The photovoltaic (PV) generation forecast is a key element to an efficient building energy management system (EMS) operation. The forecast's uncertainties and generation variabilities expose the loads to misplanning, and hence decrease building autonomy, self-sufficiency, and potential costs savings. In this paper, a novel approach for a day-ahead PV power generation probabilistic forecast is proposed that is especially optimized for building EMS applications. The model consists of several modules to develop the probabilistic forecast. Initially, a clear sky model is tuned to incorporate the system and temperature losses and partial shading. The deviation of the PV power from the clear sky power is used to train a bagging regression tree, which produces a deterministic point forecast. The probabilistic forecast is developed based on the probabilistic analysis of the point forecast and regenerating it based on the given weather conditions. The model is developed based on the available data in buildings such as the historic PV measurements acquired from the inverter and the weather forecasts. The probabilistic forecast was validated over a complete-year data set of a rooftop PV system in Munich, Germany, where the results showed its capability to provide an accurate and reliable forecast for EMS applications.