Community Outage Prediction Modeling Framework for Current and Future Climate Storm Events
Abstract
Millions of people in the US are affected by power outages annually, impacting everyday life and the economy. A substantial portion of these power outages are caused by storm events, including thunderstorms, hurricanes, tropical storms, strong winds, and heavy rains. At the University of Connecticut an outage prediction model (OPM) has been developed for forecasting storm outages. The OPM began serving utilities in the Northeastern US in 2015 using weather, infrastructure, land cover, and elevation variables to generate predictions from non-parametric machine learning (ML) ensembles. The model provides probabilistic operational forecasts to utilities which are used as part of response planning to large storm events by the utility stakeholders.
This study leverages knowledge from the UConn OPM and utilizes a similar ML framework to build a community OPM for predicting customer outages during large storms along the US Eastern Seaboard. Proxies for proprietary infrastructure and variables to account for regional diversity are utilized. Events are selected from the NOAA Storm Events Database based on a threshold for associated financial damages. Once selected, events are windowed using information on the start date in each county and data from non-utility-owned customer outage data. Correlations between customer outages and utility troublespots for the Northeast domain, where outage data directly from utilities are available, are used to validate the customer outage data. Partial dependence plots are used to evaluate input variable significance and determine their relative influence on predictions. Model performance is tested using leave-one-storm-out cross validation, where a model is trained on all storms in the dataset except for one, and predictions are generated for the withheld storm. This process is repeated iteratively until each event has been withheld from training and had predictions generated for it. Predictions are then evaluated against actual outages using absolute and percentage errors, demonstrating that the model is capable of predicting the peak number of customer outages accurately, and showing promise for the ultimate goal of determining return periods of outages under current and future climate scenarios to help the public and utilities with resiliency and response planning.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMSY31A..05T