Investigating performance of transformer health index in machine learning application using dominant features

Abstract

Transformer Health Index (HI) has become a standard tool for performing transformer health evaluations. Due to economic constraints, the recently published paper focuses on developing various techniques to identify the most dominant features for transformer HI prediction. However, the fundamental problems concerning their input features remain unresolved since most suggested features contradict industry practice. In this paper, the primary objective is to investigate the performance of the transformer HI by developing and utilizing only dominant features following the industry recommendation. The investigated dominant features in this paper using 1) CO2/CO ratio and 2) the Incipient fault for detecting temperature abnormalities, and 3) the Dissipation Factor (DF) for detecting oil contamination. The performance validation is carried out using various machine learning (ML) classifiers. Also, the performance of the ML model is validated based on 10-fold type cross-validation to avoid biases in the experiment. As a result, the proposed Artificial Neural Network (ANN) network utilizing the investigated dominant features following the industry practice has produced the highest average accuracy of 80.09% than others ML techniques as a classifier. Hence, additional studies to complement the investigated dominant features may be considered for the subsequent investigation.