Modelo preditivo de consumo industrial: Aplicação de aprendizado de máquina na previsão de consumo de um insumo em uma linha de produção localizada no Polo Industrial de Manaus.

Abstract

This study addresses the development of a predictive solution applied to the forecasting of steel strip consumption and replenishment in a razor blade manufacturing line located in the Manaus Industrial Pole, Brazil. The research aimed to propose a machine learning-based approach to support the analysis of replenishment behavior, the estimation of the time interval until the next replenishment event, and decision-making in the manufacturing environment. Methodologically, the investigation was characterized as applied research with a quantitative approach, developed through a case study.Real historical data related to steel strip supply and consumption were collected, processed, organized, and complemented through Monte Carlo simulation, employed as a methodological strategy to expand the database and generate scenarios compatible with the observed operational behavior. Predictive modeling was developed using the Extreme Gradient Boosting (XGBoost) algorithm, considering temporal and historical variables, such as previous records, lag features, and moving averages. The target variable was defined as the time interval, in days, until the next replenishment event.The results indicated that Monte Carlo simulation contributed to expanding the analytical database while maintaining statistical coherence and operational plausibility in relation to the real observations. The XGBoost model presented moderate predictive performance, with an average prediction error below one day, although its overall explanatory capacity remained limited. These findings suggest that broader historical databases and additional operational variables may contribute to improving predictive accuracy in future applications.It is concluded that the proposed approach presents methodological potential to support steel strip replenishment planning, providing an initial framework for the application of machine learning techniques in industrial decision-making processes.