Transformer fault types and severity class prediction based on neural pattern-recognition techniques

Abstract

Dissolved gas analysis (DGA) is used to diagnose power transformer fault based on the concentration of dissolved gases and the ratios between them. These gases are generated in oils as a result of electrical and thermal stresses, but these DGA techniques cannot identify the severity of the fault types. In IEEE Standard C57.104, the maintenance action is taken based on the total dissolved combustible gases, which is not sufficient because it ignores the importance of the gas type and its change rate. Thermodynamic theory using different starting decomposing materials, namely, n-octane (C8H18) and eicosane (C20H42), is used to estimate the severity of transformer fault types. Two scenarios are suggested with different data transformation techniques to enhance neural pattern-recognition (NPR) method accuracy for predicting transformer fault types and their severity classes. The proposed scenarios are built based on 446 samples collected from the laboratory and literature. Results refer to the role of the starting decomposing material on the severity of the transformer fault and illustrate that the proposed model has a higher accuracy (92.8%) compared with other DGA methods for diagnosing transformer fault types and superior accuracy (99.1%) to predict fault severity class.