AI-Enabled Condition Monitoring for Predictive Maintenance and Early Anomaly Detection in Renewable-Rich Electric Distribution Systems: A simulation-based proof-of-concept study using the IEEE 33-bus feeder
TL;DR
The findings support the use of feeder physics, temporal feature engineering, and ensemble learning for condition monitoring, while also showing that deployment-oriented predictive maintenance will require field data, higher-fidelity failure models, and explicit adaptation to unseen DER operating regimes.
Abstract
High penetration of distributed energy resources (DERs) increases the operating variability of distribution systems and complicates the separation of normal renewable-driven changes from developing equipment and network abnormalities. This study presents a reproducible simulation-based condition-monitoring framework intended to support predictive-maintenance research in a renewable-rich radial distribution network. The IEEE 33-bus feeder is solved with a backward-forward sweep load-flow model under stochastic load variation and DER penetration. Four operating classes are generated: normal operation, progressive line-resistance degradation, localized overload, and source-voltage reduction. Each 20-step observation window contains 68 electrical features comprising bus-voltage magnitudes, branch-current magnitudes, net active/reactive power, and feeder losses. Temporal statistical descriptors are used with Random Forest, radial-basis-function support vector machine (RBF-SVM), and HistGradientBoosting classifiers. On an in-distribution (IID) test set, HistGradientBoosting achieved 90.75% accuracy and 90.75% macro-F1, outperforming Random Forest and RBF-SVM. Under a high-DER stress test with 50–70% DER penetration, performance decreased to 83.50% accuracy and 83.58% macro-F1, indicating useful but imperfect generalization. Progressive line degradation was the most difficult condition to distinguish from normal operation, whereas all 100 source-voltage-reduction samples were correctly classified in the simulated IID test set. The findings support the use of feeder physics, temporal feature engineering, and ensemble learning for condition monitoring, while also showing that deployment-oriented predictive maintenance will require field data, higher-fidelity failure models, and explicit adaptation to unseen DER operating regimes.