Self-Adjusting Domain Adversarial Transfer Learning Algorithm for Power Transformer Lifetime Prediction

Accurate prediction of the remaining useful lifetime of power transformers plays a crucial role in effective operation and health maintenance of the entire power grid. Machine-learning based data-driven solution is effective on mitigating the high financial and time costs induced by traditional prediction schemes, however, this solution is limited by data distribution, data volume, data quality, and variations of prediction scenarios, making it difficult to achieve good prediction performance. To address this concern, a self-adjusting domain adversarial transfer learning (SDATL) algorithm for power transformer lifetime prediction is first proposed in this article. It extracts features from source and target domains by stacked convolution sparse auto-encoder (SCSAE), measures distributions of source and target domains by the adversarial domain adaptive module. Furthermore, it implements advantageous transfer learning after determining the transfer object. Moreover, the parametric linear rectification unit (PReLU) adaptive adjustment module is designed and constructed to further improve the prediction accuracy by optimizing the network structure through network transfer and iteration. The proposed SDATL network is evaluated by using condition monitoring data from six dry-type transformers and compared with existing prevalence data-driven algorithms. The evaluation and comparison results validate higher prediction accuracy and better adaptability of the proposed scheme.