Transformer oil temperature is a critical parameter for assessing the operational status of power system equipment. However, existing predictive models often suffer from low accuracy and poor stability. In response to these challenges, this study introduces an enhanced LSTM-ARIMA model that utilizes the trend and cyclical patterns inherent in oil temperature fluctuations. The proposed approach employs Pearson correlation coefficient to eliminate redundant load features, applies the Fourier transform to decompose the original data into distinct frequency components. The LSTM model is employed to handle low-frequency components, whereas ARIMA is utilized to extract high-frequency components. The final prediction results are derived by integrating the outputs of both models through a linear combination. The proposed approach is validated using actual transformer oil temperature data collected from a specific province in China. Experimental results demonstrate that the model achieves a MAPE of 1.74% and a R2 above 0.98. Compared with traditional deep learning networks, this proposed method improves prediction accuracy and stability, providing certain valuable reference and method support for other time-series prediction problems in power system applications.
He et al. (Wed,) studied this question.