Application of Machine Learning Algorithms for the Predictive Maintenance of Power Transformers in Electrical Transmission Networks

The increasing need for aged power transformers raises the probability of unintended transmission network failures. This research proposes a Reliability, Availability, Maintainability, and Safety (RAMS) guided machine learning paradigm that uses physically grounded, synthetically generated data based on realistic degradation processes. Five predictors, namely, thermal transients, cellulose insulation degradation, humidity variation, operating aging, and demand variation, were simulated to emulate infrequent fault behaviors. As the failure occurrence is severe (less than 1%), a continuous variable-optimized Gaussian-augmented SMOTE method was used to counter the problem. Random Forest reached an accuracy of 85%, F1-score got 0.83, and AUC got 0.73, across an ensemble of 1, 000 Monte Carlo stratified tests, significantly surpassing a Multi-Layer Perceptron, notably on the recall (84% vs. 75%) metric. Feature importance identification pointed to insulation degradation and service aging indicators. As opposed to previous research on transformer faults, the originality in this research comes about through the synthesis of RAMS-bound synthetic data used alongside interpretable ensemble predictors, allowing reproducibility under rare-event situations. The research recommends cost-effective preventive measures, which can minimize the cost related to outlier issues by up to 18-22% while improving the grid’s resilience.