Research On End Point Prediction Model Of VOD Refinery Based On LSTM

In recent years,the demand for stainless steel has increased day by day,and higher requirements have been placed on the quality of stainless steel.As the main method of smelting stainless steel,Vacuum Oxygen Decarburization(VOD)has begun to be put into use in steel production enteiprises.The process of smelting stainless steel in the VOD refining furnace is a typical multi-input,multi-output,nonlinear time-vaiying process,and also a complex multiphase pyrometallurgical process.The accurate predictive control of VOD endpoint parameters is the main control goal of smelting.Among them,the end-point carbon content and temperature are the most critical parameters,which have a nonlinear and coupled nonlinear relationship with multiple variables,and there are problems that a large number of parameters cannot be measured in real time.Establishing an accurate prediction model to predict the end point carbon temperature of VOD can increase the end point hit rate and increase production efficiency.This article first analyzes the VOD smelting process,and briefly analyzes the refining mechanism process,briefly analyzes the physical and chemical reactions that occur in each stage and the system heat budget,and determines the input parameters that affect the endpoint carbon temperature.Secondly,in view of the numerous parameters of the VOD smelting process,the nonlinearity between the parameters,and the redundancy,the in-depth study of the relevant knowledge of the feature selection theoiy,the use of Mutual Infbmiation(Mutual Information)based on the Maximum Relevance Min Redundancy(mRMR),MI)feature selection method,to select input parameters,not only can measure the correlation between variables,but also consider the nonlinearity and redundancy between variables,thereby reducing model parameters and redundant features.Then,the working principles of neural networks and recurrent neural networks were studied in depth.Aiming at the current problems of gradient disappearance and gradient explosion in recurrent neural networks,the algorithm model used in this paper was led to the long short term memory network(Long Short Term Memory Networks,LSTM).The working principle has been studied in depth,and the LSTM algorithm has been simulated experiments on CO and NOx emissions on a gas turbine.The results show that,compared with the results of extreme learning machine and random forest algorithm,LSTM has a stronger predictive effect and generalization ability in processing industrial parameter predictive control with time series characteristics.Finally,the feature selection method based on mutual information obtains the input parameters of the VOD end point carbon temperature prediction model,and the setting of the training parameters of the LSTM network model is studied,including the selection of the loss function and the method of improving the generalization ability of the model.Then,a Support Vector Regression(SVR)prediction model,an LSTM model without feature selection,and an LSTM endpoint prediction model based on mutual information were established respectively.Aiming at the problem of SVR parameter selection,genetic algorithm is used to optimize SVR*s hyper parameters to improve the SVR prediction effect.The results show that through feature selection,the number of parameters and the redundancy between parameters can be reduced,and the complexity of the model can be reduced.At the same time,the application of deep learning algorithms to the predictive control of steel making process parameters opens up a new method for steel making endpoint prediction methods.Simulation experiments show that the method proposed in this paper is better than the traditional BP neural network model and SVR model,can improve the end point carbon temperature hit rate of VOD refining furnace,can be used to guide the VOD production process,and expand to other metallurgical fields.