Residential photovoltaic (PV) systems have proliferated worldwide as cost-effective sources of clean energy, yet they are prone to latent faults and gradual performance degradation caused by fluctuating irradiance, temperature stress, and component ageing.
This study proposes a feature-driven approach for unsupervised fault detection in residential photovoltaic (PV) systems, emphasizing interpretability and model generalization. Using real-world operational data from a Swedish PV installation, the method integrates three stages: (1) feature selection via Random Forest, Lasso,and XGBoost; (2) temporal behavior analysis of the top-ranked features; and (3) anomaly detection using Isolation Forest, followed by SHAP-based interpretability. Experimental results show that “DC Power 1”, “Temperature[°C]”, and “DC Power 2” consistently rank among the most influential predictors across all models. Temporal analysis confirms these features exhibit stable and responsive trends under daily output fluctuations. The Isolation Forest model effectively detects fault-like anomalies without labeled data, and SHAP interpretations reveal consistent explanatory patterns aligned with both normal and abnormal operating conditions.
By combining unsupervised feature extraction with unsupervised anomaly detection, this study provides a transparent and generalizable solution for fault detection in residential PV systems.