Teng Mingxing^*, Wang Yanyan

Hunan Mechanical & Electrical Polytechnic

Abstract：

In the substation, the station power supply system provides uninterrupted power to critical secondary equipment. Battery reliability significantly impacts the safe and stable operation of substations. Traditional prediction methods lack accuracy when processing long-sequence complex degradation features. To address this issue, a remaining useful life (RUL) prediction model is proposed by fusing Long Short-Term Memory (LSTM) and Transformer architectures. The proposed model utilizes LSTM to extract local time-series features. The self-attention mechanism of the Transformer is introduced to capture global degradation dependencies. The prediction model is constructed and tested based on the measured charge and discharge data of substation batteries. Experimental results show that the proposed LSTM-Transformer model achieves better performance in root mean square error (RMSE) and mean absolute error (MAE). This ensures the reliable operation of  batteries and  power supply system in  the substation.

Key Words：

remaining useful life; substation battery; long short-term memory; transformer; deep learning