Aplicação de algoritmos de aprendizado de máquina na previsão do indicador de tempo de permanência de veículos em uma usina siderúrgica

Abstract

Length of Stay (LOS) is one of the key process indicators of operational efficiency in the scrap metal inbound process of a steel plant. Accurate prediction of LOS not only allows for more efficient resource management but also minimizes bottlenecks, resulting in a more agile and cost-effective operation. In this context, various machine learning models address time series forecasting. This study aims to reshape a transactional database of the scrap metal inbound process, with data from January to September 2021, into a time series of the LOS indicator and apply machine learning algorithms, both classical and fuzzy logic-based, to forecast it. Grid search methods were used to find the best hyperparameters for the models, and their generalizations were evaluated through cross-validation. The models’ performance was evaluated based on metrics such as Mean Average Error (MAE) and Root Mean Average Error (RMSE), and the results showed that the fuzzy logicbased NTSK-wRLS model demonstrated the highest accuracy, standing out for its ability to adapt to the dynamics of the LOS series and effectively capturing the seasonal and operational variations of the process. Finally, this work also contributes to the practical application of machine learning in industrial environments, highlighting its potential to enhance data-driven decision-making processes.