Comprehensive system based on a DNN and LSTM for predicting sinter composition

Abstract

Because of the lag in sinter composition detection, fluctuations in production conditions are not conducive to making timely adjustments to sintering. In this paper, the characteristics of sintering production data are studied, and three core conclusions are drawn. We find that these data have (1) noise, (2) high dimensionality, and (3) time correlation. Based on these findings, an integrated model based on a deep neural network (DNN) and a long short-term memory (LSTM) network is proposed; using a DNN and an LSTM network solves the problem of developing a system model according to the given input and output data to predict the chemical composition of sinter. Specifically, first, we use a box plot and an isolated forest (iForest) algorithm to detect and filter noise in the data preparation stage and then propose using key feature selection and the Pearson correlation coefficient to reduce the high dimensionality of the data. Then, both an online component monitoring model based on a DNN and an advance component prediction model based on an LSTM are proposed to help the field operators to control the change of the sinter composition in real time. The results of many experiments show that the proposed method performs better than the current methods: the goodness of fit (R2) score of the better model is above 0.92, and both the mean square error (MSE) and mean absolute error (MAE) are approximately zero. The deep neural network-based method proposed in this paper is more suitable for the online monitoring and advance prediction of sinter components.