Decoding pixels: A modular software prototype for cognitive image-based diagnostics of PV plants

Although aerial infrared (aIRT) imagery-based solutions for diagnostics of PV plants demonstrate impressive time-efficiency and spatial resolution, they also suffer from considerable drawbacks: limited automation (hence, expert dependence) and insufficient quantitative insights. In this paper, we introduce a software prototype, evolved from an innovative diagnostics framework researched and developed by CEA-INES over the last years, which integrates aIRT imagery with deep learning-based algorithms and physical/electrical modeling. With such an approach, unlike conventional ones, we worked on reaching both qualitative fault detection and quantitative (power loss) insights, with a focus on various spatial granularity levels within PV systems. Leveraging advanced deep learning techniques, first results show that we can achieve automated PV module detection and fault identification/classification, with associated power loss analysis at PV system, string/inverter, or module level. Further real-life validation efforts are underway, in utility-scale PV plants. Future developments aim to enhance further enhance our PV diagnostic framework, through data fusion with SCADA outputs and integration with maintenance and end-of-life (EoL) management tools.