Enhanced ensemble structures using wavelet neural networks applied to short-term load forecasting

Abstract

Load forecasting implies directly in financial return and information for electrical systems planning. A framework to build wavenet ensemble for short-term load forecasting is proposed in this work. For this purpose, data are first transformed for trend removal and normalization, then an optimal time window is calculated and a subset of features is selected. The bootstrapping, cross-validation like, inputs decimation, constructive selection, simple mean, median, mode and stacked generalization algorithms are used for the ensemble aggregation of wavenet learners. Predictions are realized through one step ahead forecasting strategy. Hourly load values from Italy in 2015 and the GEFCom competition (Global Energy Forecasting Competition) 2012 are used to test and compare the proposed framework with existing similar forecasting techniques such as a multilayer perceptron neural network with sigmoid activation functions in the hidden layer, a single wavenet, a regression tree approach, and the forecasting based on the last week mean. Cross-validated results using 10-folds demonstrate the effectiveness of the proposed forecasting framework based on wavenet ensemble, overcoming performance of the models compared.