• SVR with DC performance ratio preprocessing estimates PV power from field data. • Instantaneous DC PR screening with K-means clustering improves data quality. • Clear-sky RMSE of 0.036 kW demonstrates high predictive accuracy of the model. • SVR-estimated degradation of 1.09%/yr aligns with independent STC result of 0.97%/yr. • First-year 11.6% power loss from PID is detected and confirmed by EL/IR imaging. Accurate power prediction and degradation diagnosis are crucial for maintaining long-term reliability of photovoltaic (PV) systems. However, this remains difficult due to the inherent variability of solar energy, unpredictable environmental conditions, and the high cost and complexity of monitoring utility-scale PV systems. This study presents a machine learning-based model for power estimation and fault detection in PV systems. The model was trained and validated using one year of real-world data from two grid-connected systems in Korea. Key input variables included generated power, plane-of-array solar irradiance, and module temperature. A support vector regression algorithm was employed for power prediction, with preprocessing based on instantaneous DC performance ratio to improve accuracy. The model achieved a root-mean-square error of 0.036 kW under clear-sky conditions, demonstrating high predictive performance. The annual degradation rate was estimated at 1.09%, closely matching independent standard condition measurements showing a yearly decline of 0.97%. The model also detected significant performance anomalies, including a power loss of 11.6% within the first year of operation in one system. On-site inspection attributed this to a potential-induced degradation effect. These findings underscore the model’s utility for predictive maintenance and potential as a scalable, data-driven tool for enhancing PV system reliability.
Oh et al. (Wed,) studied this question.