Previsão de carga elétrica a curto prazo combinando métodos lineares univariados e variáveis exógenas através de algoritmos genéticos

Abstract

Short-term load forecasts are crucial to the operation of energy systems, since they allow optimizing operating costs and avoiding waste. In addition, the forecasts are needed to set the price of energy in the market and to define supply contracts. In this dissertation, we developed a system based on the combination of forecasts. First, univariate forecasts are made by four exponential smoothing methods (extensions of the Holt-Winters method); then, these forecats are combined linearly with each other, and also with some non-linear functions of exogenous variables (especially temperature), through genetic algorithms. We performed experiments with two different implementations of these algorithms and tested the system on a real database, obtained in the city of Rio de Janeiro. The results indicated that the introduction of exogenous variables did not improve the accuracy of the forecasts; in addition, that the combination using genetic algorithms did not get better forecasts than that made using an already well-known optimization method (in this case, the Nelder-Mead method). This shows that methods based on computational intelligence may not achieve better results, in all circunstancies, than the usual linear methods and that their application must be well justified in each particular problem.