Improving first-principle model accuracy using a hybrid approach: A case study to high pressure grinding rolls in mineral industry. Issue 21 (2022)

• Record Type: Journal Article
• Authors: Mohammadi, Ahad; Chioua, Moncef
• Is Part Of: IFAC-PapersOnLine. Volume 55:Issue 21(2022)
• Page Start: 7
• Page End: 12
• Publication Date: 2022
• Subjects: Hybrid Model -- Grinding-Flotation Circuit -- Machine Learning -- First-principle Model -- MultiLayer Perceptron (MLP); Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2022.09.235
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: High-pressure grinding rolls are key assets in ore processing because of their low energy consumption and high capacity. Accurately predicting the purity of the obtained product is paramount as it directly impacts the predicted economic benefits. First-principle models of grinding- flotation circuits can be used to predict the product grade. These models can, however, exhibit deviations from actual process values ("model-plant mismatch"). Deviations between first-principle models and actual processes originate mainly from unmodeled or poorly approximated relationships between process variables. This paper presents a hybrid modeling approach to improve the prediction of mineral grade in first-principle models of grinding flotation circuits. A Multi-Layer Perceptron (MLP) is combined with a first-principle model to reconstruct the mismatch due to the unmeasured effects of grain size distribution on the mineral grade. Two different backup soft sensors are proposed to predict the product mineral grade when two separate faults occur in the online grade analyzer. The hybrid model is able to estimate the mismatch with a coefficient of determination (R 2 ) of 0.97 and a mean absolute error (MAE) of 0.011.