A team of scientists in the United States has combined both spatial and temporary attention mechanisms to develop a new approach for PV -Omverterer error detection. Training of the new method on a data set made in Matlab/Simulink, the group compared it to a series of other data -driven and statistical -based methods and has found the accuracy that 97.35%was reached.
A research team led by scientists from Georgia Southern University of the United States has developed a new deep -square framework for error diagnosis in PV -Omsmenters.
For this goal, the scientists used a Dual Graph Attention Network (Dualgat), which combines both spatial and temporary attention mechanisms, in particular the Diskgat and Tempat, respectively.
“Our work introduces a dual hole that combines both spatial and temporary attention mechanisms for the first time in the diagnosis of PV -Omverter,” said the corresponding author Jakir Hossen PV -Magazine. “This approach on two levels enables the model to record complex signal correlations and to evolve dynamics under different irradiation and temperature conditions, thereby guaranteeing greater robustness and interpretability compared to previous methods.”
To collect data on which the system must be trained, the group simulated a PV inverter system in Matlab/Simulink. It consisted of a PV source connected to a 3-phase inverter with 2 levels on the grid side, with a DC-DC-boost-Omgetter placed on the source side. The inverter consisted of six insulated gate bipolar transistor (IGBT) switches, each linked with an anti -parallel diode. Two types of open circuit errors were considered, namely some IgBT-open circuit errors and double IgBT-open-circuit errors.
“The three-phase current signals are measured when the irradiation and temperature of the PV series vary to evaluate the performance of the proposed method. The irradiation is adjusted in steps of 1 W/m2, ranging from 250 W/m2 to 750 W/m2, while the temperature is varied from 25 ° C in 1 C. “Each power signal contains 5,511 samples sets, with three power signals that correspond to the three phases for each fault type. In total, the dataset includes 121,242 samples over 22 classes, including the normal operating state.”
With the help of this data, the new framework constructed a graph of spatial fout relationships to illustrate how defective switches treat each other, together with a temporary graph to indicate how errors evolve in order. Then it combined both graphs to see how errors work together in space and time, and based on that it predicted which one of the 22 defective circumstances in which the inverter is located. From the simulated data, 80% was used for training and the rest was used for testing.
The system was tested against competitive error detection methods. They tested it with the help of data-driven approaches and on statistical-based methods, namely Ann, CNN, RNN, Gat, Gru + attention, TCN, Transformer, Resnet-1D, Inceptiontime, LightGBM + SVM, KNN, RF, DT and BC.
Under the neural network-based methods, the proposed Dualgat model turned out to reach the highest results in all statistics, with a test accuracy of 97.35%, the scientists said that it shows robust power to catch both spatial and temporary error patterns. “Other temporary models, such as Gat and RNN, also show strong performance, with accuracy of 95.18% and 94.12% respectively, which exceed traditional methods such as RF and SVM, which achieve accuracy of 87.11% and 85.37%,” they added.
In addition, the academics have carried out ablation studies of the method, where parts of the models are removed. The accuracy fell to 91.27%without Diskgat; It fell to 87.62%without Tempat; Without the regular it was 90.13%; And without the component of the Crossatment, this was 92.51%.
The new framework was presented in “Double graph attention network for robust error diagnosis in photovoltaic inverters“Published in Scientific reports. The team included researchers from Georgia Southern University of the United States, Cornell University, Bangladesh’s University of Rajshahi, Rajshahi University of Engineering and Technology, and the Multimedia University of Malaysia.
This content is protected by copyright and may not be reused. If you want to work with us and reuse part of our content, please contact: editors@pv-magazine.com.
