A prediction method for power transformer state parameters based on feature attention mechanism

Precise predictions on state parameters of power transformers could provide a comprehensive assessment of current operation status and potential risks on equipment. In the proposed power transformer prediction model, a feature attention mechanism is introduced to basic long short-term memory network (FA-LSTM) to mining the association rules between the target attribute and its relevant features like grid operation and meteorological data. Compared to previous researches treating the attributes as independent ones, the relationship extracted through time iterations could correct prediction results by offering auxiliary information for target series. Additionally, the feature attention is not restricted by the preset threshold value in traditional association rule mining algorithms. In the case study of predicting the concentration of dissolved hydrogen in insulating oil, experiments show that our proposed FA-LSTM outperforms traditional methods with lower absolute percentage errors, giving a more accurate result for time series prediction of power transformer state parameters.