Predictive Condition Assessment of Distribution Transformers: Using Load, Thermal, Power-Quality, and Maintenance Indicators

Abstract

- Distribution transformers occupy a critical position between bulk-power infrastructure and end users, yet many units are still managed through periodic inspection, fixed alarms, and reactive maintenance. This study develops an interpretable predictive condition-assessment framework that integrates loading, thermal behavior, power quality, and maintenance history into a single engineer-facing reliability health score while retaining statistical and machine-learning challengers. The public-data foundation is the Kaggle Distributed Transformer Monitoring dataset, which contains five IoT data files and 54 variables collected at approximately 15-minute intervals between 25 June 2019 and 14 April 2020, including three-phase current and voltage, oil and winding temperatures, oil level, power, power factor, frequency, and harmonic-distortion measures (Putchala, 2020; Ramesh et al., 2022). Because that public release does not contain adjudicated maintenance histories or a complete failure-outcome record suitable for causal maintenance inference, the empirical evaluation deliberately separates data provenance from scenario validation. A 19,200-observation benchmark-calibrated engineering panel reproduces the source signal families and adds explicitly simulated maintenance-age, defect-backlog, and prior-alarm variables. A severe multi-indicator condition onset within the subsequent six hours is used as a controlled prospective target; it is a method-validation event, not a claim of actual transformer failure prevalence. The study compares a fixed threshold rule, an interpretable 0-100 condition-stress score, robust multivariate distance, Isolation Forest, class-weighted logistic regression, random forest, and gradient boosting under a chronological 60/20/20 development-validation-test design. On the locked temporal test, logistic regression achieved the strongest discrimination (AUROC 0.763; 95% bootstrap interval 0.744-0.780; AUPRC 0.405), while the interpretable condition score achieved AUROC 0.647 and AUPRC 0.315 and produced a monotonic risk gradient: the condition-event rate rose from 7.8% in the very-healthy band to 70.0% in the critical band. Feature-layer ablation showed that adding thermal variables to load indicators increased AUROC from 0.716 to 0.742, while adding maintenance evidence increased AUPRC from 0.366 to 0.405. Static thresholds were precise for some severe states but captured fewer emerging events, whereas anomaly methods were sensitive to temporal threshold transport. The results support a hybrid maintenance architecture in which interpretable engineering scores and hard limits provide transparent triage, statistical models provide an independent challenger, and engineers retain authority over inspection, testing, derating, repair, and replacement decisions. The contribution is therefore not a claim that machine learning replaces transformer engineering; it is a reproducible decision framework for converting heterogeneous condition evidence into prioritized, auditable maintenance action