- Department of Civil and Environmental Engineering, University of Connecticut, Storrs, CT, United States
- Eversource Energy Center, University of Connecticut, Storrs, CT, United States
- Los Alamos National Laboratory, Los Alamos, NM, United States
Abstract
Accurate and timely power outage predictions are crucial to the allocation of valuable resources and the mitigation of potential damages to power distribution systems. However, the effectiveness of outage forecasting is commonly constrained by the abundance and/or quality of the available information, with respect to the spatiotemporal development of intense storms. Due to the low frequency of occurrence associated with impactful storms, as well as the complexity of weather patterns, data-driven power outage prediction models are tasked with providing estimates based on spatially and temporally unbalanced datasets, which may significantly affect their performance. A proposed solution has been the integration of a hurdle model (see Watson et al., 2022), which can be employed to identify inconsequential events with respect to power outages and, therefore, balancing the training data and shifting the focus to impactful occurrences. In this study, we attempt to robustly identify potentially adverse weather activity and allow for targeted power outage prediction, by incorporating existing storm reports offered by the National Oceanic and Atmospheric Administration (NOAA), along with hourly ERA5 atmospheric reanalysis data over a 25-km grid and a period of 40 years. To do so, we develop an automated framework based on data-driven techniques, which is trained to provide predictions of the location and time that a winter storm occurs over the Contiguous United States (CONUS). Given the importance of real-time predictions in effective hazard mitigation, we conduct training and testing under a setting that simulates real-time conditions, thus enabling the use of the proposed framework with short-term weather forecasts. In this context, the developed approach aims at providing a versatile and scalable tool that supports and enhances operational power outage prediction, which could be potentially generalized for global applications due to the diverse climatological conditions that is trained on.
References
Watson, P. L., Spaulding, A., Koukoula, M., & Anagnostou, E. (2022). Improved quantitative prediction of power outages caused by extreme weather events. Weather and Climate Extremes, 37. https://doi.org/10.1016/j.wace.2022.100487

