Identifying High-Significance Latent Physical Anomalies in Solar Energy Systems

Data-driven physical anomaly detection for solar energy systems is studied. A fully unsupervised learning approach based on traces of solar generation data and weather data is developed. The idea is that, without needing any anomaly labels, a) predictors that output expected solar generation can be trained based on generation data under normal system operations, and b) a variety of anomalies can be identified based on analyzing the deviations between the actual and predicted solar generation. This paper focuses on identifying physical anomalies that are a) significant and sustained over long periods of time, yet b) "latent", i.e., can be easily missed by asset managers in practice. Two types of predictors - weather-based predictors and cross-inverter predictors - are developed that can provide complementary information in identifying major anomalies. Furthermore, conditional probabilities of the prediction errors are estimated for accurate probabilistic evaluation and statistical interpretations of anomalies. As such, conditional log-error-probabilities are employed as error metrics. Comprehensive experiments are conducted based on rich real-world solar energy data sets that span over 4+ years and across different states. It is demonstrated that the developed method successfully identifies a variety of high-significance physical anomalies that evade asset managers' attention for sustained periods from weeks to years.