Comparison and Application of Data Science Techniques for Anomaly Detection in Photovoltaic Systems

Photovoltaic (PV) systems in Brazil have significantly increased in the past decade, driven by resolutions, bills, and reduced technology costs. With this growth, ensuring return on investment and monitoring anomalies in PV technologies has become crucial. These anomalies, detectable via machine learning (ML) algorithms, range from issues in PV modules to energy processing equipment (power converters). Therefore, this paper, based on actual data from a PV plant, utilizes ML algorithms to evaluate their effectiveness in these scenarios. Specifically, the methods (i) Isolation Forest, (ii) One-class SVM, and statistical approach (iii) Interquartile Range were tested. The results showed that the One-Class SVM model performed better in detecting anomalies, surpassing the Interquartile Range method by 26% in detecting normalities. In conclusion, it is apparent that these models have potential for further improvements, considering the possibility of incorporating a more extensive history of plant measurements and meteorological data.