A Neural Networks Based Model for the Prediction of the Bottled Propane Gas Sales

Abstract

This work presents an application of the artificial neural networks (ANN) in the prediction of the time series of the weekly wholesales of bottled propane gas (13 kg. Bottles). For this purpose several networks with different topologies were built. In order to reduce the error of the predictions, many schemas of ensembles were applied. Additionally, given the scarce data available, it was mandatory to minimize the input dimensionality of the networks and to do this, with a rational and systematic approach, considerations about stochastic dynamical systems were made and the Deyle and Sugihara's theorems for nonlinear state space reconstruction as long the generalizations of the Takens-Mañé's theorem for non-deterministic systems were used.