Development of an Automated Property Prediction System for the Pelletizing Process

Abstract

In an integrated steel plant, the iron ore pellet properties affect the productivity and efficiency of the iron-making units. The pelletizing process is very dynamic and is affected by many raw materials and process variables, which cause fluctuations in the cold and high-temperature properties of the pellets. This necessitates an increase in the frequency of sampling and testing in case of process anomalies, which involves tedious and lengthy testing procedures leading to delays in corrective actions by the operators. To address this issue, an online fully automated machine learning-based prediction system was developed and implemented for the instantaneous prediction of the key pellet quality parameters-Cold crushing strength, Reducibility index, and Tumbler index on a continuous basis. The prediction system was developed using a Gaussian gradient boosting algorithm with more than 1,000 datasets as inputs in Python and R language. The developed model showed good accuracy with R2 > 0.95 for the pellet properties. The developed prediction system was implemented on a plant scale through dynamic dashboards with timed automated predictions without manual intervention. A novel ‘incremental updating dataset’ approach was incorporated, which enabled the model to 'self-learn' from the deviations and ongoing trends in the fresh datasets to deliver more accurate predictions. Post-deployment, the model assisted in taking corrective actions to reduce process deviations and helped reduce sampling frequency by 50% resulting in better manpower utilization and lower annual maintenance costs of the testing equipment, leading to significant monetary savings.