New deep learning tech for PV inverter fault diagnosis

A team of scientists at Georgia Southern University has combined both spatial and temporal attention mechanisms to develop a new approach for PV inverter fault detection. Training the new method on a dataset created in MATLAB/Simulink, the group has compared it to a series of other data-driven and statistical-based methods and has found accuracy reached 97.35%.

September 3, 2025 Lior Kahana

Image: Claus Ableiter, Wikimedia Commons, CC BY-SA 3.0

From pv magazine Global

A research team led by scientists from the United States’ Georgia Southern University has developed a novel deep learning framework for fault diagnosis in PV inverters.

For this purpose, the scientists utilized a dual graph attention network (DualGAT), which combines both spatial and temporal attention mechanisms, specifically the DisGAT and TempGAT, respectively.

“Our work introduces a DualGAT that combines both spatial and temporal attention mechanisms for the first time in PV inverter fault diagnosis,” corresponding author Jakir Hossen told pv magazine. “This dual-level approach enables the model to capture complex signal correlations and evolving dynamics under varying irradiance and temperature conditions, ensuring greater robustness and interpretability compared to prior methods.”

To collect data on which the system is to be trained, the group simulated a PV inverter system in MATLAB/Simulink. It consisted of a PV source connected to a 2-level, 3-phase inverter on the grid side, with a DC-DC boost converter positioned on the source side. The inverter comprised six insulated-gate bipolar transistor (IGBT) switches, each paired with an antiparallel diode. Two types of open-circuit faults were considered, namely single IGBT open-circuit faults and double IGBT open-circuit faults.

“The three-phase current signals are measured as the irradiance and temperature of the PV array vary to evaluate the performance of the proposed method. The irradiance is adjusted in increments of 1 W/m2, ranging from 250 W/m2 to 750 W/m2, while the temperature is varied from 25 C to 35 C in 1 C intervals,” the group explained. “Each current signal contains 5,511 sample sets, with three current signals corresponding to the three phases for each fault type. In total, the dataset comprises 121,242 samples across 22 classes, including the normal operating condition.”

Using this data, the novel framework constructed a graph of spatial fault relationships to illustrate how faulty switches interact with one another, along with a temporal graph to depict how faults evolve in sequence. Then, it combined both graphs to see how faults interact in space and time, and based on that, it predicted which one of the 22 faulty conditions the inverter is in. From the simulated data, 80% was used for training, and the rest was used for testing.

The system was tested against competing fault detection methods. They tested it using data-driven approaches as well as statistical-based methods, namely ANN, CNN, RNN, GAT, GRU + Attention, TCN, Transformer, ResNet-1D, InceptionTime, LightGBM + SHAP, SVM, KNN, RF, DT, and BC.

Among the neural-network-based methods, the proposed DualGAT model was found to achieve the highest results across all metrics, with a test accuracy of 97.35%, which the scientists said demonstrates its robust ability to capture both spatial and temporal fault patterns. “Other temporal models, such as GAT and RNN, also exhibit strong performance, with accuracies of 95.18% and 94.12%, respectively, surpassing traditional methods like RF and SVM, which achieve accuracies of 87.11% and 85.37%,” they added.

Furthermore, the academics conducted ablation studies of the method, which involve removing parts of the models. Without DisGAT, the accuracy dropped to 91.27%; without TempGAT, it fell to 87.62%; without the regularizer, it was 90.13%; and without the cross-attention component, it was 92.51%.

The new framework was presented in “Dual graph attention network for robust fault diagnosis in photovoltaic inverters,” published in Scientific Reports. The team included researchers from the United States’ Georgia Southern University, Cornell University, Bangladesh’s University of Rajshahi, Rajshahi University of Engineering and Technology, and Malaysia’s Multimedia University.

This content is protected by copyright and may not be reused. If you want to cooperate with us and would like to reuse some of our content, please contact: editors@pv-magazine.com.