Field photovoltaic measurements exhibit nonlinear environmental variation, heterogeneous module responses, and scarce verified fault labels, making fixed-threshold anomaly screening difficult. This study develops an unsupervised framework for operating-regime and anomaly-candidate discovery from a limited field dataset consisting of only 81 observations collected from three photovoltaic modules over five measurement days under Incremental Conductance MPPT. Ten environmental, temporal, electrical, and physics-related features were robustly normalized using the median and interquartile range and evaluated with K-Means, Fuzzy C-Means, Gaussian Mixture Model, DBSCAN, and K-Medoids. Model selection combined internal validity, minimum cluster-size constraints, repeated-run stability, ordinal-rank consensus, and composite anomaly scoring. Among the tested configurations, DBSCAN achieved the highest internal validity, with a Silhouette coefficient of 0.3751 and a Davies–Bouldin index of 0.8749, while K-Medoids showed high repeated-run stability (mean ARI=0.9655) and strong agreement with DBSCAN on non-noise observations (ARI=0.9301). The consensus identified Low, Moderate, and High operating regimes containing 16, 26, and 39 observations, respectively. The mutually exclusive anomaly categorization identified five high-score anomaly candidates, seven density-supported candidates, and 69 non-anomalous observations. Because verified fault labels and independent full I–V sweep references were unavailable, the selected observations are interpreted as candidates for targeted physical inspection rather than confirmed photovoltaic faults. The framework therefore provides a structured basis for exploratory screening of field photovoltaic measurements.
Copyrights © 2026