Examining performance of aggregation algorithms for neural network-based electricity demand forecasting

Abstract

• An ensemble of individual neural network is constructed with heterogeneous architecture. • The individual neural networks are combined using three aggregation algorithms. • Best models are selected in BESTcomb based on their MAPE. • The trim-mean is extended in TRIMcomb to determine the optimal value of trimming. • In BMAcomb the coefficients for combination are computed using the BMA method. The aim of this research is to examine the efficiency of 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 NN models are combined by three different aggregation algorithms. These aggregation algorithms comprise of a simple average, trimmed mean, and a Bayesian model averaging. These methods are utilized with certain modifications and are employed on the forecasts obtained from all individual NN models. The output of the aggregation algorithms is analyzed and compared with the individual NN models used in NN ensemble and with a Naive approach. Thirty-minutes interval electricity demand data from Australian Energy Market Operator (AEMO) and the New York Independent System Operator’s web site (NYISO) are used in the empirical analysis. It is observed that the aggregation algorithm perform better than many of the individual NN models. In comparison with the Naive approach, the aggregation algorithms exhibit somewhat better forecasting performance.