Ph.D. Proposal: Kingsley Udeh

Doctoral Dissertation Proposal

Title: Predicting Customer Outages in Spatiotemporal Utility and Weather Data with Deep Neural Networks

Student: Kingsley Udeh

Major Advisor:  Dr. Derek Aguiar

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

Date/Time: Wednesday, December 16th, 2020, 1:00PM 

Meeting link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=mf19c2306950ff7e4e9eda92aaca18f56

Meeting number: 120 108 9999

Password: Hpmd4Cvfc58

Join by phone: +1-415-655-0002

Access code: 120 108 9999

Abstract:

Electric utility emergency response personnel are increasingly required to face the impact of more frequent severe weather events on electric distribution networks. This increased frequency and severity of storms costs between $20 billion and $55 billion to the U.S. economy annually due to storm-related outages. To prepare an appropriate and timely response, the primary challenge for electric service utility managers is to accurately model the impact of extreme weather on the electrical grid prior to the event. The current emphasis on modeling severe weather impact on electric distribution networks is focused on electric network resilience and power grid modeling. However, modeling utility customer outage is a key component for advancing emergency preparedness and power outage restoration plans. Accurate customer outage modeling improves customer confidence in electric utility service providers and reduces downtime during a power outage event, but requires the collection of spatiotemporal weather and outage data and the development of robust prediction methods.

In this dissertation proposal, we consider customer outage forecasting from an Internet of Things perspective and develop deep learning prediction models from distributed temporal and spatial weather and outage data. We consider two problems: (1) forecasting local customer outages from real-time weather data and customer outages of external regions affected by similar weather patterns, and (2) forecasting customer outages solely from weather data. We develop novel multivariate and multi-step customer outage forecasting methods for both problems based on weather and utility outage data that are collected across ten New York counties. For the first problem, we include customer outages as both dependent and independent variables to predict customer outages in a multi-step manner where each forecast horizon is a distinct model across the counties. We observe significantly improved prediction accuracy in our deep learning models when compared to an autoregressive persistence model.

In ongoing work, we develop methods to forecast customer outages using only weather data, providing utility managers customer outages predictions directly from weather forecasts in preparation for emergency storm response. We develop four novel architectures based on 1-D convolutional, recurrent, and dense encoder-decoder neural networks that cluster counties at various granularity to exploit shared patterns within groups. Our first two models predict hourly customer outages (a) for each county independently and (b) by clustering all counties into a single model. Next, we combine the data from neighboring counties to train models to forecast hourly customer outages of a target county. Lastly, we incorporate the geographical weather station data of New York State to first cluster weather stations, and then build geographically specific models for each cluster. Preliminary results demonstrate that our single county architecture has better performance in 60% of the counties. Finally, we describe future work that includes building a framework for pre-training an encoder-decoder based model on weather stations with high data availability to make predictions in counties with sparse data using transfer and ensemble learning.