Investigation of Feature Selection Techniques for Improving Efficiency of Power Transformer Condition Assessment

Transformer oil tests have been conducted in utility companies as one of the major tools for evaluating the integrity of transformer insulation. However, the information obtained from different types of oil tests (the result of a particular type of oil test is termed as an oil characteristic in this paper) may have different significant degree in revealing the condition of a transformer's insulation system. This paper investigates feature selection techniques, which can identify a subset of the most informative oil characteristics amongst all oil characteristics for transformer condition assessment. This selected subset of oil characteristics can be subsequently fed into a support vector machine (SVM) algorithm for determining the health index level of the insulation systems of transformers. The major benefits of feature selection approach include (a) improving the efficiency of transformer condition assessment since only a subset of oil characteristics is used; and (b) assisting SVM algorithm to consistently attain satisfied accuracy since it can focus on the most relevant but non-redundant oil characteristics for transformer condition assessment. In the paper, two feature selection approaches namely correlation analysis based feature selection and minimum-redundancy-maximum-relevance (mRMR) based feature selection have been adopted. Case studies are provided to verify the applicability of feature selection approaches.