Abstract
The presence of metallic particle is the most adversely influencing defect in gas-insulated substation (GIS). These particles may move freely in the GIS enclosure or adhere to the solid spacer surface due to electrostatic forces. The later can initiate partial discharges (PDs) in the weakest area, i.e., the triple junction. Therefore, the investigation of PD characteristics and particle size and position on the spacer surface is the significant step toward the reliability improvement of the GIS equipments. This paper presents the use of the back-propagation artificial neural network (BP-ANN) technique supplemented with principal component analysis (PCA) as the PD pattern recognition tools for the estimation of the particle size (length) and position on the spacer surface in a simulated GIS. PD-featured acquisition is performed by collecting their fingerprints from the measurements carried out using IEC 60270 method. The role of PCA is to reduce the dimension of the collected PD fingerprint data. The obtained results show that PCA can significantly improve the BP-ANN performance in terms of execution time. Without PCA, 88 and 92 % accuracies can be achieved when BP-ANN is implemented with 1 and 2 hidden layers, respectively. With the integration of PCA, execution times are greatly reduced while retaining fairly high accuracy of 88 % in both cases. Thus, the proposed method is a contribution to development of the tool for improving the reliability of GIS.