Reliable fault detection in photovoltaic (PV) arrays, precisely and at large, is necessary to guarantee energy efficiency, system longevity, and grid stability. Traditional machine learning methods face limitations in capturing both spatial and temporal correlations inherent in PV systems, especially under compound fault conditions. In this paper, we propose a hybrid deep learning structure that incorporates a graph neural network for spatial topology learning and a Transformer with multi-head attention for temporal dependency modeling. We test the model with both real (Desert Knowledge Australia Solar Center Alice Springs, and a Belgium-based system) and synthetic fault data generated using Monte Carlo simulations and a single-diode model. The hybrid architecture achieves 98.6% classification accuracy and surpasses baselines of long short-term memory, convolutional neural network, and message passing neural network, and reduces inference time by 40%. An adaptive regression-based optimization method further enhances feature embedding and classification performance. The proposed architecture retains robustness under noisy and partially missing data scenarios and holds promise for application in real-time PV monitoring systems.
Pang et al. (Sat,) studied this question.