Occupant's behavioural patterns determine a significant level of uncertainty in building energy performance evaluation. It is difficult to account for this uncertainty in the design phase when operational and occupancy profiles are unknown. The relevant "performance gap" usually encountered between simulated and measured energy performance is clearly connected to biased assumptions in modeling, especially in the initial design phase. A probabilistic modeling approach is proposed to improve simulation reliability and robustness with respect to variability in occupancy patterns. The case study presented is the eLUX lab of the "Smart Campus" of Brescia University in Italy. Occupancy dependent input parameters such as air change rates (i.e. mechanically controlled ventilation) and internal heat gains (i.e. due to people, lighting and appliances) are described by means of probability distributions to obtain probabilistic thermal demand and load profiles as output. Probabilistic results enables a more reliable identification of energy saving strategies (operational and environmental settings) with respect to highly variable operating conditions. Further, simulation data are processed to obtain a weather-adjusted energy demand visualization, suitable for establishing a continuity between modeling in design and operation phases, with calibration purpose. Calibrated energy models can be used for several specific tasks in the operation phase, in particular condition monitoring, fault detection and diagnosis, supervisory control and energy management. For the case study presented, a detailed data acquisition scheme has been designed to enable an effective monitoring activity in the operation phase, aimed at experimenting model-based approaches for the tasks reported. The proposed research is the point-of-departure for a general activity aimed at assessing critically the issues of reliability and robustness of simulation results obtained with conventional modeling approaches, in particular with respect to occupants' behaviour, exploiting at the same time the possibility of using measured data as a direct feedback to promote behavioural change. (C) 2016 Elsevier B.V. All rights reserved.