Novel data-driven health-state architecture for photovoltaic system failure diagnosis

Accurate and cost-effective diagnosis and prognosis of photovoltaic (PV) system failures is crucial for prolonged operational efficacy and minimizing operation and maintenance costs. A key challenge in this field remains the absence of accurate, transferable, and location-independent data-driven PV diagnostic algorithms. This study addresses this fundamental challenge by proposing a unified PV system health-state architecture to predict common array failures. The proposed architecture comprises data quality routines, digital twin models, and artificial intelligence-driven failure diagnostic algorithms. The proposed architecture was validated using historical data from PV systems in hot and cold climates, demonstrating scalability and location-independency. The digital twin predictive models exhibited less than 2 % errors, while the failure diagnostic algorithms showed detection accuracies above 90 % for faults with magnitudes > 8 %. The classifiers proved robust in diagnosing commonly exhibited faults, achieving classification accuracies > 95 %. Finally, valuable information is supplied to enhance performance monitoring systems through automated functionalities that leverage analytics for utilityscale PV plants transitioning into the smart grid era.