Monitored power output data of photovoltaic (PV) installations is increasingly used for purposes such as fault detection and performance studies of distributed PV systems. The value of such datasets can increase significantly when they are paired with information about local irradiance and shading conditions, especially in urban environments. However, on-site irradiance measurements are seldom performed for small or medium-sized rooftop PV installations. This paper proposes a novel method to identify locally shaded periods of PV installations, using only measured AC power, regional irradiance data and basic information about the sites (i.e. module tilt, orientation and nominal power) as inputs. The proposed three-step method uses machine learning techniques and a grey-box PV performance prediction model to classify the visible sky hemisphere of a PV installation to obstructed and unobstructed areas. Detailed results of a moderately-shaded residential PV site in the Netherlands are shown to illustrate the working principles of the method. Finally, a successful comparison with on-site shade measurements is carried out and the ability of the method to detect shade from nearby objects is illustrated.