Research And Application Of Data-driven End-point Control Models Of Converter

BOF steelmaking is a multi-component multiphase high temperature physicochemical process with fast chemical reaction rate and many influencing factors.Due to the complexity of converter steelmaking mechanism and the diversity of influencing factors,as well as the inability to obtain direct and continuous molten steel composition and temperature information in the entire smelting process,it is difficult to accurately control the endpoint of converter steelmaking.The end point control mode of converter can be divided into three types:manual experience control,static control and dynamic control.Among them,static control is the basis of dynamic control,and dynamic control is the modification of static control based on detection information.The key of static control and dynamic control is the precision of process control model.Aiming at the problems existing in the current converter endpoint control model,combined with the existing end point control mode of converter,new static control model and dynamic control model of converter were established by combining metallurgical mechanism and artificial intelligence algorithm.In terms of converter static control model,in view of some unreasonable assumptions(90%CO and 10%CO2 generated by carbon oxidation and WFeO=15%of final slag)in the calculation of oxygen consumption in the traditional mechanism model,a converter oxygen consumption prediction model based on staged oxygen decarburization efficiency was established by integrating metallurgical mechanism and case reasoning algorithm.The results show that the hit rate of the relative error of oxygen consumption prediction within the range of[-5%,5%]is 94%,compared with the traditional data-driven model,the hit rate is increased by 16.33%.Aiming at the problem of large error of lime addition calculated by traditional basicity theory,an improved prediction model of converter lime addition was established.The results show that hit rate of the model for lime addition prediction error per ton of steel within the range of[-10 kg/t,10 kg/t]is 84%,which is 2.67%higher than that of the traditional case-based reasoning model.Because the static control model can only consider the initial state and final state,and ignores the dynamic control parameters of the blowing process,the hit rate of converter end point control is not high.Therefore,it is necessary to establish a prediction model for the converter blowing process,and conduct soft sensing for the composition(mainly carbon content)and temperature.In view of the fact that the influence of process parameters such as gun position change and oxygen supply flow on the end point is not considered in the traditional data-driven model,an endpoint prediction model of converter was established by integrating multimodal data(single value data and time series data).The results show that the hit rate of the model for carbon content prediction error within the range of[-0.02%,0.02%]is 85%,and for the end point temperature prediction error within the range of[-15℃,15℃]is 89%,which is 5%higher than that of the traditional data-driven model.In terms of the converter dynamic control model,aiming at the problem of collaborative control of process parameters in the second blowing stage,a control model for the second-blowing stage was established by integrating support vector machine and multi task learning algorithm.The results show that the the hit rate of the model for the prediction error of oxygen consumption per ton of steel in the second-blowing stage within the range of[-1.5 Nm3/t,1.5 Nm3/t]is 89%,the accuracy rate for judging the type of coolant/heat supplement is 91%,and the hit rate for the prediction error of coolant/heat supplement addition per ton of steel within the range of[-1.5 kg/t,1.5 kg/t]is 85%.In view of the delay of the traditional off-gas analysis model and the problem that the artificial intelligence algorithm model can only realize the endpoint prediction but not the dynamic prediction,a real-time dynamic prediction model for the carbon content and temperature of the converter in the late blowing period(after TSC detection)is established by combining case reasoning and long and short cycle memory network algorithm,the results show that the hit rate of the model for the prediction error end-point carbon content within the range of[-0.02%,0.02%]is 91%,and the hit rate for the prediction error of end-point temperature within the range of[-15℃,15℃]is 89.17%.Compared with the traditional model,it has increased by 6%and 5.42%respectively.Finally,based on the research results of the end-point control model mentioned above,combined with the automation system,the converter intelligent steelmaking system was developed,which can provide an accurate control scheme for the endpoint control of the converter.The result of software trial operation shows that the hit rate of the prediction error for lime addition per ton of steel with the range of[2.5 kg/t,2.5 kg/t],[-5 kg/t,5 kg/t],[-7.5 kg/t,7.5 kg/t]and[-10 kg/t,10 kg/t]were 31%,53%,66%and 80%,the hit rate of the relative prediction error for oxygen consumption with the range of[-3%,3%],[-5%,5%],[-8%,8%]and[-10%,10%]were 77%,93%,99%and 100%,the hit rate of the prediction error for end-point carbon content within the range of[-0.005%,0.005%],[-0.01%,0.01%],[0.015%,0.015%]and[-0.02%,0.02%]were 21%,44%,69%and 85%,and the prediction error for end-point temperature within the range of[-5℃,5℃],[10℃,10℃],[-15℃,15℃]and[-20℃,20℃]were 31%,70%,90%and 95%.