Converter end point temperature prediction method

Abstract

The invention relates to the technical field of converter steelmaking, in particular to a converter terminal temperature prediction method. According to the method, on the basis of machine learning and deep learning, correlation is created in data feedback in real time by utilizing historical data, and prediction and optimization of the process are realized. In the steel ladle treatment working process in the steel industry, an operator dynamically adjusts continuous casting process parameters by combining professional skills and utilizing model decisions of machine learning and deep learning design, and the stability of steelmaking production and the product quality can be remarkably improved. Converter blowing endpoint temperature and component control is important operation in the later period of converter blowing, accurate temperature forecasting is very important, and in order to improve the carbon and temperature hit rate of the converter blowing endpoint, a characteristic correlation analysis method is used for determining main input variables of a model. The converter final temperature can be accurately and quickly predicted, the quality of molten steel can be improved, and the service life of a furnace lining can be prolonged.

Description

Technical Field

[0001] The invention relates to the technical field of converter steelmaking, in particular to a method for predicting converter endpoint temperature. Background Art

[0002] The determination of the end point of converter blowing is an important link in the later stage of converter blowing. At the end of converter steelmaking blowing, the accuracy of the end point determination is very important. Improving the end point temperature and carbon hit rate of converter blowing can improve the quality of molten steel. On the contrary, inaccurate control of the end point temperature and carbon will extend the smelting time, reduce the life of the furnace lining, increase metal consumption, and affect the quality of steel.

[0003] At present, most domestic converter steelmaking production mainly relies on manual experience, furnace gas analysis measurement method or auxiliary gun detection method to observe and analyze the smelting process and control the smelting end point.

[0004] Since the reaction process in the converter cannot be observed directly, but is based on secondary information such as flame, splashing, and sound, relying on manual experience will result in large errors and high labor intensity.

[0005] The auxiliary gun method can measure directly, and the measured position has good reproducibility. The best indicator that the auxiliary gun method can achieve is a simultaneous hit rate of 80% for carbon and temperature. The control accuracy of carbon is ±0.015%, and the temperature is ±l0℃. Since its development has been relatively mature, it is difficult to improve the control accuracy. At the same time, the development of the auxiliary gun method is restricted by its own defects, which can be mainly attributed to the following aspects: (1) The auxiliary gun is limited by the furnace capacity, and the furnace capacity is preferably above 200t; (2) The auxiliary gun is an intermittent working mode of point measurement, lacking the means of online detection and prediction of manganese and phosphorus components in steel, and it is difficult to meet the closed-loop automatic control process of the converter: (3) It cannot predict the splashing and slag conditions, and cannot provide continuous dynamic information.

[0006] The commonly used online detection system for furnace gas analysis is generally composed of three parts: sampling system, analysis system, and data communication system. Since the characteristics of the sample gas are high temperature, high dust, and rapid changes in composition, and general high-precision analyzers cannot directly analyze this type of sample, in addition to sample gas collection, the sample must also be cooled, dusted, dehumidified, and interfering components filtered out.

[0007] The above methods have large experimental errors and require subsequent calculations. The experimental costs are high, the time is long, and the labor intensity of the operators is high. Summary of the invention

[0008] In order to overcome the shortcomings of the prior art, the present invention provides a method for predicting the end temperature of a converter, which saves time and effort and can accurately and quickly predict the end temperature of the converter.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A method for predicting converter end point temperature, the method specifically comprising the following steps:

[0011] S1: Environment construction:

[0012] Install the data analysis library on your computer;

[0013] S2: Collect the following converter data:

[0014] 1) Composition of molten iron, weight percentage of Si and P in molten iron, %;

[0015] 2) Molten iron temperature, °C;

[0016] 3) Molten iron loading, tons / tank;

[0017] 4) Furnace age, years;

[0018] 5) End point carbon content, wt%;

[0019] S3: Data preprocessing:

[0020] 1) Handling missing values, outliers and duplicate values ​​in data:

[0021] Ⅰ Delete missing values. The deletion process is divided into sample deletion and feature deletion. When the number of samples with missing values ​​in the data set is less than 10% and the missing values ​​are randomly distributed, the rows containing missing values ​​are directly deleted;

[0022] If less than 10% of the samples have missing values, and there is no obvious pattern in the distribution of missing values ​​in the data set, then the samples with missing values ​​are deleted;

[0023] If the missing value of a feature exceeds 80%, delete the feature column;

[0024] Ⅱ Fill in missing values. Filling is divided into mean filling, median filling, mode filling, fixed value filling, regression filling and multiple filling. For numerical features, if the data is normally distributed, the mean is used to fill in missing values. When the data is skewed or has outliers, the median is used to fill in missing values. If fixed business data is missing, it is filled with fixed data.

[0025] III uses methods of I and II to handle outliers and duplicate values;

[0026] 2) Perform feature scaling:

[0027] Using the formula Scale the feature values ​​of different scales to the same range [0, 1];

[0028] Where: x is a sample value in the original data, xmin and x max are the minimum and maximum values ​​of the feature, respectively, x max is the new value after normalization;

[0029] S4: Data segmentation and training model:

[0030] 1) Use the formula Pearson correlation coefficient r was calculated between variables;

[0031] Where: n is the number of samples, x i and i are the sample means of the input and output values ​​respectively;

[0032] Select the Pearson correlation coefficient r with an absolute value between 0.7 and 1, and finally select the participants x i Variables for model training;

[0033] 2) Divide the data into training data and test data in proportion, where the training set accounts for 80% of the data samples and the test set accounts for 20% of the data samples;

[0034] Use decision tree as the base model to select and adjust parameters. Grid search is used to select parameters, that is, to obtain the optimal hyperparameter settings by exhaustively enumerating all possible hyperparameter combinations.

[0035] 80% of the hyperparameter combinations were randomly selected for evaluation and trained on the training data to obtain different algorithms;

[0036] By training the machine learning algorithm on furnace parameters, steel grade composition, smelting process and casting process and actual tapping temperature data, a model reflecting the relationship between these parameters and tapping temperature is constructed;

[0037] S5: Optimization model and evaluation model:

[0038] Based on the machine learning library, the steel tapping temperature prediction is realized. First, data cleaning and segmentation are performed, then training and testing are performed, and finally the model is output for comparison. The model with the smallest mean square error and root mean square error is the optimal model, or the model with the determination coefficient closest to 1 is the optimal model.

[0039] S6: Steel tapping temperature prediction:

[0040] Substitute multiple sets of process data into the trained prediction model and solve the prediction value to obtain the predicted value of the converter tapping temperature.

[0041] Furthermore, in S1, the data analysis library is one of Numpy, Pandas, Sklearn, Seaborn, Matplotilb, and tensorflow.

[0042] Furthermore, in S3, before training the AI ​​model, the data training set is analyzed and preprocessed, and the training data set is loaded through the read_csv() function of pandas to obtain 79997 rows and 77 columns of data, and the pandas DataFrame object is returned.

[0043] Furthermore, in S4, the data X and Y of the training set are divided into feature data and probability data. The variable data is a DataFrame object. By calling the drop function to discard a certain column, the converter tapping temperature is taken as Y, and the remaining data is taken as feature data X. R2 and decision tree regression are used to create and compare models with different depth values. Grid search is used to find the best model, and finally the converter tapping temperature is predicted.

[0044] Furthermore, in S4, the decision tree is as follows:

[0045] from sklearn.tree import DecisionTreeRegressor

[0046] regressor=DecisionTreeRegressor()

[0047] X_train = X

[0048] y_train = Y

[0049] regressor.fit(X_train,y_train)

[0050] First, import the DecisionTreeRegressor class to create a decision tree regression object regressor;

[0051] Then, prepare the training data x_train and y_train, where x_train is the input feature and y_train is the corresponding output value;

[0052] Use the fit() function to fit the training data to the decision tree regression model; use the predict() function to predict the input feature x_train to get the predicted output value y_pred.

[0053] Furthermore, in S5, N different models are created through N different depth values ​​to compare the performance of the N models. First, data cleaning and segmentation are performed to obtain F% of the data for training and (100-F)% of the data for testing. A function is defined to test the model performance, the output models are compared, and the optimal model is selected.

[0054] Furthermore, in S5, four different models are created through four different depth values ​​to compare the performance of the four models, which are displayed in graphs. First, data cleaning and segmentation are performed to obtain 90% of the data for training and 10% of the data for testing. A function is defined to test the model performance, and the depth values ​​are set to 1, 3, 6, and 10. The output models are compared to select the optimal model.

[0055] Furthermore, in S5, the process data include carbon content, alloying elements, furnace type, furnace capacity, oxygen blowing time, blowing time furnace type parameters, steel grade composition information, smelting process parameters and casting process parameters.

[0056] Compared with the existing method, the beneficial effects of the present invention are:

[0057] The present invention is based on machine learning and deep learning, using historical data to create correlations in data feedback in real time to achieve process prediction and optimization. In the ladle processing workflow in the steel industry, operators combine professional skills and use model decisions designed by machine learning and deep learning to dynamically adjust continuous casting process parameters, which can significantly improve the stability of steelmaking production and product quality. The terminal temperature and composition control of converter blowing are important operations in the later stage of converter blowing. Accurate temperature prediction is very important. In order to improve the carbon and temperature hit rates of converter blowing, the feature correlation analysis method is used to determine the main input variables of the model.

[0058] After the implementation of this method, the steel tapping temperature of the converter under different operating conditions can be effectively predicted. Operators can combine their professional skills and use model decisions designed by machine learning and deep learning to dynamically adjust the continuous casting process parameters, which can significantly improve the stability of steelmaking production and product quality.

[0059] The invention discloses a method for predicting the steel-out temperature of a converter based on historical data, which saves time and effort, can accurately and quickly predict the converter terminal temperature, and is beneficial to improving the quality of molten steel and extending the service life of a furnace lining. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the optimal model of the embodiment of the present invention.

[0061] Figure 2 It is the prediction model of the embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention discloses a method for predicting the end temperature of a converter. Those skilled in the art can refer to the content of this article and appropriately improve the process parameters to achieve it. It is particularly important to point out that all similar substitutions and modifications are obvious to those skilled in the art and are considered to be included in the present invention. The method and application of the present invention have been described through preferred embodiments, and relevant personnel can obviously modify or appropriately change and combine the methods and applications described herein without departing from the content, spirit and scope of the present invention to implement and apply the technology of the present invention.

[0063] [Example]

[0064] 1. Environment construction:

[0065] The hardware development environment of the system computer must meet the following conditions: CPU: i5 2.4GHz, hard disk 40G or more, memory 8GB or more, 64-bit operating system.

[0066] Software environment requirements for system operation: Windows 10 64-bit operating system.

[0067] Development language and program version number: Python 3.10.2

[0068] Install Numpy, Pandas, Sklearn, Seaborn, Matplotilb or tensorflow data analysis libraries in your computer.

[0069] 2. Collect converter data:

[0070] The converter tube information includes tens of thousands of converter records, from which the key factors affecting the converter tapping temperature are selected for data cleaning as the feature vectors of the data set, which are:

[0071] (1) Molten iron composition, the weight percentage of Si and P in molten iron, %; Si and P in molten iron are strong heat-generating elements. If the content is too high, it can increase the heat, but it will also bring many problems to smelting.

[0072] (2) Molten iron temperature, °C; the temperature of molten iron is related to the amount of physical heat, so when other conditions remain unchanged, the temperature of the molten iron entering the furnace affects the final temperature.

[0073] (3) The amount of molten iron charged, tons / tank; the increase or decrease in the amount of molten iron charged causes changes in its physical heat and chemical heat. Under the condition that other conditions remain constant, the higher the molten iron ratio, the higher the terminal temperature.

[0074] (4) Furnace age, years; The temperature of a new converter is low and the tapping port is small, so the final temperature in the early stage of the furnace campaign is 20℃-30℃ higher than that of a normal furnace.

[0075] (5) Key carbon content, wt%; Carbon is an important heat-generating element in converter steelmaking.

[0076] 3. Data preprocessing:

[0077] Before training the artificial intelligence model, the data training set is analyzed and preprocessed. The training data set (database) is loaded through the data reading function to obtain 79,997 rows and 77 columns of data. First, data cleaning is performed to deal with missing values, outliers, and duplicate values ​​in the data by interpolation, deletion, or replacement. Then feature scaling is performed: feature values ​​of different scales are scaled to the same range. The specific method used is normalization to scale the data between 0 and 1. The implementation method is as follows:

[0078] From sklearn.preprocessing import MinMaxScaler

[0079] scaler = MinMaxScaler()

[0080] normalized_data=scaler.fit_transform

[0081] 4. Data segmentation and training model:

[0082] The data X and Y of the training set are divided into feature data and probability data. The variable data is a DataFrame object. By calling the drop function to discard a column, the converter tapping temperature data is taken out as Y, and the remaining data is used as feature data. The decision tree regression is used to create and compare models with different depth values. The best model is found using grid search, and finally the converter tapping temperature is predicted. The specific implementation method of the decision tree is as follows:

[0083] from sklearn.tree import DecisionTreeRegressor

[0084] regressor=DecisionTreeRegressor()

[0085] X_train = X

[0086] y_train = Y

[0087] regressor.fit(X_train,y_train)

[0088] First, import the DecisionTreeRegressor class to create a decision tree regression object regressor. Then, prepare the training data x_train and y_train, where x_train is the input feature and y_train is the corresponding output value. Use the fit() function to fit the training data to the decision tree regression model. The input feature x_train to be predicted is predicted using the predict() function to obtain the predicted output value y_pred.

[0089] 5. Optimization model and evaluation model:

[0090] Based on the machine learning library, the steel tapping temperature prediction is realized. Four different models are created through four different depth values ​​to compare the performance of the four models. The performance is displayed through charts. First, the data is cleaned and segmented to obtain 90% of the data for training and 10% of the data for testing. A function is defined to test the model performance. The depth values ​​are set to 1, 3, 6, and 10. The output models are compared and the optimal model is selected. Figure 1 shown.

[0091] The mean square error (MSE) and root mean square error (RMSE) were used as evaluation indicators. 2 Evaluate the goodness of fit of the converter prediction model, which indicates the explanatory power of the model's explanatory variables on the dependent variables, R 2 The value range is from 0 to 1. The closer it is to 1, the better the model's ability to explain the variables, and the closer it is to 0, the poorer the model's ability to explain the variables.

[0092] Calculate R 2 The method is as follows: First, calculate the total sum of squares (SST), Among them, yi is the value of the dependent variable of the observed data, is the mean of the dependent variable of the observed data, and then the regression sum of squares (SSR) is calculated. in, is the predicted value of the regression model corresponding to the i-th observation data. Finally, the residual sum of squares (SSE) is calculated. R 2 =1-(SSE / SST). Thus the optimal model is selected.

[0093] 6. Steel tapping temperature prediction:

[0094] Substitute multiple sets of process data into the prediction model and solve the prediction value such as Figure 2 shown.

[0095] 7. When this method is applied in the experiment, the consistency rate between the predicted results and the true results is greater than 95%, as shown in Table 1.

[0096] Predicted value True value error percentage 1673.27 1702 -28.73 -0.016880141 1675.62 1680 -4.38 -0.002607143 1675.56 1665 10.56 0.006342342 1676.71 1662 14.71 0.008850782

[0097] The present invention is based on machine learning and deep learning, using historical data to create correlations in data feedback in real time to achieve process prediction and optimization. In the ladle processing workflow in the steel industry, operators combine professional skills and use model decisions designed by machine learning and deep learning to dynamically adjust continuous casting process parameters, which can significantly improve the stability of steelmaking production and product quality. The terminal temperature and composition control of converter blowing are important operations in the later stage of converter blowing. Accurate temperature prediction is very important. In order to improve the carbon and temperature hit rates of converter blowing, the feature correlation analysis method is used to determine the main input variables of the model.

[0098] After the implementation of this method, the steel tapping temperature of the converter under different operating conditions can be effectively predicted. Operators can combine their professional skills and use model decisions designed by machine learning and deep learning to dynamically adjust the continuous casting process parameters, which can significantly improve the stability of steelmaking production and product quality.

[0099] The invention discloses a method for predicting the steel-out temperature of a converter based on historical data, which saves time and effort, can accurately and quickly predict the converter terminal temperature, and is beneficial to improving the quality of molten steel and extending the service life of a furnace lining.

[0100] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting converter end point temperature, characterized in that: The method specifically comprises the following steps: S1: Environment construction: Install the data analysis library on your computer; S2: Collect the following converter data: 1) Composition of molten iron, weight percentage of Si and P in molten iron, %; 2) Molten iron temperature, °C; 3) Molten iron loading, tons / tank; 4) Furnace age, years; 5) End point carbon content, wt%; S3: Data preprocessing: 1) Handling missing values, outliers and duplicate values ​​in data: Ⅰ Delete missing values. The deletion process is divided into sample deletion and feature deletion. When the number of samples with missing values ​​in the data set is less than 10% and the missing values ​​are randomly distributed, the rows containing missing values ​​are directly deleted; If less than 10% of the samples have missing values, and there is no obvious pattern in the distribution of missing values ​​in the data set, then the samples with missing values ​​are deleted; If the missing value of a feature exceeds 80%, delete the feature column; Ⅱ Fill in missing values. Filling is divided into mean filling, median filling, mode filling, fixed value filling, regression filling and multiple filling. For numerical features, if the data is normally distributed, the mean is used to fill in missing values. When the data is skewed or has outliers, the median is used to fill in missing values. If fixed business data is missing, it is filled with fixed data. III uses methods of I and II to handle outliers and duplicate values; 2) Perform feature scaling: Using the formula Scale the feature values ​​of different scales to the same range [0, 1]; Where: x is a sample value in the original data, x min and x max are the minimum and maximum values ​​of the feature, respectively, x max is the new value after normalization; S4: Data segmentation and training model: 1) Use the formula Pearson correlation coefficient r was calculated between variables; Where: n is the number of samples, xi and yi are the sample means of input and output values ​​respectively; Select the Pearson correlation coefficient r with an absolute value between 0.7 and 1, and finally select the variables involved in the xi model training; 2) Divide the data into training data and test data in proportion, where the training set accounts for 80% of the data samples and the test set accounts for 20% of the data samples; Use decision tree as the base model to select and adjust parameters. Grid search is used to select parameters, that is, to obtain the optimal hyperparameter settings by exhaustively enumerating all possible hyperparameter combinations. 80% of the hyperparameter combinations were randomly selected for evaluation and trained on the training data to obtain different algorithms; By training the machine learning algorithm on furnace parameters, steel grade composition, smelting process, casting process and actual tapping temperature data, a model reflecting the relationship between these parameters and tapping temperature is constructed; S5: Optimization model and evaluation model: Based on the machine learning library, the steel tapping temperature prediction is realized. First, data cleaning and segmentation are performed, then training and testing are performed, and finally the model is output for comparison. The model with the smallest mean square error and root mean square error is the optimal model, or the model with the determination coefficient closest to 1 is the optimal model. S6: Steel tapping temperature prediction: Substitute multiple sets of process data into the trained prediction model and solve the prediction value to obtain the predicted value of the converter tapping temperature.

2. A method for predicting converter end point temperature according to claim 1, characterized in that: In S1, the data analysis library is one of Numpy, Pandas, Sklearn, Seaborn, Matplotilb, and tensorflow.

3. The method for predicting converter end point temperature according to claim 1, characterized in that: In S3, before training the artificial intelligence model, the data training set is analyzed and preprocessed, and the training data set is loaded through the read_csv() function of pandas to obtain 79997 rows and 77 columns of data, and the pandas DataFrame object is returned.

4. The method for predicting converter end point temperature according to claim 1, characterized in that: In S4, the data X and Y of the training set are divided into feature data and probability data. The variable data is a DataFrame object. By calling the drop function to discard a column, the converter tapping temperature is taken as Y, and the remaining data is taken as feature data X. R2 and decision tree regression are used to create and compare models with different depth values. Grid search is used to find the best model, and finally the converter tapping temperature is predicted.

5. A method for predicting converter end point temperature according to claim 4, characterized in that: In S4, the decision tree is as follows: from sklearn.tree import DecisionTreeRegressor regressor=DecisionTreeRegressor() X_train = X y_train = Y regressor.fit(X_train,y_train) First, import the DecisionTreeRegressor class to create a decision tree regression object regressor; Then, prepare the training data x_train and y_train, where x_train is the input feature and y_train is the corresponding output value; Use the fit() function to fit the training data to the decision tree regression model; use the predict() function to predict the input feature x_train to get the predicted output value y_pred.

6. A method for predicting converter end point temperature according to claim 1, characterized in that: In S5, N different models are created through N different depth values ​​to compare the performance of the N models. First, the data is cleaned and segmented to obtain F% of the data for training and (100-F)% of the data for testing. A function is defined to test the model performance, the output models are compared, and the optimal model is selected.

7. A method for predicting converter end point temperature according to claim 6, characterized in that: In S5, four different models are created through four different depth values ​​to compare the performance of the four models. The performance is shown in the chart. First, the data is cleaned and segmented to obtain 90% of the data for training and 10% of the data for testing. A function is defined to test the model performance. The depth values ​​are set to 1, 3, 6, and 10. The output models are compared to select the optimal model.

8. The method for predicting converter end point temperature according to claim 1, characterized in that: In S5, the process data include carbon content, alloy elements, furnace type, furnace capacity, oxygen blowing time, blowing time furnace type parameters, steel grade composition information, smelting process parameters and casting process parameters.

Images