Automated Transformer Health Classification: Feature weight Optimisation using Differential Evolution

The reliability of electrical power systems heavily relies on the proper functioning of transformers, which are crucial components that step up or down the voltage level of electricity in the transmission and distribution networks. Transformer failure can lead to widespread power outages, financial losses, and potential harm to human life. It is, therefore, imperative to detect any defects in transformers as early as possible to minimize the damage and prevent dangerous situations from occurring. To achieve this, predictive maintenance has become an increasingly popular approach in the power industry. Predictive maintenance involves monitoring the performance of transformers using various techniques, including temperature, oil quality, and vibration analysis, to predict the health of the equipment and schedule maintenance activities accordingly. Predictive maintenance can be enhanced by developing models that can predict the transformer Health Index (HI) accurately, which can help identify potential issues before they become critical. However, developing accurate models that can predict transformer health is a challenging task. The data collected from transformer maintenance records is often complex and multidimensional, and traditional methods may not be sufficient to extract meaningful insights. In this research, the authors focus on using Error-Correcting Output Coding- Support Vector Machine (ECOC-SVM) models to predict transformer health. Adaptive Differential Evolution (ADE) was deployed to optimized the weights for each feature. Weighting the features is a useful approach in optimizing any type of machine learning algorithms. ADE is selected due to its ability to find good solutions without prior parameter tuning and able to work in highly multi-model problems (in this case, selecting weights). Result shows that there is substantial improvement by optimizing the weight using ADE.