Home > Published Issues > 2015 > Vol. 4, No. 4, October 2015 >

PV arc-fault feature extraction and detection based on bayesian support vector machines

Author(s): Yuan Gaoabc, Jianfei Dongabc, Yaojie Sund*, Yandan Lind, Rui Zhangd
a Changzhou Institute of Technology Research for Solid State Lighting, Changzhou, 213161, China
b State Key Laboratory of Solid State Lighting, Changzhou, 213161, China
c Beijing Research Center, Delft University of Technology, Beijing, 100083, China
d Department of Light Sources and Illuminating Engineering, Fudan University, Shanghai, 200433, China
International Journal of Smart Grid and Clean Energy, vol. 4, no. 4, October 2015: pp. 283-290
ISSN: 2315-4462 (Print)
ISSN: 2373-3594 (Online)
Digital Object Identifier: 10.12720/sgce.4.4.283-290

Abstract: In a PV system, DC arc is regarded as a serious fault, which might cause circuit damage and trigger fires. The arc fault, however, is hard to detect due to the special fields of photovoltaic systems: constant direct current without zero-crossing point, sophisticated components leading to noise interruption, and usually occupying large area. Therefore, detectable characteristics are of great importance to diagnosis and alarm of fault arcs in PV systems. In this paper, we presented a classification method of separating arcing and non-arcing in the feature space. First, data sets of current signal were sampled by designing field experiments with “pull apart” method for arc ignition. Then seven features in both time and frequency domains were defined and two of them in each domain were selected to train BSVM. In order to simplify the computation, the trained BSVM network was replaced by a separating line, which was proved to have a better performance of classification. Testing results showed that this method could diagnose fault arcs with high accuracy. But whether this method is suitable for other PV systems needs to be verified in further work.

Keywords: Photovoltaic systems, arc fault detection, bayesian support vector machines, PV testing

Full Paper.pdf