Título: Inferência Bayesiana No Desenvolvimento De Previsores Neurais De Vazão Diária Utilizando Informações De Precipitação
Autores: Leocádio, Caio Monteiro; Ferreira, Vitor Hugo
Resumo: This paper deals with Bayesian neural models applied to daily water inflow forecasting, including automatic algorithms for input selection, structure stabilization and complexity control. The available database includes water inflow and precipitation in a daily basis from Grande River basin, enabling tests of the applicability of the inclusion of precipitation information in the daily water inflow forecasting models in order to decrease the forecasting error of the developed neural models. Six forecasting strategies are compared, all of them using Bayesian neural models but different input space representation, including models that use only past water inflow information and models that combine precipitation and past values of the water inflows. The use of seasonal dummy variables in order to represent the dry, wet and transition periods of the year is also tested. The obtained results show the viability of the application of Bayesian neural models combining water inflows and precipitation as inputs for daily water inflow forecasting, with these models showing the best results for the test periods considered.
Palavras-chave: Daily water inflow forecasting; energy planning; neural networks; Bayesian inference applied to neural networks
Código DOI: 10.21528/CBIC2011-01.5
Artigo em pdf: st_01.5.pdf
Arquivo BibTex: st_01.5.bib