DGA Warning Values for PD Defects Obtained from Normal Data Analysis on a Transformer Experimental Platform

This paper presents an approach to determine the DGA warning values for a transformer experimental platform by using the normal data from on-line DGA. In order to collect normal data and PD-defect data, an experimental platform of a 10 kV transformer with a capacity of 100 kVA is established for long-term operation. Five sets of DGA data are obtained, 1 set from the normal data, 2 sets from the PD-defect in developing-discharge stage (D1-1) and 2 sets from PD-defect in severe-discharge stage (D1-2). All sets of data are analyzed and found better suitable for the lognormal distribution, in which there are 2 parameters, µ and σ. It is found that the position parameters µ for 5 sets of data are nearly linear with the PD values and the scale parameters σ for 5 sets of data are nearly equal for certain discharges. Therefore, PD-defect data distribution can be approximately deduced from the normal data. In general, there is a common domain between the distributions of either PD-defect data and normal data or D1-1 and D1-2. A ROC curve, widely used in medical diagnosis, is employed to determine the warning values in common domain data. The optimal cut-point in common domain between the normal data and the deductive data of D1-1 is chosen as the Typical Value, and the cut-point between the deductive data of D1-1 and that of D1-2 is chosen as the Alarm Value.