Applications of latent variable methods in primary aluminium production

Abstract

The aluminum industry is currently facing challenges on multiple fronts. One of them is the raw material quality and their impact on process stability. In modern aluminum smelter plants, there are increasing environmental and cost efficiency demands while raw material quality decreases. The digitalization of the manufacturing industry, dubbed Industry 4.0, opens up new avenues of exploration to meet the demands of the 21st century aluminum production. In this research, big data analytics methods were employed to determine the stability of a potline. An anomaly detection tool was built using principal component analysis. The prototype showed that drifts in process operations can be correctly identified earlier than with existing univariate statistical process control tools. The multivariate prototype was later implemented in production at Alcoa Fjardaal. Using the principal component analysis model, a ranking system was developed as the prioritization method of the problematic reduction cells. The variables’ contributions plots were used to help identify the issues in the problem cells. The model performance was assessed with the help of operators attending to the pots on the shop floor. A comparison with existing plant tools showed the potential of advanced notice of the onset of process anomalies. Through an anode tracking system, smelter data was collected to estimate the net carbon consumption (NCC) of individual anodes using latent variable methods. The properties of individual anodes were thus linked to their corresponding reduction cell performance. A multivariate statistical analysis of the linked carbon and pot operation databases, was compared to the existing net carbon consumption formula developed by R&D Carbon. The comparison demonstrated that the prediction for net carbon consumption, for individual anodes, significantly increased using the anode tracking system and latent variable methods. The anode tracking data was split into several blocks determined by manufacturing steps sequence. The impact of each block onto the net carbon consumption was determined using the sequential multi-block partial least squares algorithm. This methodology helped establish a link between green anode properties and the net carbon consumption of individual anodes.