Unsupervised Graph-Generative Network-based PV Condition Monitoring Systems

Condition monitoring plays a pivotal role in the photovoltaic (PV) system's operation. To overcome issues in conventional PV condition monitoring, in this paper, an unsupervised graph-generative network-based PV condition monitoring system is proposed by considering Gaussian/non-Gaussian noises. The designed network has the ability to handle an imbalanced dataset and high dimensional signals, establish the correlation between time-varying signals in PV systems, and learn spatial-temporal features. The proposed system is validated through an experimental dataset by considering seven abnormal conditions, including PV array faults, PV control system faults, and grid faults. The proposed method appears to be effective under Gaussian/non-Gaussian noises with an accuracy higher than 98%. The superiority of the proposed graph-generative network is demonstrated by comparing with well-known deep- and shallow-based PV condition monitoring systems with 5%- 30% improvements in terms of accuracy and reliability.