Classification of anomalies in photovoltaic systems using supervised machine learning techniques and real data

Abstract

In a photovoltaic (PV) plant, various types of anomalies can lead to a decrease in energy conversion efficiency. These anomalies represent deviations from the normal behavior of the plant. Understanding the system’s behavior and identifying deviations from normality are crucial for implementing preventive and corrective measures to ensure the expected economic return on investment in the plant. One effective approach for anomaly detection and classification is the utilization of Supervised Machine Learning Techniques (SMLT). However, this approach comes with several challenges, including the classification of the training dataset, temporal variations, uncertainties, model interpretability, data scarcity, and generalization of the model to different systems and locations, among others. In this way, this study aims to explore the application of SMLT in PV systems and compare different methods. To achieve this, a methodology was developed to create a training and testing synthetic dataset based on real irradiance data, followed by a new process flow for ensemble SMLT. Finally, the new process flow for ensemble SMLT was tested on a real PV plant and synthetic dataset. The methodology for the proposed algorithm was an ensemble of Random Forest with K-nearest neighbors (k-NN) and an inference machine for specific classes. The results indicate that the algorithm successfully classified anomalies, achieving an AUC of 0.9815 for the synthetic dataset and an AUC of 0.9861 for the real dataset, as well as an accuracy of 0.9647 for the real dataset. It is evident that SMLT can serve as valuable tools for detecting and addressing issues, particularly due to the abundance of data and the diverse characteristics inherent to PV plants.