Prediction for Dissolved Gas in Power Transformer Oil Based on Temporal Convolutional and Graph Convolutional Network

Abstract

Dissolved Gas Analysis (DGA) is an important method to identify internal faults of transformers. A novel predictive method for dissolved gas content in transformer oil based on temporal convolutional network (TCN) and graph convolutional network (GCN) is proposed in this paper. First, a Temporal Convolutional Network based on dilated causal convolution algorithm is designed to extract feature information from both current data and historical data of each gas content. Since the correlation coefficients among different gases may affect the accuracy of the prediction, a topological structure diagram is constructed to describe the relationship among different gases. Then, a GCN is designed to predict the gas content in the desired time horizon by using the information obtained from the topological structure diagram. The elements in the adjacency matrix of GCN are replaced by Pearson correlation coefficients to improve the accuracy of the prediction. The test results show that the proposed method in this paper achieves high accuracy for the prediction of dissolved gas content in transformer oil. The Mean Absolute Error (MAE) of the predicted gas content by the proposed method is reduced by 11.18% compared with the Long Short-Term Memory (LSTM) network and 12.23% compared with the traditional Back Propagation Neural Network (BPNN).