Delineamento de experimentos no treinamento de redes neurais artificiais para o problema de previsão de séries temporais não lineares

Abstract

A time series is defined as a collection of observations of a variable over time, whose data order has a fundamental importance due to the dependence between these consecutive values. The analysis of these data, and the understanding of this correlation, is an important tool in understanding phenomena in various sciences, such as Economics, Engineering and Operations Management, where prices, demands and values are these variables. The modeling of this data sequence provides its use in order to, based on historical data, make predictions for future periods. This consecutive relationship can be considered complex and, not uncommonly, non-linear. The use of Artificial Neural Networks has proven increasingly effective in establishing pattern recognition, modeling and predicting future values. The statistical programs available on the market provide user-friendly tools and results demonstrated in several scientific available in publications, but the number of factors and levels that are available for use during the training of Artificial Neural Networks, which may indicate the need for hundreds of years to execute every possible combination. In this study, the statistical methodology of Design of Experiments (DOE) is applied in order to determine the best parameters of an Artificial Neural Network for the prediction of non-linear time series and, thus, significantly reduce the time needed to point out the choice of the best Artificial Neural Network capable of solving our prediction problem. Instead of using the most common technique for training an Artificial Neural Network, that is, the empirical method, DOE is proposed to be the best methodology. The main motivation for this dissertation was the prediction of non-linear seasonal time series - which is related to many real problems, such as short-term electrical load, daily prices and returns, water consumption, etc. A case study is presented. The objective was fulfilled when it was proved to reach error results, between prediction and real value, smaller for the Artificial Neural Network than the error reached with the model.