Soiling identification and forecasting in urban environment

Urban Photovoltaics (PVs) plays an important role in harnessing power generation for sustainable development, but its performance is impaired under an increasing level of environmental pollutants. This paper presents the development and verification of a Stacked Ensemble Learning (SEL) and Decision Trees (DTs) based algorithm for forecasting PV output under conditions of significant air pollution. It uses a small dataset for training and validation, including data on the production of an urban PV system and meteorological data commonly available from local weather services. The developed algorithm was tested on data collected from an urban location in Niš, Serbia, which is exposed to significant air pollution during the regular heating season, for three different panel mounting configurations. The SEL model achieves a near-zero error metric in predicting PV yield for vertical monocrystalline panels, maintaining high accuracy even under soiling, which is a key feature for real-world applications. Similar conclusions can be drawn for the cases of horizontal and optimally positioned panels. Using the developed SEL algorithm, it is shown that the soiling ratio obtained from the estimation of PV production of clean and soiled panels can be used as a reliable indicator for scheduling the optimal period for cleaning PV panels.