Algoritmos de aprendizado de máquinas aplicados no dimensionamento e controle de estoque na indústria de bebidas

Abstract

Inventory planning is an extremely important task within any industry. This importance becomes even more remarkable for industries that deals with complex and intricate supply chains, with plants spread around many regions of the country, varying lead times and different suppliers. In this work, we will use data science and machine learning techniques to carry out forecasting and inventory control of various supplies in different plants, in a large beer industry. For the consumption forecast, Holt-Winters, Gradient Boosting (LGBM implementation), Dense Neural Networks and a simple persistence model (predicting the future as being exactly the past values) were used. The predictive models generated were validated using the mean absolute error (MAE) metric and the residuals were tested for normality (Shapiro-Wilk), zero mean (t-test) and autocorrelation (Ljung-Box). A Python software was developed in order to simulate a predictive inventory control system using the prediction of each of these models alongside with a heuristic inventory policy provided by the company. The resizing of the inventory policy was also tested (lower bound threshold), taking into account the predictive performance of the models for each time series. Finally, using inventory metrics from simulation and the Pareto front technique for multiobjective optimization, the best candidates were selected for further validation in production stage.