Researchers in China have made a data set of different PV errors and normalized these to meet different array sizes and typologies. After testing the new approach in combination with the 1D-CNN Machine Learning technology under different circumstances, they found an accuracy of more than 99%.
Researchers from Chinese Shandong University have developed a new method for error diagnosis in PV arroys, using function engineering and one-dimensional conventional neural networks (1D-CNN).
The novelty of the technology lies in the step of the Engineering function, which normalizes raw inputs and makes it independent of the size or layout of the system. The 1D-CNN is a kind of neural network architecture that is specifically designed to process one-dimensional sequential data. It expands the traditional CNN concept, usually used for image recognition, to process successive data.
“The proposed method is intended to improve the generalization of the model in clean or noisy data, the adaptability of the method to diagnose the errors in PV arrays and topologies on different scale, and the transferability of the pre-trained model to diagnose the errors in different PV arrays scales,” said the research. “The current voltage (IV) Curve characteristics are normalized on the basis of the PV system scale, which creates a standard standardization approach for various scales of PV arrays and the generalization capacity of the proposed method improves.”
The first step in the study was to make a dataset. For this purpose, the scientists have built a PV array in Matlab/Simulink, consisting of 20 modules organized in four parallel strings, each containing five consecutive modules. They were simulated in six scenarios, one of the regular effect and five defective, namely line-to-line error, arch error, partial shadow condition, open circuit error and demolition error. With varying irradiation, temperature and resistance in each of those scenarios, 16,200 IV curve monsters were made.
In the next step, the team used their new function standardization technology to develop the power on maximum power (Imp), the voltage on maximum power (VMP), the short-circuit current (ISC) and the open circuit voltage (VOC). They were usually normalized by coordinating the raw electrical data in a percentage, which represents how much performance has been lost. After this, the team carried all processed data in 1D-CNN, also with the help of the ADAM optimization algorithm. The database was trained at 80% of the data and then tested for the remaining 20%.
The data was tested both clean and noise, where Gaussian noise was added. In order to investigate the adaptability of the method, various array sizes were tested, including 10 × 3, 10 × 4 and 15 × 4, together with the front topologies, namely Series-Parallel (SP), Total Cross-Tedied (TCT), Honingraat (HC) and Bridges (BL). Finally, the new model was also compared to four other models for machine learning and artificial intelligence, including MLP, KNN, Ann and 1D-CNN.
“The diagnostic model offers a higher accuracy in clean and noisy environments in the SP-PVA data set, with an accuracy of 99.90% and more than 99% for clean and noisy data respectively. These results emphasize the meaning of the formalized functions in optimizing model performance and improving the robustness. “The proposed method offers a considerable advantage by demonstrating exceptional adaptability and achieving more than 99.63% accuracy for TCT and BL topologies and 99.90% for HC topologies.”
The new approach was described in “Error diagnosis in photovoltaic Arrays: a robust and efficient approach using function engineering and 1D-CNNPublished in the International Journal of Electrical Power & Energy Systems.
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.
