Detect Anomalies in Photovoltaic Systems Using Isolation Forest (Preliminary Results)

Abstract

Solar energy from Photovoltaic (PV) systems is one of the greatest growing renewable sources of energy. PV systems, although generally quite robust, are subject to failures that can adversely affect energy conversion or even pose safety concerns. Being able to promptly and automatically detect fails/anomalies is essential to improve PV systems reliability while maintaining the expected efficiency. In this paper are investigated five unsupervised Machine Learning (ML) methods for PV plants’ faults detection. The tested methods include Isolation Forest (IF), one-Class Support Vector Machines, Local Outlier Factors, Deep Learning Autoencoders, Gaussian Mixtures Models (GMM). Dealing with anomalies in PV means to handle a highly unbalanced dataset, as anomalies are relatively rare events, hopefully. Moreover, it is usually quite challenging to have precise labels for each fault event. For this reason in this paper are analysed only unsupervised ML models, that does not require labelled data. However, being able to quantify models’ performance precisely is a quite challenging task as it requires expert support or labels existence, even if they can be imprecise. To accurately represents the model’s performance for such a highly unbalanced dataset are reported metrics more suitable for such a task as balanced accuracy, recall, F1, Matthews Correlation Coefficient and Cohen Kappa. For the reason outlined above regarding the available labels, the results summarised in this paper can be considered as preliminary and require a more suitable dataset with precise labels. These preliminary results show that GMM could be highly effective to operate in the PV anomaly detection field; however, some models as IF and Autoencoders, that have proven to be very effective in different but demanding fields, deserve further investigation.