Prediction Model of Converter Oxygen Consumption Based on Recursive …
97
add some other materials to help remove the impurity elements in the furnace and
make them meet the requirements of molten steel composition and stability. In this
process, there are many random and uncertain factors, such as the operation of the
staff, the spatter of molten iron in the furnace, the specific time of adding oxygen,
and so on. They all may affect the consumption of oxygen in varying degrees.
There are many factors affecting converter oxygen consumption, among which
there are different degrees of correlation. Based on the converter steelmaking mechanism, 12 main influencing factors are selected as feature variables. As shown in
Tables 1 and 2, for the correlation analysis of the historical data of dephosphorization and decarbonization converter in a steel enterprise, the numbers in the table are
Pearson correlation coefficients between the two variables. The range of its variation
is [–1, 1], and the absolute value represents the correlation degree, and the positive
and negative, respectively, represent the positive and negative correlation. There are
the correlation coefficients among feature variables in the first 12 rows in the table,
and the underlined number marks the biggest four correlation coefficients. Both tables
indicate that C and Si contents in molten iron, P and Mn contents in molten iron, endpoint P and Mn contents, molten steel quality and steel scrap quality are four pairs of
highly correlated variables, but it is difficult to be explained from the mechanism. The
last row contains the correlation coefficients between feature variables and oxygen
consumption , in which the bold numbers are the biggest two correlation coefficients.
They are corresponding to molten steel quality and steel scrap quality, respectively,
in Table 1, and corresponding to P and Mn contents in molten iron, respectively, in
Table 2. Therefore, the main influencing factors of different converters are different,
and the related mechanism is not clear.
To sum up, the oxygen consumption process of the converter has features with
large data dimensions, strong correlation, unclear mechanism, and influence of
converter function, which makes the data distribution complex. Therefore, it is difficult to select a suitable combination of feature variables and a single model for accurate prediction through simple mechanism analysis. Therefore, we hope to reduce the
complexity of data distribution by recursively dividing the feature space, and then
the fitting pressure of the model will be reduced. Then in the divided subspaces, the
suitable combination of feature variables is selected by the RFE method, and then
further modeling and prediction are carried out.
Modeling Methods
Recursive Division of Feature Space
In this paper, the feature variables are the influence factors of BOF oxygen consumption selected above. The feature space is composed of feature variables, and the
number of feature variables is the dimension of feature space. The purpose of the
recursive division of feature space is to divide the feature space with a complex
97
add some other materials to help remove the impurity elements in the furnace and
make them meet the requirements of molten steel composition and stability. In this
process, there are many random and uncertain factors, such as the operation of the
staff, the spatter of molten iron in the furnace, the specific time of adding oxygen,
and so on. They all may affect the consumption of oxygen in varying degrees.
There are many factors affecting converter oxygen consumption, among which
there are different degrees of correlation. Based on the converter steelmaking mechanism, 12 main influencing factors are selected as feature variables. As shown in
Tables 1 and 2, for the correlation analysis of the historical data of dephosphorization and decarbonization converter in a steel enterprise, the numbers in the table are
Pearson correlation coefficients between the two variables. The range of its variation
is [–1, 1], and the absolute value represents the correlation degree, and the positive
and negative, respectively, represent the positive and negative correlation. There are
the correlation coefficients among feature variables in the first 12 rows in the table,
and the underlined number marks the biggest four correlation coefficients. Both tables
indicate that C and Si contents in molten iron, P and Mn contents in molten iron, endpoint P and Mn contents, molten steel quality and steel scrap quality are four pairs of
highly correlated variables, but it is difficult to be explained from the mechanism. The
last row contains the correlation coefficients between feature variables and oxygen
consumption , in which the bold numbers are the biggest two correlation coefficients.
They are corresponding to molten steel quality and steel scrap quality, respectively,
in Table 1, and corresponding to P and Mn contents in molten iron, respectively, in
Table 2. Therefore, the main influencing factors of different converters are different,
and the related mechanism is not clear.
To sum up, the oxygen consumption process of the converter has features with
large data dimensions, strong correlation, unclear mechanism, and influence of
converter function, which makes the data distribution complex. Therefore, it is difficult to select a suitable combination of feature variables and a single model for accurate prediction through simple mechanism analysis. Therefore, we hope to reduce the
complexity of data distribution by recursively dividing the feature space, and then
the fitting pressure of the model will be reduced. Then in the divided subspaces, the
suitable combination of feature variables is selected by the RFE method, and then
further modeling and prediction are carried out.
Modeling Methods
Recursive Division of Feature Space
In this paper, the feature variables are the influence factors of BOF oxygen consumption selected above. The feature space is composed of feature variables, and the
number of feature variables is the dimension of feature space. The purpose of the
recursive division of feature space is to divide the feature space with a complex