A Supervised Learning Approach to Outage Prediction Using Large Outage Management System Data
Proceedings Paper
Shees, A, Amaan, A, Mohammed, J et al. (2026). A Supervised Learning Approach to Outage Prediction Using Large Outage Management System Data
. ELEVENTH INTERNATIONAL CONFERENCE ON COMMUNICATION NETWORKS, ICCN 2015/INDIA ELEVENTH INTERNATIONAL CONFERENCE ON DATA MINING AND WAREHOUSING, ICDMW 2015/NDIA ELEVENTH INTERNATIONAL CONFERENCE ON IMAGE AND SIGNAL PROCESSING, ICISP 2015, 283 1521-1531. 10.1016/j.procs.2026.06.229
Shees, A, Amaan, A, Mohammed, J et al. (2026). A Supervised Learning Approach to Outage Prediction Using Large Outage Management System Data
. ELEVENTH INTERNATIONAL CONFERENCE ON COMMUNICATION NETWORKS, ICCN 2015/INDIA ELEVENTH INTERNATIONAL CONFERENCE ON DATA MINING AND WAREHOUSING, ICDMW 2015/NDIA ELEVENTH INTERNATIONAL CONFERENCE ON IMAGE AND SIGNAL PROCESSING, ICISP 2015, 283 1521-1531. 10.1016/j.procs.2026.06.229
Modern power grids are becoming more complex, and our reliance on digital systems keeps growing. This makes it essential to develop smarter and more resilient ways to manage outages. In this study, we use supervised machine learning to predict when outages are likely to occur, based on real-world data from an Outage Management System. By creating useful features like substation IDs, device types, customer counts, and hazardous load indicators the models can identify patterns that signal potential outage events. The dataset, drawn from over 600 substations at Florida International University's ADMS testbed, was divided into training (80%) and testing (20%) sets, and the three classifiers like Logistic Regression, Random Forest, and XGBoost were evaluated using accuracy, precision, recall, F1-score, and AUC. To enhance trust and interpretability, SHAP (SHapley Additive exPlanations) is employed to explain model decisions and identify the most influential outage predictors. Results show that XGBoost achieved perfect performance (Acc = 1.00, PR = 1.00, R = 1.00, F1 = 1.00, AUC = 1.00), outperforming Random Forest (Acc = 0.99, F1 = 0.97) and Logistic Regression (Acc = 0.81, F1 = 0.20). SHAP analysis revealed that the number of customers affected, and outage device type were the most influential features.
