The installed capacity of PV plants has increased dramatically in the past years. A common approach to determine the actual power of an ensemble of PV systems within a specific region typically employs data from measured reference plants. Obviously the precision of the power estimation depends on having representative reference plants, which are not influenced by strong individual characteristics. The goal of this contribution is to detect such apparently atypical behavior of PV systems by comparing their measured power to simulations based on a nearby weather station and clear sky irradiance. Deviations are studied in the course of each day for the year 2012 and 48 PV systems, indicating systematic characteristics independent from meteorological conditions. Additionally, an approach is presented to detect such unexpected deviations automatically. This can be the basis for a dynamic nowcasting algorithm, which selects the reference units based on their (temporal) suitability.