Graph neural networks for fault diagnosis of geographically nearby photovoltaic systems

Abstract

Faults in photovoltaic (PV) systems significantly reduce their efficiency and can pose safety risks. Nevertheless, most residential PV systems are not actively monitored, because existing methods often require expensive sensors, which are only cost-effective for large PV systems. Therefore, we propose a graph neural network (GNN) to monitor a group of nearby PV systems without relying on dedicated sensors. Instead, the GNN compares 24 h of current and voltage measurements obtained from the inverters. Four GNN variants are experimentally compared using simulated data of six different PV systems in Colorado. Results show that all GNN variants outperform a state-of-the-art PV fault diagnosis method based on gradient boosted trees. Moreover, some GNN variants can even generalize to PV systems which were not in the training data, enabling monitoring of new PV systems without retraining.