Métodos univariados e multivariados para previsão da demanda de energia elétrica em curto prazo: um estudo comparativo

Abstract

Short-term demand forecasts a are vital part of the production plan and control on electrical power systems. As it is not possible to keep large inventories to meet sudden demand increases, the generation scheduling is based on forecasts made for some hours ahead. Throughout the years, many methods have been proposed in order to solve the problem. Among the most popular are the univariate ones, on which the demand is written as a linear function of its historical behavior and forecast by statistical techniques. It is also common to use multivariate methods, which take into account also the nonlinear effects produced on the demand by weather-related variables, such as the air temperature. For this case, recent papers suggest the use of computational intelligence devices, such as artificial neural networks. Although some authors claim that multivariate methods must be considered, some others state that, on a short-run (lead-times up to a few hours), adding weather-related variables brings little benefits, because its effects might take a longer time to affect the demand. On this work, we experiment a large amount of univariate and multivariate methods aiming to compare its performance over a dataset from the city of Rio de Janeiro. For these data, we show that is possible to obtain, via a simple linear univariate method (a load curve model where the standard load is forecast by the Holt-Winters-Taylor smoothing), results that are close enough to those achieved by more complex techniques, but bringing the advantages of more robustness, parsimony and computational economy.