Dual graph attention network for robust fault diagnosis in photovoltaic inverters

This paper presents a novel deep learning framework based on a Dual Graph Attention Network (DualGAT) to enhance the accuracy and robustness of fault diagnosis in photovoltaic (PV) inverters operating under diverse environmental and operational conditions. Given the critical role of PV inverters in ensuring stable energy conversion, early and reliable detection of open-circuit faults is essential to prevent performance degradation and equipment failure. To address this, a detailed simulation model of a grid-connected PV inverter was developed in MATLAB/Simulink, incorporating variations in irradiance and temperature to generate realistic fault scenarios. Discrete Wavelet Transform (DWT) was employed to extract energy-based fault signatures from the inverter's current signals, forming a rich dataset for model training. The proposed DualGAT architecture combines spatial and temporal attention mechanisms through two complementary modules-DisGAT and TempGAT-enabling the model to learn both structural dependencies and time-evolving patterns of inverter faults. Experimental results show that the model achieves a test accuracy of 97.35%, significantly outperforming traditional machine learning and recent deep learning approaches. Furthermore, the model demonstrates strong resilience under noisy conditions, maintaining high diagnostic performance even with signal distortion. These findings underscore the effectiveness of the DualGAT framework in capturing complex spatio-temporal fault characteristics, offering a promising solution for intelligent condition monitoring in PV-based power systems.