INTEGRATION OF SCORING-WEIGHTING METHODS AND MACHINE LEARNING ALGORITHM IN MODELING HEALTH INDEX TRANSFORMERS AT VARIOUS TYPES OF OPERATING VOLTAGE

• Main Author: Akmal Afibuddin Putra, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87636
• Institution: Institut Teknologi Bandung

Power transformers are vital and costly components of the electrical power system. Ensuring the reliability and safety of these systems necessitates consistent monitoring and maintenance of transformers. The Health Index (HI) is a widely used method for assessing transformer conditions. This study models a method for calculating the transformer Health Index using both conventional (scoring-weighting) and non-conventional (machine learning) approaches. The developed method considers transformer operating voltages and integrates dissolved gas analysis interpretations using multiple methods (Multi-methods). The findings indicate that higher operating voltages correlate with faster rates of transformer condition degradation. The non-conventional method with machine learning demonstrates higher prediction accuracy and reduces calculation subjectivity without requiring expert involvement. The Random Forest algorithm outperforms others in failure prediction based on DGA results, while Neural Networks excel in predicting HI categories with numerous input parameters. The application of the Synthetic Minority Oversampling Technique (SMOTE) improves model performance by balancing dataset classes, achieving a prediction accuracy of 99% for DGA and 97% for the Health Index. The results of the method development were then integrated into the development of transformer health index prediction software. This research provides accurate insights into transformer conditions across various operating voltages, supports asset managers in designing effective maintenance strategies, and helps prevent sudden transformer failures.