Anomaly detection in photovoltaic systems using unsupervised learning (machine learning and neural networks)

Abstract

This paper entitled “Anomaly detection in photovoltaic systems using unsupervised learning (machine learning and neural networks)” provides an overview of photovoltaic installation monitoring through artificial intelligence techniques, addressing aspects related to early detection of faults and degradation.

First, the author emphasizes the importance of unsupervised machine learning in the face of the insufficiency of traditional visual inspection techniques, as a basis for identifying unusual patterns without requiring labeled data from previous failures.

One of the central aspects of the work is the implementation of algorithms such as Isolation Forest, DBSCAN and autoencoder neural networks on power, voltage, current and solar irradiance variables, demonstrating that autoencoders show the best overall performance, with detection rates above 85%. This enables applications in identifying module failures, partial shading, inverter degradation and unplanned disconnections.

Additionally, principal component analysis (PCA) results are presented, demonstrating the separation between normal and abnormal operating states in the reduced feature space.

Finally, the study discusses challenges related to the interpretation of system-generated alerts, including identifying the most relevant indicators for each type of anomaly, and highlights the importance of integrating these models into real-time monitoring platforms to promote more sustainable energy management.