One-Sided Relational Autoencoder With Seasonal-Trend Decomposition to Extract Process Correlations for Molten Iron Quality Prediction
Duojin Yan, Chunjie Yang, Shuming Sun, Siwei Lou, Liyuan Kong, Yuyan Zhang
Abstract
Accurate prediction of molten iron quality (MIQ) is significant for monitoring the production status in the blast furnace (BF) ironmaking process. While most previous data-driven models focus on the nonlinearity and dynamics of process variables, few works take the problem of multicomponent mixing and poor correlation between quality and process variables into account. To tackle these issues, we propose a novel one-sided relational autoencoder with seasonal-trend decomposition (OSRAE-ST) method to predict MIQ. Initially, the process variables are adaptively decomposed using the frequency domain analysis-based decomposition module, which effectively decouples the multicomponent. Subsequently, the one-sided relational autoencoder that models the process and quality association relationship is designed to capture quality-related latent features. In OSRAE-ST, the latent features and historical quality variables data are input in parallel into prediction networks accounting for nonlinearity and dynamics. By enhancing the correlation between quality variables and latent features obtained from decoupled process variables, this framework captures more quality-related information for accurate predictions. Experimental results on typical quality variables of BF ironmaking demonstrate the superior prediction performance of this method.