Short-term fault prediction method for a transformer based on a CNN-GRU combined neural network

To explore the relationship between transformer state parameters and quantify the impact of the external environment on a transformer, this paper proposes a short-term fault prediction method for a transformer based on a convolution neural network (CNN) and gating cycle unit combined neural network (GRU). First, mining the correlation between transformer state parameters through association rules, and incorporating variable weight method to evaluate the status of the transformer, an exponential function is introduced to establish the fault rate model representing the operational state of the transformer. This is used as the prediction state parameter. Secondly, considering the influence of the external environment on the operational status of the transformer, a fault prediction feature set is constructed based on the date, meteorological and production process factors. Then, the convolution neural network extracts the feature vectors between the feature set and the fault rate in the high-dimensional space, and inputs the result into the gating cycle unit for optimization training, so as to predict the development trend of the transformer fault rate. Finally, the feasibility and effectiveness of the proposed method are verified by the fault prediction trend analysis of a transformer on an offshore platform. Compared with the long short-term memory (LSTM), GRU, CNN-LSTM and support vector machine models, the proposed method has higher prediction accuracy and higher prediction efficiency. © 2022 Power System Protection and Control Press.