Neural network modeling of distribution transformers with internal short circuit winding faults

Abstract

To detect and diagnose a transformer internal fault an efficient transformer model is required to characterize the faults for further research. This paper discusses the application of neural network (NN) techniques in the modeling of a distribution transformer with internal short-circuit winding faults. A transformer model can be viewed as a functional approximator constructing an input-output mapping between some specific variables and the terminal behaviors of the transformer. The complex approximating task was implemented using six small simple neural networks. Each small neural network model takes fault specification and energized voltage as the inputs and the output voltage or terminal currents as the outputs. Two kinds of neural networks, back-propagation feedforward network (BPFN) and radial basis function network (RBFN), were investigated to model the faults in distribution transformers. The NN models were trained offline using training sets generated by finite element analysis (FEA) models and field experiments. The FEA models were implemented using a commercial finite element analysis software package. The comparison between some simulation cases and corresponding experimental results shows that the well-trained, neural networks can accurately simulate the terminal behaviors of distribution transformers with internal short circuit faults.