Previsão de séries temporais utilizando modelos clássicos e redes neurais artificiais.

Abstract

This work presents the main concepts that involve the analysis of time series, as well as the main models used in the time series forecasting. At the end of the work, a case study was carried out, in which the load Northeast region was predicted using a classical forecasting model, specifically the seasonal ARIMA model, and an artificial neural network (ANN). The methodology used to identify the ARIMA model is based on the analysis of autocorrelation and partial autocorrelation functions of the time series. The ANN was implemented in the Matlab® software through the Neural Network Time Series Tool toolbox. The quality of adjustment of models and forecasts was measured by performance indexes. It was concluded that both forecasting methods are satisfactory, since they have good performance indices. However, artificial neural networks have some advantages over the ARIMA model, such as the possibility of incorporating regression variables into the model. This work has as main contribution to initiate a research line in the High Voltage Laboratory of the UFCG, serving as a guide for the graduating students who will work on the subject from now on.