Machine learning analysis of electric arc furnace process for the evaluation of energy efficiency parameters

Abstract

The electric arc furnace has been the subject of extensive research due to its complex and chaotic nature. Machine learning methods provide a powerful forensic examination of industrial processes as they exclude numerous assumptions and involve irregularities present in industrial conditions. In this study, different machine learning and data processing methods were used to evaluate the energy efficiency parameters of the electric arc furnace process. The dataset was collected over five years, in a steelmaking factory, with 42 features. This data was split into training and test sets, which were used for training and evaluation, respectively. With extensive data management, the data quality and machine learning model performance were improved. It was found that selected models display similar performance, yet the artificial neural network shows greater flexibility when changing targets. The results indicate that a data-centric rather than model-centric approach is better for improving model performance. Using the partial dependence plot and SHAP method, insight was gained into each parameter's correlation with the target. It was found that the amount of hot heel (melted steel left in the furnace, to be re-heated) was the main factor disturbing the data quality and model performance. It was also demonstrated that data for total oxygen consumption should be divided from the oxygen used into refining and natural gas. This highly improves model performance. Employing a data-centric machine learning model to control and optimize main process parameters (with a small capital investment) leads to lower energy consumption for industrial processes.