The hot metal pretreatment desulfurization process in modern converter steelmaking process can improve the efficiency of impurity removal,reduce the burden of converter blowing,and shorten smelting time.It is a necessary process for smelting variety steel and clean steel.The pretreatment process of molten iron undergoes complex high-temperature physical and chemical reactions,which is a black box process,making precise control of the smelting process very difficult.Establishing an accurate process control model is the core of achieving precise control of the hot metal pretreatment process,which is of great significance for enterprises to reduce steel production costs,promote digital and green transformation.This paper summarizes the modeling principles,characteristic and research progress of various models,such as mechanism model,statistical regression model,expert system and machine learning model,established by domestic and foreign researchers for hot metal pretreatment desulfurization process.Based on the different uses of the model,the development process of application in practice and prospects of the hot metal pretreatment desulfurization model were proposed,focusing on the prediction of endpoint sulfur content and desulfurization rate,prediction and optimization of smelting process parameters,and prediction of desulfurization agent consumption and utilization rate.

The prediction of sulfur content of liquid steel in ladle furnace (LF) is of great significance for the precise control of the composition of refined molten steel and the improvement of product quality.In view of the complex mechanism,multi-variable,and nonlinear characteristics of LF,a prediction method of sulfur content in LF was proposed based on autoencoder-back propagation neural network (AE-BPNN).Firstly,the effects of noise and missing values are eliminated through AE network for data reduction and feature extraction.Then,the BPNN is used for predicting the sulfur content of liquid steel in LF.The actual on-site data validation show that the E_RMS,E_MA and the correlation coefficient R² are respectively up to 1.403,1.083 and 0.824,which has good prediction performance.

The converter heat loss rate is one of the important parameters that affect the prediction accuracy of material consumption.Based on the calculation of heat loss rate,the historical production data of 1 900 heats of a 150 t converter in a steel plant was applied to accurately predict the converter heat loss rate using machine learning algorithm.The prediction results show that light gradient boosting machine (LightGBM) algorithm has the highest prediction accuracy compared with support vector regression (SVR) and random forest (RF) algorithms.Considering the influence of the last furnace,after adding the final smelting temperature variable of the last furnace,the determination coefficient R² of LightGBM algorithm increases from 0.89 to 0.93,and within the range of ±0.005 and ±0.01,the prediction hit rate of heat loss rate increased from 85%,89% to 90%,93% respectively.In addition,the prediction accuracy of the model can be further improved by optimizing the internal parameters of the algorithm.For the LightGBM algorithm,the determination coefficient R² and root mean square error (RMSE) further reach 0.94 and 0.009,and the prediction hit rate of the heat loss rate further increases to 91% and 94% within the range of ±0.005 and ±0.01.Based on historical data of converter smelting,intelligent algorithms can be used to predict the heat loss rate of the converter,which can provide support for the prediction of the material consumption of the converter.

The accurate prediction of the Linz Donawitz converter gas (LDG) holder level in iron and steel enterprises can provide an important basis for gas system scheduling.Given the impact of LDG recovery,scheduling workers are particularly concerned about the problem of exceeding the upper limit of the holder level.Based on a large amount of on-site actual data,a cost-sensitive learning-based support vector machine (SVM) method for predicting the level of LDG holder was proposed,which can improve the prediction accuracy of the situation when the level exceeds the upper limit.This method takes the recovery and consumption flow of LDG as inputs,and the future holder level value as output.By using the KKT equation,the original constraint conditions are transformed into equation constraints,different costs are set for the gas holder level exceeding the limit and false alarms.Finally,by minimizing the model′s false alarm error,the original prediction problem is transformed into a series of linear equations and solved.A simulation experiment was conducted on data　from a domestic steel plant,and the results showed that the proposed method can effectively reduce the false alarm rate of converter gas holder level exceeding the limit to 0.16%,providing more rapid and effective guidance for scheduling workers to develop reasonable scheduling strategies.

The precise control of the end-point carbon content and temperature of Consteel electric arc furnace is extremely important to ensure the high-quality liquid steel. Different machine learning models including prediction models and hyper-parameter optimization models were used to fit the actual production data from the smelting process of Consteel electric arc furnace in a domestic steel plant. Based on the prediction results of machine learning models,the end-point carbon content and temperature prediction model of the Consteel electric arc furnace based on Bayesian optimization algorithm (BOA) and gradient boosting decision tree (GBDT) was established. Through the verification by using the collected data,this model achieves better prediction performance for the end-point carbon content and temperature of Consteel electric arc furnace smelting,which can provide certain guidance for the actual production process.

Aiming at the rescheduling problem after the failure of a smelting equipment in the steelmaking-continuous casting production process,in order to ensure the stability of production and reduce the variation of the rescheduling scheme compared with the initial scheduling scheme,a rescheduling model was proposed with the objective of minimizing the weighted variance of the start time,operation cycle and equipment assignment,which was established by mathematical planning method.By analyzing the production operation mode and production process of a steel plant,a rescheduling algorithm was proposed,which consists of an equipment assignment algorithm based on the "furnace-caster coordinating" scheduling strategy rule and a time adjustment algorithm based on the process flexible buffer control strategy rule.Taking the converter equipment failures or refining furnace equipment failures that often occur during the actual production of steelmaking and continuous casting in a large domestic steel plant as simulation samples,the experimental results indicate that the total weighted differences between the start time,operation cycle and equipment assignment before and after scheduling were 0.29,1.43,and 1.21,respectively,which can effectively maintain the consistency between the rescheduling plan and the initial scheduling scheme,and ensure the stability of production.And the solution time was less than 0.6 s,which can quickly provide better solutions to the rescheduling problem after smelting equipment failures.In the actual production process with frequent failures of smelting equipment,the orderly,stable,and efficient operation of steelmaking-continuous casting can be ensured.

Continuous casting slabs are raw materials of steel production.The defect of slab will lead to quality defect of final steel product.The low and high frequency data collected from continuous casting process on site were studied.Cleaning methods of complex process industrial data and feature extraction methods of high frequency industrial data were proposed.Based on machine learning theory,four kinds of slab surface defect prediction models,namely classification and regression tree (CART),AdaBoost,random forest (RF),and optimal classification tree (OCT) were established.Feature selection were carried out using Relief and RF model.The prediction accuracy of different models was compared and analyzed through a large number of experiments.The experimental results show that the RF model gives the best prediction accuracy.The top 10 features,such as liquidus temperature,tundish (TD) lower limit temperature and TD target temperature,which play a key role in slab surface defects are found out.The method in this paper can be extended to industrial data analysis and utilization modeling in other scenarios,which has important reference value for using industrial data to improve product quality.

During the continuous casting process,the instantaneous abnormal mold level fluctuation has great detrimental effects on slab quality,thus the mold level fluctuation is a key parameter for continuous casting process of high-quality steel.In the present study,the slab continuous casting process data for low carbon steel,medium carbon steel,hypo-peritectic steel and peritectic steel were collected,the fast Fourier transform (FFT) and continuous wavelet transform (CWT) were used to analyze the data characteristics,and then the influence of process parameters on the instantaneous abnormal mold level fluctuation was studied.The results of FFT analysis indicate that bulging has no significant effect on the instantaneous abnormal mold level fluctuation.The time-frequency characteristics of instantaneous abnormal mold level fluctuation and stopper-rod position were analyzed by CWT,and the results show that under different steel grades and casting speeds,before the instantaneous abnormal mold level fluctuation occurrence,the CWT coefficients in high-frequency region of stopper-rod position shows a linear increase trend.Therefore,by conducting CWT analysis of the high-frequency region of the stopper-rod position,it is possible to predict the instantaneous abnormal　mold level fluctuation.