Neural Network Based Approaches for Fault Diagnosis of Photovoltaic Systems

Abstract

Faults in photovoltaic (PV) systems due to manufacturing defects and normal wear and tear are practically unavoidable. The effects thereof range from minor energy losses to risk of fire and electrical shock. Thus, several PV fault diagnosis techniques have been developed, usually based on dedicated on-site sensors or high-frequency current and voltage measurements. Yet, implementing them is not economically viable for common small-scale residential systems. Hence, we focus on cost-effective techniques that enable introducing fault diagnosis without incurring costs for on-site sensor systems. In this chapter, we will present in particular two machine-learning-based approaches, built on recent neural network models. The first technique relies on recurrent neural networks (RNNs) using satellite weather data and low-frequency inverter measurements for accurate fault detection, including severity estimation (i.e., the power loss caused by the fault, usually not quantified in state-of-the-art methods in literature). The second technique is based on graph neural networks (GNNs), which we use to monitor a group of PV systems by comparing their current and voltage production over the last 24 h. By comparing outputs from multiple (geographically nearby) PV installations, we avoid any need for additional sensor data. Moreover, our results suggest that the GNN-based model can generalize to PV systems it was not trained on (as long as nearby sites are available) and retains high accuracy when multiple PV systems are simultaneously affected by faults.