Título: Previsão de Cargas Elétricas por Redes Neurais: Uma Investigação Empírica do Problema de Overfitting
Autores: Hippert, Henrique S.; Bunn, Derek W.; Souza, Reinaldo C.
Resumo: In the last few years, models based on neural networks have been proposed with increasing frequency for load profile forecasting. Many forecasters are still sceptical about the performance of such models, arguing that most of them are based on very large neural networks that seem to be overfitted, and also, that most of them have not been properly validated, as their performance has not been compared to that of standard linear models. However, despite the scepticism, some of those NN-based models seem to have been tried out in practice, with satisfactory results. Could these models be really overfitted, and still be able to produce good forecasts? In this paper we investigate the effects of overfitting in the accuracy of NN-based forecasting systems, by comparing the performance of one such system to that of standard linear methods, over a series of actual load data.
Código DOI: 10.21528/CBRN2001-094
Artigo em pdf: 5cbrn_094.pdf
Arquivo BibTex: 5cbrn_094.bib