An Approach Based on Transfer Learning to Lifetime Degradation Rate Prediction of the Dry-Type Transformer

Abstract

Lifetime prediction of the power transformer plays an important role in maintaining the stable operation of power equipment. However, due to the complexity of insulation structure degenerative process, the existing methods featuring high cost and low precision are not effective enough in transformer life time prediction. Meanwhile, how to effectively and promptly respond to a new prediction scenario of insufficient and limited data is a common challenge for all the data-driven prediction methods. To address these concerns, a prediction approach of a back adoptive adjustment transfer learning scheme (BAATL) is proposed for lifetime degradation prediction of the dry-type transformer. The power transformer condition monitoring data of Supervisory Control and Data Acquisition system is conducted as the data driven. A deep neural network, a transfer learning module and a back adjustment module are constructed to realize feature extraction, domain adaptation and prediction network optimization. The proposed scheme is able to improve prediction accuracy and resolves the problems and drawbacks of traditional prediction methods, and presents its superior portability and application potential in the case of data shortage and scenario change. With authentic datasets, simulation tests performed on the condition monitoring data of dry-type transformers prove the effectiveness of the proposed scheme.