Performance Analysis of Bagging Feed-Forward Neural Network for Forecasting Building Energy Demand

Abstract

Forecast models play a fundamental role in anticipating the effects of the energy demand in buildings to addressing the energy crisis. A forecast model for anticipating from one to three days every 30 min of the building energy demand is presented. In this model, a feed-forward artificial neural network (ANN) is combined with bootstrap aggregation techniques, using a Box–Cox transformation, seasonal and trend decomposition using loess, and a moving block bootstrap technique. An analysis was conducted using the data provided by a building’s energy demand; the data were collected during a period of four months, with readings every 10 s and averages of the values obtained every 30 min. The feed-forward neural-network method combined with bootstrap aggregation techniques consistently outperformed the forecasting accuracy of the original feed-forward neural network through cross-validation in the root mean square error (RMSE) and the mean absolute percentage error. From cross-validation in-sample period, used for the initial parameter estimation and model selection, it is concluded that a feed-forward neural network with the original data gives a slightly lower RMSE compared with the data generated by bootstrapped versions. However, in the cross-validation out-of-sample period used to evaluate forecasting performance, it has better consistency throughout all horizons for the ANN model combined with bootstrap aggregation techniques than for the ANN original model. The results are statistically significant according to the Ljung–Box test, which verifies that the forecast errors are not correlated and validates the proposed model.