This paper presents a comparative study of neural network (NN) efficiency for the detection of incipient faults in power transformers. The NN was trained according to five diagnosis criteria commonly used for dissolved gas analysis (DGA) in transformer insulating oil. These criteria are Doernenburg, modified Rogers, Rogers, IEC and CSUS. Once trained, the neural network was tested by using a new set of DGA results. Finally, NN diagnosis results were compared with those obtained by inspection and an analysis. The study shows that NN rate of successful diagnosis is dependant on the criterion under consideration, with values in the range of 87-100%.

A computer model that describes the dynamic arc behavior in the high- and low-current regions before current zero is proposed. The model divides the current and voltage waveform in two regions. A differential equation for both regions which unifies current and voltage time derivatives is obtained by means of a generalized function method. The computer waveforms reproduced with the model show good agreement with measured results published in in the low and high current regions, but further comparison with other test measurements are required to know if the model has any feature of predictability.