New Deep Learning Framework for PV Inverter Fault Diagnosis Developed at Georgia Southern University
In a significant advancement for the renewable energy sector, scientists at Georgia Southern University have pioneered a deep learning framework aimed at improving fault detection in photovoltaic (PV) inverters. This innovative approach leverages both spatial and temporal attention mechanisms to enhance the robustness and interpretability of fault diagnosis, marking a notable milestone in the ongoing quest for reliability in solar energy systems.
The Dual Graph Attention Network: An Overview
The core of this new methodology is the Dual Graph Attention Network (DualGAT), which integrates two distinct attention mechanisms: DisGAT for spatial attention and TempGAT for temporal attention. This dual approach allows for a thorough analysis of complex signal interactions over time, particularly under varying environmental factors such as irradiance and temperature, as explained by Jakir Hossen, the corresponding author of the study.
Hossen highlights the uniqueness of this framework, stating, “Our work introduces a DualGAT that combines both spatial and temporal attention mechanisms for the first time in PV inverter fault diagnosis.” This innovation enables the model to capture intricate relationships among signals and their dynamic evolution, providing greater assurance in fault detection than previous methodologies.
Data Collection and Simulation
To train the DualGAT model, the research team employed MATLAB/Simulink to create a simulated PV inverter system. The setup included a PV source linked to a two-level, three-phase inverter, complemented by a DC-DC boost converter on the source side. The study considered two primary fault types: single and double IGBT (Insulated-Gate Bipolar Transistor) open-circuit faults.
As a part of experimentation, the team measured three-phase current signals while systematically varying the irradiance and temperature. The irradiance was altered in increments of 1 W/m² within the range of 250 W/m² to 750 W/m², and the temperature was adjusted from 25°C to 35°C in 1°C increments. The resulting dataset consisted of 121,242 samples across 22 classes, including a representation of normal operating conditions.
Constructing the Fault Detection Framework
The novel framework constructs two types of graphs: a spatial graph that displays relationships among faulted switches and a temporal graph that illustrates the sequence in which faults evolve. The interaction of these graphs allows for a comprehensive understanding of how faults operate both in space and time. With an 80-20 split of the data—80% for training and 20% for testing—the model effectively predicts which of the 22 faulty conditions the inverter might experience.
Performance Evaluations Against Competing Methods
To rigorously assess the DualGAT’s performance, the research team conducted comparative tests against a variety of traditional data-driven and statistical-based fault detection methods, including ANNs, CNNs, RNNs, GATs, GRUs with attention mechanisms, TCNs, Transformers, ResNet-1D, and additional approaches like SVM and Random Forest.
Remarkably, the DualGAT model achieved a test accuracy of 97.35%, outperforming other neural network models and significantly surpassing traditional methods. For instance, the GAT and RNN models achieved accuracies of 95.18% and 94.12%, respectively, while traditional models like SVM posted lower accuracies of 85.37%.
Understanding Model Robustness Through Ablation Studies
To further validate the framework’s capabilities, the scientists performed ablation studies where individual components of the model were removed to assess their contribution to overall accuracy. They found that omitting the DisGAT component alone resulted in a drop to 91.27%, while removing TempGAT brought it down to 87.62%. Other essential components, such as the regularizer and the cross-attention mechanism, also played crucial roles in achieving high accuracy levels.
Publication and Collaborative Efforts
The findings from this research have been documented in detail in the article “Dual graph attention network for robust fault diagnosis in photovoltaic inverters,” published in Scientific Reports. The project was a collaborative effort that included experts not only from Georgia Southern University but also from Cornell University, the University of Rajshahi, Rajshahi University of Engineering and Technology, and Multimedia University in Malaysia.
These collective efforts pave the way for more reliable and advanced fault detection methods in solar energy systems, potentially transforming the efficiency and effectiveness of solar power operations worldwide.