Benchmarking a Physics-Based Approach for Anomaly Detection at Utility PV Plants

• Main Authors: Sheppard, Scott, Dickey, Keith A., Koskey, Steven, Teasley, Corson, Perullo, Christopher, Fregosi, Daniel, Li, Wayne
• Format: Conference Proceeding
• Language: English
• Subjects: Anomaly detection; Benchmark testing; Fault detection; Hardware; Monitoring; Pattern recognition; Photovoltaic systems; Software; Training data
• ISSN: 2995-1755
• DOI: 10.1109/PVSC57443.2024.10749158

Many utility monitoring and diagnostic centers have adopted advanced pattern recognition software to aid in anomaly detection and diagnosis. Due to the wide variety of electricity generation methods and associated supporting hardware, utilities choose software that is applicable to the broad category of industrial hardware. As a result, these tools excel at detecting large deviations from normal operation but struggle to identify subtle shifts in performance that are indicative of the onset of degradation and failure. At worst, these tools can be oversensitive and raise false alarms when the deviations are explained by operation outside of what was observed in the tool's training data. Recent developments in physics-based modeling have resulted in models that are capable of accurately detecting faults in the DC collector field that, individually, results in a less than 5% power loss at the combiner box level. These new models are benchmarked against current state-of-the-art utilities tools, with models designed to match the physics-based approach as much as is feasible. The applied physics-based models improve fault detection capabilities over the standard utility tool, detecting approximately twice as many real faults for a given false positive rate.