Improving Solar Energy Reliability with Data-Driven Anomaly Detection Techniques

This study investigates unsupervised machine learning (ML) for anomaly detection in solar photovoltaic (PV) power generation data from 2019 to 2023. An unsupervised approach is selected to overcome the absence of pre-labeled fault data, enabling the autonomous identification of operational patterns. Following data preparation, K-means clustering (k=3) identifies distinct operational patterns, specifically characterizing regimes such as optimal performance (Cluster 2) and low energy output attributed to adverse weather conditions (Cluster 1). These clusters are subsequently visualized using principal component analysis (PCA) to validate their distinct separation. An isolation forest model is then employed for anomaly detection, identifying 17 significant deviations. These anomalies occur most frequently in 2020, coinciding with the COVID-19 pandemic period. Many fall outside the typical energy range of 2.0–3.2 kWh/day and are associated with non-ideal weather conditions. This finding demonstrates that unsupervised ML provides a scalable framework for monitoring PV system health, enhancing reliability, and supporting preventive strategies.