Neural network model for building electric energy consumption forecasting in densely populated tropical areas

Abstract

Electric energy consumption forecasting is a relevant issue to design and implement public policies related to energy generation and distribution at urban scale. This problem has been addressed from different standpoints, including traditional time series analysis and machine learning techniques, focused on the prediction of future energy demand according to metered data. This work proposes a hybrid model, based on computer simulation results from the City Energy Analyst toolbox and metered environmental and energy consumption data from a representative set of buildings located in Singapore. Such model is intended to provide reliable energy demand forecasts by using regression models. Three different regression methods were evaluated: a traditional SARIMA model and both NARX and Recurrent neural network architectures. The results obtained using this approach point out that RNN models provide accurate forecasts for 1 and 24 hours, outperforming NARX based models, while the SARIMA is, in general, unable to represent the electrical demand time series patterns.