Forecasting of iron ore sintering quality index: A latent variable method with deep inner structure

Accurate and real-time estimation of iron ore sintering quality index is essential for the stability of the production process. However, the sintering process data is generally characterized by high dimensionality, collinearity, nonlinearity, and dynamic features, which seriously hinder the modeling performance. To cope with the complex properties mentioned above, an effective fusion of recurrent neural network with the gated recurrent unit and partial least squares (GRU-PLS) is introduced for ferrous oxide (FeO) prediction of the finished sinter. The proposed GRU-PLS model takes the advantage of conventional latent variable method but incorporates the deep inner structure between each pair of latent variables that captures the nonlinear and dynamic information simultaneously. The modeling performance of the proposed model is evaluated by the actual data collected from the iron ore sintering process in a large iron and steel group in South China. The results show the lowest prediction error of the GRU-PLS model in comparison with other counterparts. More specifically, the rootmean-square error of the GRU-PLS model is decreased by 35.29% in comparison with that for the recurrent neural network with the gated recurrent unit. • A latent variable model (LVM) with deep inner structure is proposed. • Recurrent neural network (RNN) is introduced for each iteration of LVM. • There is a relationship of mutual promotion between LVM and RNN. • Evaluate modeling performance through a real iron ore sintering process. • The proposed model shows the highest prediction accuracy.