Neural Network Approach to Dissolved Gas Analysis for Fault Analysis in Power Transformers

In this paper, an application on artificial neural network (ANN) named Levenberg Marquardt backpropagation training algorithm as multi-layer reverse diffusion ANN approaching to Dissolved Gas Analysis (DGA) was implemented to utilize a computational and automatic internal fault prediction of power transformer oil insulation. The DGA method applying Key Gas method and Duval Triangle method is used to identify such type of faults. The RadBas, Tan-Sigmoid, Log-sigmoid, Purelin linear transfer functions were chosen to find mean absolute error and least learning time during the analysis. The DGA results of 244 power transformers in distribution system were used as training data while additional DGA results of 100 power transformers were tested and compared to manual analysis. The results show that there is 6.4% absolute error from manual fault prediction. The method and tool are appropriate and less time consuming for condition-based maintenance of power transformer.