The forecasting of the thermal power demand is essential to support the development of advanced strategies for the management of local resources on the consumer side, such as heating ventilation and air conditioning (HVAC) equipment in buildings. In this paper, a novel hybrid methodology is presented for the short-term load forecasting of HVAC thermal power demand in smart buildings based on a data-driven approach. The methodology implements an estimation of the building's activity in order to improve the dynamics responsiveness and context awareness of the demand prediction system, thus improving its accuracy by taking into account the usage pattern of the building. A dedicated activity prediction model supported by a recurrent neural network is built considering this specific indicator, which is then integrated with a power demand model built with an adaptive neuro-fuzzy inference system. Since the power demand is not directly available, an estimation method is proposed, which permits the indirect monitoring of the aggregated power consumption of the terminal units. The presented methodology is validated experimentally in terms of accuracy and performance using real data from a research building, showing that the accuracy of the power prediction can be improved when using a specialized modeling structure to estimate the building's activity. (C) 2018 Elsevier B.V. All rights reserved.