Conference Paper

Proceedings of Building Simulation 2023: 18th Conference of IBPSA

Unsupervised estimation of residential electricity outages from aggregated meter data

Praveen Radhakrishnan, Kingsley Nweye, Ting-Yu Dai, Zoltan Nagy
University of Texas at Austin, United States of America

DOI: https://doi.org/10.26868/25222708.2023.1403
Abstract: Winter Storm Uri left a devastating impact on Austin’s power grid and highlighted the need for preparation in the face of extreme weather events.To better understand the load lost during Uri and identify vulnerable areas, an ensemble approach was used, combining K-means clustering, Local outlierfactor algorithm for possible outage prediction. Linear regression and random forest models were used for load lost percentage estimation. Current researchdoes not focus on residential outage analysis during Winter Storm Uri and most of the studies are supervised learning which fails in the absence of truelabels. Our work fills this gap by using unsupervised learning and ensemble methodology to obtain robust estimates. Our results indicate that from Feb15th- 17th, 25 zip codes showed indications of possible power outages with an average load lost of around 60%. We also found 11 vulnerable zip codes with anaverage load lost of over 50%. Our work contributes to estimating demand losses during extreme weather events, which can help understand electricity supplydemand for potential future events. It also helps identifying vulnerable areas (by zip codes) which can help prioritize resources. Our methods use aggregated residential meter data to avoid privacy concerns and can be applied to other locations as well.
Keywords: Machine Learning, Winter Storm Outages, Load Lost, Unsupervised Learning, Power Outage Modeling
Pages: 1015 - 1022
Paper: