Research On Molten Iron Quality Prediction Of Blast Furnace Based On ARMAX-LSTM Model

The iron and steel industry is the foundation of national economy.Molten iron quality(MIQ)parameters such as silicon,manganese,phosphorus and sulfur content,are key indicators to reflect the performance of iron making and ensure high energy efficiency in smelting process.Addressing the multi-operating modes,multi-phase coupling and nonlinear characteristics in blast furnace,MIQ prediction has always been a difficult project in metallurgical automation area.In this paper,the research is carried out based on the reaction characteristics of the blast furnace:First of all,a reasonable amount of process-related variables is determined by mech-anisms of chemical reaction and statistical analysis of actual production data.According to the characteristics of the system,ARMAX model and LSTM model are selected as the main modeling algorithms.Then,the ARMAX model is utilized to identify the system by implementing the online-ARMAX identification algorithm.Aiming to decrease unsolvable risk of the online algorithm,a weak stationary Bayesian criterion is proposed.Then the LSTM model is applied to predict the MIQ.Based on the previous research of the algorithms,an ARMAX-LSTM framework is proposed to model the industrial process.The framework firstly uses the ARMAX model to extract linear information in original data,and then uses the LSTM algorithm to model the residuals of the former.By using the framework,the LSTM scale is largely reduced and prediction accuracy is significantly improved.At last,in order to solve the multi-operating mode switching problem in blast furnace iron making process,the research proposed 2 algorithmo?In response to significant changes in operating conditions,an empirical algorithm based on continuous step-size is introduced.The algorithm monitors the first-order difference(1-diff)of the process variable.When a large 1-diff value appears and the subsequent 1-diff value maintains small value of sufficient step size,ARMAX-LSTM model is triggered to re-set.In response to implicit mode switching,the ARMAX-LSTM model is then modified to sliced multi-model,which means to connect multiple ARMAX-LSTM model with Softmax function.In this model,one ARMAX-LSTM represents one corresponding hiden mode.Remarkably,the whole framework gets better prediction performance than previous ones,indicating a bright potential in industrial applications.