Predição não-linear de series temporais usando redes neurais RBF por decomposição em componentes principais

AUTOR(ES)
DATA DE PUBLICAÇÃO

2001

RESUMO

This thesis proposes a new technique for non-linear time series forecasting based upon Radial Basis Function Neural Networks and the Karhunen-Loeve Transform. A significant performance improvement is obtained with the novel technique in comparison with usual prediction methods. By obtaining the neural network centers from the data set sub-spaces - or data set principal components - the new method yields lower prediction error and requires less previous known samples than the usual technique that applies the own training set vectors to the centers. Additionally we present a possible solution to the problem of dynamically adapting the neural network architecture to the time-varying series statistics

ASSUNTO(S)

redes neurais (computação) telecomunicações teoria da previsão analise de series temporais analise de componentes principais

Documentos Relacionados