Identification of Partial Discharges at DC Voltage using Machine Learning Methods

In order to guarantee the operational reliability of high voltage direct current
equipment, the partial discharge measurement plays an important role. For AC systems,
numerous works on partial discharge interpretation have been carried out for several
decades. However, these methods for AC systems cannot be applied directly to DC
systems. In this work, partial discharge data for typical defects in gas-insulated systems
were accumulated and several identification methods on partial discharges at DC voltage
were compared. As for the experiment, specially designed test cells were used, including
three kinds of typical defects: floating electrode, protrusion on a high voltage electrode and
free metallic particle. The experiments were performed in SF6 gas at DC voltages with both
positive and negative polarity. The experimental results showed that even in the same
category of the defect, such as the floating electrode, the partial discharge patterns greatly
varied depending on test conditions or the shape of the defect. As for the partial discharge
identification, three kinds of input data (statistical features, raw partial discharge data and
the pixel data of the NoDi* Pattern mappings) and two kinds of identification algorithms (the
artificial neural network and the decision tree) were combined. The performances of all
these methods were compared. Using the statistical features or the pixel data of NoDi*
Pattern mappings as the input data showed good performances and were able to correctly
identify PD defects with more than 95 %.