Evaluation of unsupervised anomaly detection approaches on photovoltaic monitoring data

Abstract

Fault detection in photovoltaic is a challenging task, our project <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">"</sub> PV Digital 4.0” addreses this challenge by using machine learning techniques on monitoring data from different systems to detect faults on photovoltaic modules. This paper is embedded in this project and addresses faults for which we have little data or which may occur in the future. For this we use different state of the art unsupervised anomaly detection approaches (e.g. LSCP, VAE or MOGAAL) and compare them in terms of their applicability to our problem. We combine the monitoring data of 70 PV plants with weather data and preprocess them for the usage in multiple models. Each model is elaborated with the same dataset and afterwards compared according to its accuracy. In our scenario we only want to detect new occurring faults and not long term existing faults like shadows. The models should be able to detect the actual behavior of the PV system.