Neural Networks Ensemble: Evaluation of Aggregation Algorithms for Forecasting

The aim of the thesis is to examine and analyze different aggregation algorithms to
the forecasts obtained from individual neural network (NN) models in an ensemble. In
this study an ensemble of 100 NN models are constructed with a heterogeneous
architecture. The outputs from the individual NN models were combined by four
different aggregation algorithms in NNs ensemble. These algorithms include equal�weights combination of Best NN models, combination of trimmed forecasts,
combination through Variance-Covariance method and Bayesian Model Averaging.
The aggregation algorithms were employed on the forecasts obtained from all
individual NN models as well as on a number of the best forecasts obtained from the
best NN models. The output of the aggregation algorithms of NNs ensemble were
analyzed and compared with each other and with the individual NN models used in
NNs ensemble. The results of the aggregation algorithms of NNs ensemble are also
compared with the Simple Averaging method. The performances ofthese aggregation
algorithms ofNNs ensemble were evaluated with the mean absolutepercentage error
and symmetric mean absolute percentage error.
In the empirical analysis, the methodologies developed were tested on the
Universiti Teknologi PETRONAS load data set of five years from 2006 to 2010 for
forecasting. It can be concluded from the results that the aggregation algorithms of
NNs ensemble can improve the accuracy of forecast than the individual NN models
with a test data set. Furthermore, in the comparison with the Simple Averaging
method, the aggregation algorithms of NNs ensemble demonstrate slightly better
performance than the Simple Averaging. It has also been observed during the
empirical analysis that; reducing the size of ensemble increases the diversity and,
hence, accuracy. Moreover, it has been concluded that more benefits can be achieved
by the utilization of an advanced method for forecast combinations.