Simultaneous Multispot Temperature Prediction of Traction Transformer in Urban Rail Transit Using Long Short-Term Memory Networks

Abstract

Hot-spot temperature (HST) is a typical indicator of transformer’s health status, accurate prediction of HST is critical for prognosis and health management (PHM) of traction transformer in urban rail transit (URT), where dramatically fluctuating loads and complex climates pose serious challenges. This paper proposed a novel deep-learning-enabled method to predict multi-spot temperatures of transformer simultaneous using long short-term memory (LSTM) neural networks. Real-world operation data collected from Qingdao Metro over a year-long period were used to train the prediction model and verify its validity. The most appropriate transformer-related parameters were selected as the inputs of model by Pearson correlation coefficients (PCC) to improve the accuracy and efficiency of the model. Besides, the dropout and early stopping techniques are used to prevent model from overfitting. Furthermore, the robustness and versatility of proposed method were verified by testing the data on different seasons of multiple transformers, and the superiority of the method was also proved by comparing with other methods. The results show that the proposed model can achieve accurate on-line temperature prediction of transformer, and the mean-relative-error of sixteen minutes ahead HST prediction is less than 0.26%. The proposed method can provide a reference for PHM of transformers.