Método automático híbrido SSA-ARIMA-NEURAL para previsão multi-step de séries temporais estocásticas

The development of forecasting methods fundamentally aims to reduce the uncertainty inherent in predicting non-deterministic future events. It is necessary because the information about the future behavior of variables allows a better planning of situations to come, independent of what they are. For...

Descripción completa

Detalles Bibliográficos
Autor: Gaio, Gionei
Tipo de recurso: tesis de maestría
Estado:Versión publicada
Fecha de publicación:2018
País:Brasil
Institución:Universidade Estadual do Oeste do Paraná (UNIOESTE)
Repositorio:Biblioteca Digital de Teses e Dissertações do UNIOESTE
Idioma:portugués
OAI Identifier:oai:tede.unioeste.br:tede/4116
Acceso en línea:http://tede.unioeste.br/handle/tede/4116
Access Level:acceso abierto
Palabra clave:Forecasting
Modelos híbridos
Decomposição SSA
Redes neurais artificiais
ARIMA
Hybrid models
SSA decomposition
Artificial neural networks
CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
Descripción
Sumario:The development of forecasting methods fundamentally aims to reduce the uncertainty inherent in predicting non-deterministic future events. It is necessary because the information about the future behavior of variables allows a better planning of situations to come, independent of what they are. For that, using the idea that real world time series are neither pure linear nor non-linear but instead a combination of those, a new hybrid automatic method called "SSA-ARIMA-Neural" is proposed. This method consists in decomposing the original time series by means of Singular Spectrum Analysis and forecasting each component not classified as noise independently. The trend component, which is understood as more purely linear, is modeled by the Box-Jenkins methodology and the oscillatory components, in turn, having non-linear behavior, are approximated by Artificial Neural Networks. In the end, all independent forecasts are summed, generating the final prediction. Aiming at validate this method, a computational experiment was performed using a data set obtained from instrumentation of a large concrete gravitational dam, where the resulting forecasts where compared with the ones generated using consecrated methods from the literature. Those comparisons showed a relevant improvement.