Fault Detection in Solar Power Plants Based on Energy Production Data
Synopsis
This research addresses the critical need for the timely identification of faults in solar power plants to minimize electricity loss. The study analyses energy production data from a Lithuanian solar power plant, comprising 143 strings distributed across 12 inverters, over a 19-month period. During data preprocessing, 16 key features were extracted from each string's time series data to represent the global structure of the data. The extraction process resulted in a transformed dataset, where each time series is represented as an object with 16 features, enabling for more effective analysis. Statistical and machine learning techniques — including PCA + $\alpha$-HULL, Isolation Forest (iForest), and Local Outlier Factor (LOF) — were employed to identify systems exhibiting abnormal behavior. The results demonstrate that a combination of these methods can help effectively identify outliers, with a combined anomaly score providing a comprehensive assessment of string performance. Additionally, RANSAC and DBSCAN methods were used to construct fault profiles, which enabled a more in-depth analysis of each system's performance and provided further confirmation of previously identified systems exhibiting abnormal behavior.






