A Self-Supervised Machine Learning Approach for the Estimation of Open-Circuit Voltage Degradation in Photovoltaic Systems

Abstract

In this work, we present a machine learning (ML) approach that integrates physics-based knowledge with data-driven techniques to estimate the open-circuit voltage ( $v_{oc}$) of photovoltaic (PV) systems. This self-supervised approach allows the detection of anomalies in operating systems, enabling the identification of potential faults and degradation related to $v_{oc}$ without requiring labelled data. Deviations in $v_{oc}$ values are detected by analysing the measurements recorded in the supervisory control and data acquisition (SCADA) system. This analysis is performed by a combination of clustering and regression algorithms. The proposed approach is validated on three different PV installations, showcasing its ability to detect variations in open-circuit voltage and to predict performance degradation. The proposed method achieves an $R$-squared (r2) value larger than 0.9 when trained on experimental $v_{oc}$ data from three distinct PV systems. Moreover, it estimates $v_{oc}$ from SCADA data with an average error below 5% compared with $I$– $V$ curve measurements. The results of this study demonstrate that physics informed ML techniques can significantly enhance the performance and reliability of PV systems, enabling early fault detection and degradation forecasting.