Ressources

Oumayma Mabrouk, Abdérafi Charki, Nizar Chatti, Xavier Sidambarompoulé, Robust reference curves generation for photovoltaic systems performance diagnosis using a data-driven approach, Solar Energy, Volume 317, 2026, 114958, ISSN 0038-092X, https://doi.org/10.1016/j.solener.2026.114958.

(Téléchargez l'article : https://www.sciencedirect.com/science/article/pii/S0038092X2600647X)

Abstract: The rapid deployment of photovoltaic (PV) systems demands reliable and scalable methods for early detection of abnormal events. Most existing fault diagnosis approaches rely on supervised learning, requiring large labeled datasets and often resulting in black-box models with limited physical interpretability. This paper proposes a regime-aware, fully data-driven framework for anomaly detection in PV systems that models normal PV generation behavior without fault labels. The main contribution of this work is the explicit separation between meteorological regime identification and power-based anomaly detection to prevent degraded production from biasing the reference model. Daily power curves are normalized and clustered into representative weather regimes (clear sky, partially cloudy, overcast) using k-means based solely on environmental variables. For each regime, a robust reference corridor is constructed from the median and interquartile range (IQR), capturing intrinsic variability while remaining resistant to outliers. New daily profiles are aligned to the appropriate regime using Limited Dynamic Time Warping (LDTW) to ensure physically consistent comparison. Two complementary indicators—a confidence score and an IQR-based deviation ratio—are then used for anomaly detection. Thresholds are calibrated via Receiver Operating Characteristic (ROC) analysis on operational data with synthetically injected faults, achieving an area under the curve exceeding 0.80. The proposed framework provides an interpretable and scalable solution for large-scale PV performance monitoring.

Keywords: Photovoltaic systems; Anomaly detection; Unsupervised learning; Meteorological regimes; K-means clustering; Time-series analysis; Reference curve modeling; Performance monitoring.