Prediction Of Hearth Activity Based On Data Mining

The hearth is an important core of the iron-making blast furnace,and its activity affects the steel output and tapping quality.Quantitative analysis of hearth activity can ensure the stable operation of blast furnace and improve the economic benefits of steel industry.In the field of metallurgical engineering,quantitative calculation of hearth activity relies on empirical formulas.The calculation process is complicated and it highly relies metallurgical experience.The calculation lost its meaning when some parameters are difficult to obtain.In response to these problems,a calculation model of hearth activity based on machine learning has been proposed in this article.Considering that the data of blast furnace contains time-series information,a prediction model based on deep network is proposed.The specific work is as follows:(1)Data preprocessing.First,the background knowledge of blast furnace ironmaking and blast furnace parameters are introduced.Then the collected data is analyzed,including filling in missing values,removing outliers and calculating the temperature of core dead stock column according to the existing parameters.Finally,the data characteristics are standardized and the evaluation indexes of the model performance in this article are given.(2)Establish a model based on distance correlation coefficient and ridge regression.For the preprocessed data,the main features are extracted based on the distance correlation coefficient.Performing redundancy analysis on these features and deleting redundant features to obtain the feature set.Using ridge regression to eliminate the influence of collinearity and learn the relationship between features.Then analyzing the generalization ability of the model.Listing the regression equation and proposing the theory for regulating the hearth activity of the blast furnace based on the ratio of the coefficients in the equation and the technical experience.(3)Create a long and short-term memory network model based on deep learning.For the preprocessed data,the change of each parameter is a continuous time series.In view of this,analyzing how many moments of data before the predicted target are related to the predicted target firstly.Using one-dimensional grouped convolution to extract the time series information.Splicing output tensor on the channels and then using the two-dimensional convolution to extract the deep semantic features contained in the sequence.The extracted features will be input into the long and short-term memory network to obtain the predicted value of the temperature of core dead stock column.Finally,the superiority of the method proposed in this article is comprehensively compared in terms of performance and application value.Through experimental verification,the ridge regression model established in this article can accurately handle the correlation between variables and it has good predictive performance.Its mean absolute error is 4.01,the mean square error is 25.95,the mean absolute percentage error is 0.282% and the R-squared value is 0.81.The long and short-term memory network model based on deep learning uses the time series information of the data.The predictive performance and generalization ability of the model are improved.Its mean absolute error is 0.96,the mean square error is2.74,the mean absolute percentage error is 0.067% and the R-squared value is 0.985.Compared with other works,the parameters of the hearth activity prediction model in this article are easy to collect.It can make full use of the information of the relevant parameters and optimize the performance of the model.So it can improve the monitoring ability of the hearth state and with good practical significance.