Missing data imputation using an iterative denoising autoencoder (IDAE) for dissolved gas analysis

• A data imputation method is proposed to address transformers with multiple missing DGA data. • An iterative imputation method is demonstrated, based on a denoising autoencoder. • Features are extracted from both gas trend and composition ratios for better imputation. • The proposed method is validated using a vast amount of field data. With the expansion of the energy market, safe and stable operation of the electrical power system has become an important issue. In an effort to achieve this goal, much research has been conducted on diagnosis approaches suitable for core components of the electrical power system. Transformers are one such core component. Most of the research on transformers has focused on developing a diagnosis model; less effort has been devoted to the data, in spite of the fact that such models require data of sufficient quantity and quality, which is not usually readily available for transformers. Thus, in this paper, we propose a way to fully exploit the valuable transformer data, using a data imputation approach called the iterative denoising autoencoder (IDAE) method. The proposed method imputes missing values of dissolved gas analysis (DGA) data, which is frequently lost, for various reasons. IDAE can help diagnose the health state of transformers accurately by estimating the missing values of DGA data. The proposed method is verified in this research through three comparative studies that examine field data provided by an electric power corporation. Specific studies provide: (1) a comparison with conventional methods on imputation performance for a single gas, (2) examination of imputation performance between multiple missing values, and (3) documentation of diagnosis accuracy before and after imputation. The results of the case studies show that the proposed method is effective for imputation of the missing DGA data.