Smart Trafo: A Random Forest and LLM-Based Decision Support System for Power Transformer Fault Diagnosis via Dissolved Gas Analysis

Abstrak

Manual management of Dissolved Gas Analysis (DGA) data for power transformers at PT. PLN (Persero) UPT Manado has been identified as a critical operational bottleneck, with existing spreadsheet-based workflows susceptible to human error, poor historical traceability, and limited scalability. Prior studies on DGA-based transformer diagnosis have been predominantly confined to standalone classification models without integration into operational management systems, leaving a significant gap in practical field deployment. This research contributes a novel integrated Decision Support System named Smart Trafo, which is the first to combine a Random Forest classification model, Duval Pentagon visualization, historical trending analysis, and an LLM-based conversational assistant (Volty AI) within a unified full-stack web platform. The Random Forest model was trained on 375 DGA samples across six fault classes using five gas parameters conforming to IEEE C57.104, achieving an overall accuracy of 84% and a macro-average F1-score of 0.83. Feature importance analysis revealed Hydrogen (H₂) as the dominant diagnostic indicator at 26.2%. The system successfully automates DGA fault classification, eliminates manual calculation errors, and provides real-time technical recommendations, thereby enabling more efficient and data-driven preventive maintenance decisions at PT. PLN (Persero) UPT Manado.