Prediction Of Blast Furnace Temperature Based On Skewed Depth Censored Data

As an important pillar industry of China's economic development and national defense construction,iron and steel industry is also an important indicator of China's economic strength.In industry,Blast Furnace(BF)ironmaking is one of the main ways of pig iron production.It has become an important topic in the new industrial era to ensure the quantity,increase production and stabilize the quality of iron production.In the iron making process,the running condition of BF is often monitored by the fluctuation of furnace temperature.In view of the characteristics of BF system,such as time-varying,hysteresis,chaos and nonlinearity,it is difficult to meet the needs of artificial online closed-loop control,which makes the real-time prediction and stable regulation of furnace temperature an unsolved problem.The silicon content of molten iron can indirectly feedback the thermal reaction information in the BF,which can be used to measure the fluctuation level of furnace temperature.Therefore,the accurate prediction of silicon content has an important guiding significance for furnace temperature monitoring and BF stability.Furnace temperature fluctuation is often affected by many uncertain factors,including internal and external factors.According to a large number of experimental records,there is a strong non-linear characteristic between the process parameters that affect the change of silicon content.The dynamic prediction of silicon content is always traceable,and the non-stationary of BF condition often brings great trouble to the furnace temperature prediction.When the BF condition fluctuates violently,the generation of abnormal data cannot be avoided.The effective classification of abnormal data is very important for the improvement of the prediction performance of the model.Considering the advantages of Elman neural network,such as strong fault tolerance,adaptability,ability to process information in parallel,this paper uses this method to study the relationship between the change of process parameters and the change of silicon content,and makes some improvements in the research methods.Based on the previous research,this paper analyzes the online production data of Baotou Steel Group,and constructs two kinds of furnace temperature prediction models from the new perspective of deep learning.In the research idea,firstly,considering the influence of one-stage lag between the parameters,the data is processed by difference;Secondly,here combines the censored features of the data.,the biased projection depth intelligent algorithm is used to effectively divide the data into stable and outlier classes,and Elman network silicon content prediction model and Logistic furnace temperature fluctuation prediction model are established for two kinds of samples respectively.It is found that the sample data after classification still inherits the characteristics of time series and can realize the real-time prediction of silicon content series.In terms of content,this paper discusses the innovation and deficiency of the new method,and compares the prediction results with Elman network model based on traditional time series analysis,and summarizes the advantages of biased depth intelligent algorithm and depth classification network performance.The empirical results show that Elman network based on biased depth truncation data performs better than traditional time series Elman model in terms of mean square error and hit rate.Among them,on the 156 heats of test samples,the prediction accuracy of traditional time series Elman model reaches 80.1%,and the network mean square error reaches 0.386,while the prediction accuracy of truncated Elman network model reaches 85.3%,and the network mean square error is only 0.375.In addition,this paper discusses the influence of BF condition outliers on furnace temperature fluctuation,and establishes a Logistic model for predicting the direction of furnace temperature fluctuation for outliers.The overall hit rate of the Logistic regression model is as high as 82.6%,in the down direction,the model recall rate is 66.7%,in the up direction,the model recall rate is 100%.Finally,according to the different fluctuation directions of the furnace temperature of the censored data,the maximum sample depth points of the upward and downward furnace temperature are calculated respectively,which provides the best guidance direction for the control of the furnace temperature of the outliers,and it has certain reference value based on the perspective of machine learning.