As the automation and intelligence levels of the power industry continue to improve, the importance of monitoring the operational status of power system equipment has become increasingly evident. As core equipment in power systems, the operational status of ultra-high voltage transformers directly impacts the safe and stable operation of the power grid. Therefore, accurate status prediction methods are needed to help power plants prepare in advance for sudden accidents. However, traditional transformer condition prediction methods face challenges such as high data missing rates, multi-source heterogeneous feature associations, and complex fault temporal evolution. To address these issues, this paper proposes a transformer condition temporal prediction framework that integrates physical mechanisms. First, a variational autoencoder (VAE) is used to robustly interpolate missing values in monitoring data. Second, a multi-scale spatio-temporal feature extraction module is designed by integrating Res2Net and Graph Convolutional Networks (GCNs), constructing a variable graph based on mutual information, and aggregating spatial features. Finally, a physically constrained Long Short-Term Memory Network (LSTM) prediction model is constructed, incorporating physical laws into the training process, significantly improving prediction accuracy. Specifically, the Mean Absolute Error (MAE) for leakage current prediction is reduced by 34.6% compared to traditional LSTM models, providing reliable support for intelligent power grid operations and maintenance.
Long et al. (Mon,) studied this question.