Modelo de proyección de la demanda basado en redes neuronal recurrentes de memoria a corto y largo plazo (LSTM) a través del análisis multivariante para la industria Láctea

Abstract

The dairy industries of Ecuador have a slow growth in the use of new technologies for the development of multivariable forecasts, the same ones that provide greater reliability in their results. The use of traditional forecasts generally presents high levels of error in the predictions, causing problems such as overstock, inventory breaks, and even many of its products having to be discarded due to the end of their useful life. It should be noted that the preparation of these forecasts is the first step for planning supply, production, labor and even the number of vehicles needed to transport the product. For this reason, the use of Deep Learning is proposed for the elaboration of demand forecasts, integrating in the model of Recurrent Neural Networks - Short and Long Term Memory (LSTM) some independent variables that affect the behavior of sales, establishing a comparative between the use of the univariate model versus the multivariate model, by evaluating performance metrics. With the objective of demonstrating the reduction of the error in MAPE using the multivariate model. The methodology used is divided into three general phases that correspond to the review of related works and theoretical framework, data collection and preparation, and finally the analysis of the results obtained in each model. Both endogenous and exogenous independent variables were used, achieving a reduction in the MAPE ranging from 1% to 8% compared to the univariate model, this improves the predictions and increases the level of confidence, since characteristics that interfere in the analysis are taken into account. the demand. However, it must be taken into account that these predictions can be improved by using a larger amount of data than that used in this study