Research On Prediction Of Blast Furnace Temperature Based On Machine Learning And Deep Learning Algorithms

Blast furnace smelting is one of the core links in steel production.Despite the rapid development of blast furnace smelting technology,there is a problem that has always plagued blast furnace smelting production practice,namely real-time monitoring and robust control of blast furnace temperature.At present,the prediction and control of blast furnace temperature largely rely on the experience and judgment of practitioners.Due to insufficient understanding and the complexity of the blast furnace smelting process,the result may be that the blast furnace temperature is in an unstable state,which may have three possible adverse effects.Firstly,the quality of molten iron produced by smelting does not meet the standard,which poses a hidden danger to the quality of a series of subsequent steel products.Secondly,it causes huge waste of raw materials such as ore and fuel.Thirdly,due to the loss of control during the smelting process,excessive emissions and environmental pollution can result,which is detrimental to achieving China’s"30.60"target of carbon peaking and carbon neutrality.Therefore,accelerating the application of big data technology in the steel industry can more effectively explore real-time monitoring and effective control of blast furnace temperature,thereby achieving the goals of meeting the quality standards of molten iron,rational use of raw materials,and carbon emission control within a reasonable range.Modern blast furnace production uses chemical heat as the main reference for blast furnace temperature,which utilizes the positive correlation between silicon content in molten iron and blast furnace temperature to establish a prediction model for silicon content in molten iron,indirectly achieving prediction and control of furnace temperature.This article uses artificial intelligence algorithms to establish a data-driven model for the silicon content of molten iron in the blast furnace smelting process,in order to accurately predict the trend and amplitude of furnace temperature fluctuations,and thus achieve more effective control of the smelting process.The main research content is summarized as follows:(1)Regression prediction of silicon content in blast furnace molten iron based on improved machine learning algorithm.Aiming at the problems of single traditional machine learning algorithm,low prediction accuracy,more feature selection and insufficient weight of manual selection,this paper proposes two improved machine learning algorithms.Firstly,the widely used xgboost,lightgbm and stacking model fusion algorithms were introduced.Based on this,two improved stacking model fusion algorithms are further elaborated.Compared with the single machine learning algorithm and the fusion algorithm with equal weight,the three evaluation indexes of r square(R~2),mean absolute error(MAE)and root-mean-square deviation(RMSE)can show that the two improved stacking model fusion algorithms can improve the regression prediction effect of molten iron silicon content.(2)Regression prediction of silicon content in blast furnace molten iron based on deep learning algorithm.Aiming at the problem of base model selection in fusion algorithm,this paper proposes two deep learning algorithms based on autoencoder and maximizing mutual information.Firstly,the self encoder algorithm model and the maximum mutual information algorithm model based on xgboost,lightgbm and weighted stacking as regressors are constructed respectively.Secondly,in order to better train the deep learning network model,two stages of training optimization were adopted,namely the dynamic adjustment of learning rate training strategy warm-up and model fine-tuning.The results showed that compared to the fine-tuning of the warm-up stage model,all three models achieved certain improvements.Moreover,through the testing of the same dataset in the first part,it was found that two deep learning models can effectively predict the silicon content in molten iron.In addition,in terms of model performance,both models have shown certain improvements compared to traditional machine learning regression models.(3)Classification prediction of blast furnace temperature based on maximum mutual information representation learning technology.Aiming at the low practicability of accurate prediction of silicon content in molten iron,this paper proposes a blast furnace temperature classification model based on maximizing mutual information feature learning technology.Firstly,the temperature of blast furnaces is classified based on a spatiotemporal graph network.In addition,the rationality of classification is evaluated through the contour coefficient evaluation method.Secondly,a classification model based on maximum mutual information feature learning technology is constructed.Utilize the clustering results as the target classification,and ultimately establish a classification model.A classification model with better performance was obtained through two step-by-step methods,pre training and model optimization.The model is finally evaluated through classification report,receiver operating characteristic(AUC area)and confusion matrix.The evaluation results show that the established classification model has good predictive ability.In summary,through horizontal and vertical comparisons and in-depth analysis of relevant typical models,it can be concluded that the maximum mutual information deep learning algorithm can effectively learn the potential features of the data,and the algorithm has an ideal effect on the regression prediction of silicon content in molten iron.This also fully demonstrates that the business data generated in the production practice of blast furnaces plays a very important role in regulating the silicon content of molten iron.The classification model of blast furnace temperature established on this basis can further accurately control the blast furnace temperature.The research results confirm that by utilizing the relationship between silicon content in molten iron and blast furnace temperature,the ideal prediction of silicon content in molten iron can achieve real-time control of blast furnace temperature,ultimately allowing continuous changes in blast furnace temperature within a reasonable range,thereby ensuring efficient blast furnace smelting.This can also achieve the"dual high"goals of smelting and environmental protection.