J4.3 Improved Power Outage Restoration Predictions using Neural Networks

Monday, 29 January 2024: 5:00 PM
347/348 (The Baltimore Convention Center)
Brennan Joseph Stutsrim, Univ. at Albany, Albany, NY; and N. P. Bassill and K. J. Sulia

Estimated time of restoration (ETR) for electric outages is a key driver of utility customer satisfaction. Creating and improving the tools to calculate ETR can provide a clear benefit to utility customers, regulators, and the electric utility industry. Accurate, timely ETR values help utility decision makers to properly gauge the need for outside mutual assistance resources, which represent a significant financial commitment during large-scale outages. ETR accuracy is also a benchmark for gauging restoration performance, as defined by NYS regulators. During "blue sky" days, Central Hudson Gas and Electric's current ETR system defaults a set prediction for all outages before relying on restoration crews in the field to update the prediction. During “storm” mode, ETR’s are established based on feedback from field crews, and expert analysis from the storm incident command team. In an effort to improve on Central Hudson’s current system, neural networks (NN) were trained to predict the ETR of ongoing outages in real time. NN’s were trained on features built from Central Hudson’s existing outage management system, including information on the location, type of equipment, context within the ongoing outages, and status in the restoration process. Shapely values were used to analyze feature importance and study what factors produce extended power outages. Results show that using NN’s can produce improvements over Central Hudson’s current ETR system on the scale of hours, with long duration outages showing the largest improvement.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner