Answers are grounded in this paper and cite their sources.
Experimental modeling study demonstrates varying detectability thresholds across irradiance profiles in photovoltaic systems, suggesting condition-specific distributions improve fault detection.
Photovoltaic (PV) output degradation is difficult to detect early using outdoor operating data because it can overlap with normal output variations caused by solar irradiance and module temperature. This study evaluated PV output degradation detectability using irradiance-condition-specific relative-error distributions constructed from 1-min electrical and meteorological data from a 3 kW PV testbed. Irradiance-condition classes were defined using KD and POPD to reflect the daily irradiance levels and variability. For each class, a class-specific reference distribution was constructed from the normal-state relative-error distribution using Gaussian or kernel density estimation (KDE)-based density models according to goodness-of-fit. Output degradation conditions were simulated by reducing the measured power by 10–50%, and the separability between normal and output degradation conditions was evaluated using the negative log-likelihood (NLL) and right-tail (RT) scores. Performance was analyzed using the area under the precision–recall curve (AUC-PR), area under the receiver operating characteristic curve (ROC-AUC), recall, and F1 score. The results showed that 10–20% degradation was highly detectable under high-irradiance and low-variability conditions, whereas 40–50% degradation was required under low-irradiance or highly variable conditions. The RT score improved the threshold-based detection performance in some variable classes by reflecting the directional rightward shift caused by output degradation. These findings indicate that PV degradation detection should use irradiance-condition-specific reference distributions and directional degradation information, rather than a single reference distribution.
No takes yet. Share an insight, caveat, or question.
Oh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: