Name: LUCAS LUPPI AMORIM
Publication date: 17/06/2026
Examining board:
Name |
Role |
|---|---|
| DAVID MENOTTI GOMES | Examinador Externo |
| DIEGO ROBERTO COLOMBO DIAS | Examinador Interno |
| FILIPE WALL MUTZ | Coorientador |
| FRANCISCO DE ASSIS BOLDT | Examinador Externo |
| THIAGO OLIVEIRA DOS SANTOS | Presidente |
Pages
Summary: Controlling the hot metal temperature (HMT) in blast furnaces is a critical task in the iron and steel manufacturing industry, as it significantly influences product quality, operational safety, and process efficiency. Due to the high number of variables involved and the complex dynamics of the system, effective control is traditionally performed by experienced human operators. This work presents a data-driven control system that leverages historical operator data through a hierarchical nearest-neighbor search strategy. The system selects past control sequences that (i) led to a desired temperature variation, (ii) began from a furnace state similar to the current one, and (iii) are consistent with recently applied actions. Based on the selected sequence, a new multi-step control plan is generated. The method does not rely on an explicit model of the furnace, making it suitable for evaluation in simulated environments. As direct testing in an operating blast furnace involves significant operational and safety constraints, furnace behavior was simulated using machine learning models, including Linear Regression, Partial Least Squares, Random Forest, XGBoost, LightGBM, and LSTM neural networks. Experimental results show the system is capable of accurately controlling the temperature under a variety of conditions, including different initial states and target values.
Additionally, this work introduces a control system configuration that incorporates a filtering mechanism based on the extrapolation of control variables across the historical dataset. This refinement helps reduce the impact of operational control limits when generating control sequences. As a complementary study, an alternative similarity filtering strategy based on learned embeddings is also investigated, in which historical samples are compared in a low-dimensional representation learned by a neural network. Moreover, this work also presents an extensive stability evaluation of the system with the addition of different levels of noise in the output of the furnace models, allowing a thorough assessment of its stability and robustness. Experiments with different machine learning models that simulate the blast furnace show that this novel method yields better stability and accuracy for all tested models and for different levels of noise.
