Prediction of Sinter Properties Using a Hyper-Parameter-Tuned Artificial Neural Network

Abstract

The present work aims at performing prediction validation for the physical properties of coke layered and nonlayered hybrid pelletized sinter (HPS) using artificial neural networks (ANNs). Physical property analyses were experimentally performed on the two HPS products. The ANN model was then trained to obtain the best prediction results with the grid-search hyper-parameter tuning method. The learning rate, momentum constant, and the number of neurons varied over specified ranges. The binary variable conversion was utilized to assess the two sintering processes. The nonlayered HPS product of 4 mm micropellets at basicity 1.75 and using 8% coke shows a good combination of physical properties, whereas HPS of 4 mm micropellets at 1.5 basicity using 4% coke as fuel and 2% coke as layering gives a radical improvement in physical properties. The yield of the HPS product is 96.07%, with the shatter index (SI), tumbler index (TI), and abrasion index (AI) values being 86.12, 79.60, and 5.74%, respectively. Hence, HPS can be preferred by implementing the layering of coke powder. The prediction analyses showed that the multilayer perceptron model (MLP) network with a 4-29-5 structure showed prediction accuracies of over 99.99% and a mean squared error (MSE) of 2.87 x 10-4. It verifies the accuracy and prediction effectiveness of the hyper-parameter-tuned ANN model.