Resumo:
This work proposes a new approach to short-term load forecast for power distribution
substations using the artificial neural network ensemble. In this sense, the main objective
of this approach is to make predictions for the same time series using different tools that
are competent for this type of problem and then combine the solutions, obtaining a better
solution compared to those tools used individually. For the construction of the ensemble,
the experimental planning methodology (DOE) was used initially to identify the influence
of 6 factors related the parameterization of the ANN and from the desirability optimization
method to obtain a parameterization to determine the architecture of the neural networks
that formed the ensemble . Then, the normal intersection optimization method (NBI)
combined with the technique of exploratory factor analysis (based on mixture design of
experiments) was used to establish a set of optimal Pareto solutions for combining the
outputs produced by neural networks, forming the output of the ensemble. As a criterion
of choice, the maximum ratio between Shannon’s entropy and global percentage error was
used and based on the 72 out of sample data the ensemble of artificial neural networks
presented better results compared to each individual network.
