Artificial neural network based method for temperature correction in FDS measurement of transformer insulation

Abstract

Frequency domain spectroscopy (FDS) measurement has become an important method for the assessment of the condition of the insulation of oil transformers. In recent years, numerous researchers have found that temperature variation affect FDS results. The master curve technique is commonly used to correct the effect of temperature on FDS results. In this paper, an FDS experiment is carried out on a sample transformer. Then, for this transformer, insulation model parameters are determined by using a genetic algorithm based on the FDS results. Then, by using the insulation model parameters, tanδ curves are simulated and compared to real results. Finally, an FDS experiment is conducted on two other transformers at 22 °C, 30 °C, 40 °C, 50 °C, 60 °C, and 70 °C (in order to give sufficient information for a training neural network) and insulation model parameters are calculated via the genetic algorithm. In one of the transformers, the effect of temperature on the FDS curves is corrected by using the master curve technique and the FDS curves are transmitted over the reference curve. It is also shown that the transformer insulation is an Arrhenius-type dielectric. The error of this method is calculated by using a mean square error technique. In the two other transformers, the insulation model parameters related to 22 °C are considered as the target parameters. The insulation model parameters of the other temperatures are fed into an artificial neural network as input, to train it to transfer insulation model parameters related to other temperatures to the reference parameters. Finally, the errors of both methods are compared, and it is shown that this latter method for correcting the temperature effects in the FDS method is the more effective.