Metode SMOTE ENN untuk Imbalanced Data dalam Klasifikasi Health Index Transformator Distribusi

Abstract

In the power distribution system, transformers play a crucial role in delivering electricity reliably to end users. To ensure that these assets continue to operate effectively, regular condition monitoring is necessary. One widely used approach involves calculating the Health Index (HI), which is derived from various inspection parameters and helps categorize the condition of transformers into good, fair, poor, or bad. However, a common issue found in such datasets is class imbalance, where most records fall under the “good” category. This imbalance tends to skew the performance of classification models, making them less accurate for minority classes.In this study, I explored two resampling techniques—SMOTE and SMOTE ENN—to address the imbalance problem. These were combined with Random Forest and Logistic Regression to classify the health index. Based on the experiments, SMOTE with Random Forest yielded the highest accuracy at 99%, slightly outperforming SMOTE ENN which reached 98% with the same model. One possible reason is that SMOTE increases minority samples without removing any data, keeping the dataset rich in variation. In contrast, SMOTE ENN removes potential noise, but this can also reduce diversity. Overall, the results show that both techniques are valuable, and the insights can support the implementation of condition-based maintenance strategies for distribution transformers