Análisis de ciclo de vida de paneles solares con IA y visión para detectar fallas

Abstract

Thermal, physical, and electrical failures accelerate the degradation of solar panels' useful life, to the point of compromising their energy generation function and, in turn, the investment made in the photovoltaic system. In response to this need, the research developed and validated an embedded prototype aimed at monitoring the life cycle, capable of detecting external and internal failures with traceability and immediate visualization. An architecture was implemented on NVIDIA Jetson Xavier NX that ran artificial intelligence models and presented results with interaction on a local web panel. For external faults, the system integrated a thermal model for detecting hot spots and a multi-class visual model for identifying cracks, surface coverage, dust accumulation and normal status, where accuracies of 90% were recorded for cracks and 84% for the normal class, with intermediate performance for coverage and dust. For internal faults, Random Forest and Gradient Boosting were compared, with the latter being selected for offering more consistent predictions, achieving accuracies using the R² metric of 0.9997 and an average MAE error of 117.65 W, reflecting the impact of larger point errors. The solution incorporated inspection reports and cloud storage of Amazon web services using AWS S3 organized by users and plants, with anonymized identifiers and a status file that was updated when changes were made to the platform. In addition, a PETG printed casing was designed and printed to facilitate protection, internal organization and ventilation of the equipment. Finally, the prototype was validated in the field through data acquisition with a drone, demonstrating technical feasibility as a replicable tool for diagnosis and maintenance in photovoltaic installations. Keywords: photovoltaic system, artificial intelligence, failures, monitoring, embedded