The presence of an air gap between a photovoltaic (PV) module and roof facilitates ventilation cooling under the device and consequently reduces cell temperature and improves its performance. In case of rack-mounted PV installation, the Nominal Operating Cell Temperature (NOCT) method could be effectively used to predict the temperature of the module for various environmental conditions. Many countries, for esthetic purposes, offer economic advantages (tax deductions, incentives, etc ... ) for the installation of building integrated photovoltaic modules (BIPV), with water-tightness capability and adequate mechanical resistance in order to substitute tile covering or part of it. Nevertheless, poor or absent ventilation under BIPV panels could cause them to overheat and reduce their efficiency. Lack of validated predictive tools for the evaluation of BIVP energy performance could be another barrier to their widespread application. In this study, we investigated the thermal performance of PV modules installed in a real scale experimental building over a traditional clay tile pitched roof in Italy for almost one year (from August 2009 to June 2010). One PV module was rack-mounted over the roof covering with a 0.2 m air gap; the others were fully integrated and installed at the same level of the roof covering (one with an air gap of 0.04 m, the other mounted directly in contact with the insulation). Temperature and heat flux measurements for each panel, and environmental parameters were recorded. Two temperature prediction models, NoCT model and SNL (Sandia National Laboratory) model were used to predict BIPV temperature and energy efficiency so that their suitability for BIPV could be evaluated. SNL model takes into account also the wind speed. Experimental results demonstrate that even though the rack-mounted PV module constantly maintains cell temperature below that of the other full-building integrated modules, due to the presence of a higher air gap, the difference in the energy produced by the BIPV modules estimated for the entire monitoring period is less than 4%. The two predictive models, NOCT and SNL, cause the differences in predicted and calculated temperature up to 10 degrees C. However, subsequent percentage variations on the energy predicted compared to that arising from the temperature measured generally turn out to be lower than 5%. An optimization of empirical coefficients used for calculations based on the SNL method allows for the reduction of this value below 2.5%. (C) 2014 Elsevier Ltd. All rights reserved.