Forecasting Feeder-Level Outages with Hybrid Time-Series/ML Models: Accuracy, Explainability, and Maintenance Prioritization

Feeder-level outages are a real headache for utilities nowadays, as climate variability, fast vegetation growth, and aging infrastructure are making the network more vulnerable. The current article gives a discussion on popular statistical time-series models like ARIMA, ETS, and Prophet and compares them to tree-based ensembles and LSTM hybrids in predicting outages in the short term. It is a combination of historical ticket, weather radar, vegetation indices, and asset data. The condition evaluation methodology is a rolling-origin condition evaluation methodology whereby the performance is measured in terms of accuracy, calibration, error statistics, and reliability plots. Findings indicate the existence of the hybrid models by performing better than the classical models, particularly where the nonlinear interaction between vegetation and the weather is considered. The outcomes of the calibration test also show that the hybrids are more reliable when in a hybrid format. To bridge the gap between the model performance and application, the study comprises the combination of SHAP-based attribution to graphically highlight the risk drivers in a clear manner, such as vegetation density and adverse weather. Such interpretability gives more straightforward priorities in terms of maintenance and regulatory compliance. The results indicate that hybrid plus explainable pipelines have the potential to make outage predictions turn into a viable utility-level decision-support system.