the prediction.

[0083] S45, using the pre-loaded model file to predict the DataFrame data frame, calculate the predicted value of static oxygen supply;

[0084] S46, the inference server returns the predicted value of static oxygen supply to the converter L2 process control online system, which displays it on the screen terminal of the converter L2 process control online system for reference by steel plant operators.

[0085] Example

[0086] A steel plant at a major steel enterprise implemented a data model from this embodiment for predicting static oxygen supply in converter steelmaking, achieving good results. The predicted oxygen supply deviation for the first steelmaking unit was within ±300 Nm³. 3The hit rate within the model remains consistently above 80%. The following description, in conjunction with accompanying figures and embodiments, further describes the application method of a data model in predicting the static oxygen supply for converter steelmaking.

[0087] See again Figure 5 As shown, the static oxygen supply for heat 115186 of steel is predicted, and the specific steps are as follows:

[0088] 1) The online process control system for converter L2 obtains the characteristic variable information of the heat number 115186. Some characteristics are shown in Table 3 and are stored in a map dictionary.

[0089] Table 3. Partial characteristic variables of steel grade 115186.

[0090]

[0091] 2) The online system for converter L2 process control converts the map dictionary into a JSON string format and sends the JSON string to the inference server via HTTP protocol using the POST method.

[0092] 3) After receiving the JSON string containing information on 115,186 furnace cycles, the inference server parses it into a DataFrame format.

[0093] 4) The inference server verifies that all fields of the DataFrame meet the input conditions of the model.

[0094] 5) The inference server loads the trained Bo-catboost model file, performs predictions, and obtains a predicted static oxygen supply value of 13420 Nm³. 3 .

[0095] 6) The L2 online system receives the inference results returned by the inference server and displays them on the L2 operation screen.

[0096] The final screen displays as follows Figure 6 As shown, the actual oxygen blowing and shut-off amount for steel batch 115186 was 13410 Nm. 3 The AI ​​model predicts an oxygen supply of 13420 Nm³. 3 The error is 10 Nm 3 .

[0097] Taking a random sample of 500 heats from a steelmaking plant at a certain steel plant's base as an example, this paper tests an implementation method of a data model for predicting static oxygen supply in converter steelmaking. The predicted oxygen supply is compared with the actual oxygen supply. Figure 6 As shown, in these 500 furnaces, the oxygen content prediction error is within ±300 Nm. 3There were a total of 412 furnaces tested, with a hit rate of 82.4%.

[0098] In summary, the data model of this invention can be further extended to predict the static oxygen supply of converter steelmaking, both domestically and internationally. This can improve the automation level of converters, stabilize the quality of molten steel, and reduce the risk of over-blowing in converters, with broad prospects for application.

[0099] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for applying a data model to predict the static oxygen supply in converter steelmaking, characterized in that, Includes the following steps: S1, preprocessing of data for the entire process of steelmaking converter; S2, Create a data model; S3, online deployment of data models; S4, the online system for the L2 process control of the converter calls the data model online.

2. The method for applying the data model according to claim 1 to predict the static oxygen supply in converter steelmaking, characterized in that, Step S1 specifically includes the following process: S11, remove data containing null values, or add null values ​​according to business logic; S12 uses LabelEncoder to convert the steel tapping mark into a numerical variable; S13, Perform statistical analysis on each numerical variable and observe the distribution of each numerical variable based on the histogram; S14. The correlation between the characteristic variables and the oxygen supply is analyzed using mutual information, and some weakly correlated characteristics are eliminated. S15. The five-number summary method and the isolated forest algorithm are used to find outliers in the data. After exporting the outliers, it is confirmed whether the data needs to be removed based on the process knowledge.

3. The method for applying the data model according to claim 2 to predict the static oxygen supply in converter steelmaking, characterized in that, Step S1 further includes: S16, Standardization and Normalization.

4. The method for applying the data model according to claim 3 to predict the static oxygen supply in converter steelmaking, characterized in that, Step S1 further includes: S17, Feature Construction.

5. The method for applying the data model according to claim 2 to predict the static oxygen supply in converter steelmaking, characterized in that, Step S2 specifically includes the following process: S21. Capture oxygen content characteristic data from a steelmaking database to form a static oxygen content prediction dataset, and process the dataset according to step S1. S22, the dataset is divided into training and test sets in a 4:1 ratio; S23 uses the Catboost ensemble learning model to fit the data on the training set; S24. The hyperparameters of the Catboost ensemble learning model are optimized using a Bayesian optimizer to find the hyperparameters that are most suitable for the training set, thus obtaining the Bo-Catboost model. S25. The Bo-Catboost model is tested using the test set data. The test results are visualized, and scatter plots and error analysis plots of the prediction results are drawn. S26. Evaluate the performance of the o-Catboost model, calculating the model's R², RMSE, MAE, and prediction error within ±300 Nm on the test set. 3 With ±500Nm 3 Hit rate within the range.

6. The method for applying the data model according to claim 2 to predict the static oxygen supply in converter steelmaking, characterized in that, Step S3 specifically includes: On the model development machine, the model files, model dependencies, and Flask service framework are packaged into a container image and migrated to the inference server for execution. The online server of the converter L2 process control online system reads the feature data of the converter database, calls the API interface of the inference server, obtains the prediction results returned by the inference server, and displays the data model prediction results on the screen terminal of the converter L2 process control online system.

7. The method for applying the data model according to claim 6 to predict the static oxygen supply in converter steelmaking, characterized in that, Step S4 specifically includes: S41, the online process control system of converter L2 obtains the characteristic variables required for oxygen content prediction based on the steel grade HEATNO of the furnace and stores the values ​​in the map dictionary; S42, the online process control system for converter L2 uses the C++ nlohmann library to convert the map dictionary into a JSON string, and then transmits the JSON string to the inference server via HTTP protocol using the POST method; S43, the inference server parses the received JSON string into a DataFrame; S44, the inference server determines whether the DataFrame meets the input conditions of the model. If it does, proceed to step S45. If it does not, return an error code to the converter L2 process control online system and stop the prediction. S45, using the pre-loaded model file to predict the DataFrame data frame, calculate the predicted value of static oxygen supply; S46, the inference server returns the predicted value of static oxygen supply to the converter L2 process control online system, and displays it on the screen terminal of the converter L2 process control online system.

Images