Ph.D. Defense: Kingsley Udeh

Title: Computational Approaches for Predicting Customer Outages from Spatiotemporal Electric Utility and Weather Data

Ph.D. Candidate: Kingsley Udeh

Major Advisor:  Dr. Derek Aguiar

Associate Advisors:  Dr. Sanguthevar Rajasekaran, Dr. Sheida Nabavi, Dr. Emmanouil  Anagnostou, Dr. Diego Cerrai

Date/Time: Friday, September 9, 2022, 12:30 PM

Location:
 Meeting link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=mac27539ebec707b8af214b8836734085
Meeting number:  2620 418 6182
Password:  W4iMpfTXK43
Join by phone:  +1-415-655-0002 US Toll
Access code:  2620 418 6182

Abstract:

The increased frequency and severity of storms costs the U.S. economy between $20 billion and $55 billion annually. To prepare an appropriate and timely response, electric service utility managers must accurately estimate the impact of extreme weather on the electrical grid prior to a storm event. Most prior work focuses on modeling the physical power grid infrastructure and network resilience in the context of storms. However, accurate estimation of electric utility customer outages is integral for developing an effective emergency preparedness strategy, reducing power downtimes, and ultimately, improving customer confidence in electric utility service providers. In this dissertation, we consider customer outage forecasting from distributed temporal and spatial weather data across ten New York counties. We consider two primary computational problems: (1) autoregressive forecasting of county-level customer outages from historical hourly weather and customer outage data, and (2) forecasting county-level customer outages solely from retrospective and prospective weather data.

For the first problem, we consider both the autoregressive and covariate-dependent signatures of customer outage variation. We develop a novel neural network architecture based on convolutional and recurrent layers for county-level customer outage prediction (a) independently for each county, (b) simultaneously across all counties, and (c) for each county conditioned on data from all counties. We compare our methods against traditional autoregressive statistical models (ARIMA, ARIMAX and VARMAX) and a persistence-based model. Our methods achieve the highest performance in terms of root mean square error, median absolute error, Pearson correlation, and average relative error.

For the second problem, we consider customer outage forecasting using only weather data, which is important for modeling nonstationary phenomena like severe weather. We partition the problem into two components: blue skies and storms. Our method for blue skies borrows from previous work for autoregressive modeling of stationary non-storm outages. Our model for storms first computes outliers in customer outage data using a statistical framework based on a Poisson mixture model and deviations from an empirical null Poisson or Negative Binomial distribution. We call outliers in the weather data using isolation forests and match outage outlier intervals to weather outliers and vice versa using a greedy algorithm. After calling storms, we train Poisson regression, random forests, and k-nearest neighbors storm regression models whose outputs are combined with blue skies models in a single architecture. We demonstrate that careful modeling of both the blue skies and storm components yields an effective strategy for electric utility emergency preparedness.