A Prediction Model for Time Series of Dissolved Gas Content in Transformer Oil Based on LSTM

Abstract

By combining with dissolved gas analysis, time series prediction of dissolved gas content in oil provides a basis for transformer fault diagnose and early warning. In the view of that, a prediction model based on long short time memory (LSTM) network for time series of dissolved gas content in oil is proposed, which takes advantage of LSTM network’s ability to deal with long-sequence prediction problems. Five characteristic gas concentrations are used as input to the model, and the hyper parameters of the model is optimized by Bayesian optimization algorithm to further improve prediction accuracy, then a LSTM prediction model is constructed. By case study, it is verified that the proposed model can precisely predict time series of dissolved gas content. Compared with gray model, BP neural network and support vector machine, the proposed model has higher prediction accuracy and can better track the trend of time series of dissolved gas content in oil.