Broadly speaking, building energy use prediction can be classified into two categories based on modeling approaches namely engineering and Artificial Intelligence (AI). While engineering approach requires solving physical equations representing the thermal performance of systems and components that constitute the buildings, the AI-based approach uses historical data to predict future performance. Although engineering approach estimates energy use with greater accuracy, it falls short in the overall complexity of model building and simulation in which detailed data that represent the building geometry, systems, configurations, and occupant schedule is needed. Whereas, the AI-based approach offers a rapid prediction of building energy use and, if appropriately trained and tested, may be used for quick and efficient decision making of energy use reduction. Nevertheless, for robust integration with and to improve automated building systems management and intelligence, the need for consistent, stable, and higher prediction accuracy cannot be understated. To alleviate the instability issue, and to improve prediction accuracy, we have exploited and tested an ensemble learning technique, 'Ensemble Bagging Trees' (EBT), using data obtained from meteorological systems and building-level occupancy and meters Results showed that the proposed EBT model predicted hourly electricity demand of the test building with improved accuracy of Mean Absolute Prediction Error that ranged from 2.97% to 4.63%. Additionally, results showed that proposed variable selection method could reduce the computation time of EBT by 38-41% without sacrificing the prediction accuracy. The proposed ensemble learning model that exemplifies improved prediction accuracy over other AI techniques can be used for real-time applications such as system fault detection and diagnosis. (C) 2017 Elsevier B.V. All rights reserved.