Monitoramento e ajuste de modelos de previsão baseado em gráficos de controle

Abstract

Industrial processes are commonly subject to stochastic effects such as operational failures and leading time variations. The main purpose of this research is the development of an error monitoring system for the re-estimation of parameters and readjustment of forecasting models in the recurrence of errors out of statistical control. An approach was developed using ARIMA models as predictors of future system behavior, and EWMA control chart for error monitoring. The proposed methodology was implemented in the programming language 'R' and evaluated using a test case comprising a key performance indicator of an unloading process at a large port terminal. It was demonstrated an average gain of 114% in the time interval until the predictions show recurrence in out-of-control errors out-of-control compared to the traditional approach, which considered only one forecast model over the whole horizon of the study. It was also demonstrated a global gain in terms of accuracy, indicating that the proposed approach is a viable alternative for monitoring prediction errors in an industrial context.