Automated Transformer Health Prediction: Evaluation of Complexity and Linearity of Models for Prediction

Transformer failure is a significant concern in electrical power systems as it can result in costly damages and endanger human lives. Early detection of defects can minimize damage before it becomes dangerous to work with. Predicting transformer health is essential for ensuring continuous quality of service to consumers through a predictive maintenance approach. Maintenance records of transformers often include temperature, oil quality, and vibration. The challenge lies in developing models that enable the prediction of the transformer health index (HI) from these maintenance records. Several research studies have formally reported various implementations of machine learning algorithms to predict transformer health index (HI). In this research report, authors introduce a machine learning algorithm and classification on transformer health detection approach using a support vector machine (SVM) and an artificial neural network (ANN). The direction of this study is to evaluate the complexity of this prediction domain using both machine learning models. Both SVM/ANN are commonly deployed machine learning models in most application domains. The authors investigate this problem from both regression and classification perspectives, implementing various kernel functions associated with the SVM such as radial basis function (RBF) and artificial neural network (ANN). Despite the good separation between classes and good regression on the training set using more nonlinear models, it is observed that overfitting occurs when evaluating on independent test sets for both regression and classification (especially in cases involving SVM). The authors found that more nonlinear kernels (SVM) yielded better performance thereby indicating that future research may benefit from more linear models. The study postulates that machine learning models should be chosen based on the general suitability, data linearity, and complexity to achieve accurate predictions of transformer health index (HI). The results of this study may provide insights for future research in developing models that can accurately predict transformer health.