Short-term load forecasting by artificial neural networks specified by genetic algorithms – a simulation study over a Brazilian dataset

Abstract

This paper studies the application of genetic algorithms in helping to select the proper architecture and training parameters, by means of evolutionary simulations done on a series of real load data, for a neural network to be used in electric load forecasting. Particularly, we investigate the application of a novel fitness function to the genetic algorithms, instead of the usual ones, based on the sum of the squares of the errors. We compare the results of the neural networks thus specified with that of four benchmarks: two naive forecasters, a linear method, and a neural network in which the parameter values are found by means of a grid search.