Prediction Method for Transformer State Based on GRU Network

Dissolved gases content in power transformer oil can provide important information for transformer condition assessment. The recurrent neural network (RNN)has always performed well in the data labeled in the form of sequence but it has long term dependency and vanishing gradient problem. Gated Recurrent Unit (GRU) is an improved version of RNN which deal with problems in RNN. This paper proposed a new forecasting model based on GRU network. The studies show that the model can effectively predict the change state of oil gas in power transformer. Compared with the Long Short Term Memory (LSTM) networks which is also improved recurrent neural network, the accuracy rate is comparable, and the training time is shorter and the universality is better. When the interval and fluctuation of data are greater, the prediction of the GRU network is also more stable.