INTRODUCTION
Title 1
Title 2
Title 3
Title 4
Title 5
Title 6
Projected changes in hurricane-induced power outages in a future climate
Keep Scrolling
Hurricanes, or more generally tropical cyclones, are known to cause extensive damage to electric power system infrastructure, as high winds and heavy precipitation often leave felled trees and downed power lines in their aftermath. This can result in long-duration power outages following a storm, as utility crews work to repair damaged poles and lines as quickly as possible. There is a growing body of research focused on how climate change will affect future storm characteristics in terms of frequency, intensity, and location. While studying the change in hurricane behavior itself is the first step, the motivation for this case study lies in gaining a better and more nuanced understanding of what a change in storm characteristics might mean for power system impacts. Namely, how might climate change affect the risk of hurricane-induced power outages?
As an initial demonstration resulting from a collaboration between EPRI and PNNL, we’ve coupled synthetic storm tracks, created under both current and future climate conditions, with a prediction of power outages resulting from each storm to characterize broad trends at the county scale across U.S. Gulf and Atlantic coast states. Explore the map below to see how outage events – looking at different measures of frequency and magnitude – are projected to change in a future climate. Each layer offers a different picture of risk and can be used to conceptualize a community’s lived experience, inform a utility company’s planning decisions, or bound expectations for increasing risk. This map provides a way to dive into these results at the county level, showing several different metrics of interest:
Hurricane Outage Map
1.
EPRI subject matter experts facilitate discussions to launch project and during each step.
2.
Assessment team sets the number and type of locations and the project timeline.
3.
Assessment team identifies the weather and climate hazards for each location.
4.
Assessment team defines critical climate metrics, such as design thresholds for each hazard and location.
5.
Analysts evaluate historical baselines, incorporate future projections, and characterize uncertainty.
6.
Project team communicates results to different and diverse audiences.
EPRI 3420 Hillview Avenue, Palo Alto, California 94304-1338 USA • 800.313.3774 • 650.855.2121 • askepri@epri.com • www.epri.com
© 2024 Electric Power Research Institute (EPRI), Inc. All rights reserved. Electric Power Research Institute, EPRI, and TOGETHER…SHAPING THE FUTURE OF ENERGY are registered marks of the Electric Power Research Institute, Inc. in the U.S. and worldwide.
For further details regarding this research or Climate READi, reach out to ClimateREADi@epri.com or visit epri.com/research/sectors/readi.
These results are subject to limitations. For one, these projected power outages assume a static system without hardening investments. In addition, the future period modeled here represents the 2066–2100 timeframe under a high emission scenario (SSP5-8.5), and there is still significant uncertainty as to how climate conditions will evolve. For these reasons, these projected results represent a highly conservative estimate of changing risk, i.e., how bad might things get if we don’t proactively invest to prevent damages and if climate change continues at an aggressive pace.
This work was done through a partnership between EPRI and the team of researchers at the Pacific Northwest National Laboratory (PNNL) responsible for the development of the Risk Analysis Framework for Tropical Cyclones, or RAFT. RAFT combines physical modeling with statistics and machine learning to create synthetic tropical cyclones. This tool takes climate conditions into account to produce track, wind speed, and precipitation data for each storm initiation. By creating large numbers of synthetic storms in the Atlantic Basin under both current and future climate conditions, the PNNL team has explored projected changes in hurricane risks for affected communities [1, 2].
Using the RAFT data as a starting point, EPRI and PNNL worked together to explore what it might look like to couple climate-informed tropical cyclones with predicted power outages for each synthetic storm in both a current and future climate. This combination enables a deeper look into not only how storms may change but how the impact of storms may change, with implications for power system planning in areas of grid hardening, crew preparedness, and proactive investment. Although more work is needed to fully understand the drivers of change, these initial results offer an added dimension to our understanding of future risks by providing an assessment of changing consequences in addition to the changing hazard.
(b.) Change in number of major hurricanes per decade
(a.) Change in number of tropical storms per decade
These maps show the change in modeled frequency with which each county experiences winds of at least Tropical Storm strength (≥18 m/s, shown in (a)) and Major Hurricane strength (≥50 m/s, shown in (b)). The change is computed between the future period (2066–2100) and historical period (1980–2014) under the SSP5-8.5 emissions scenario. Each map shows the frequency change averaged over the synthetic tropical cyclone climatology generated in 9 different CMIP6 environments. The methodology is similar to Balaguru et al. (2023), with an expanded ensemble of CMIP6 models.
Creating a Power Outage Prediction Model
Projecting power outages for synthetic storms first requires a model that can translate storm characteristics, such as wind speed and precipitation at a given location, into resulting power outages, often represented as the number of customers or fraction of the population without power. For this work, we build upon statistical learning methods used previously in this space to train a model using historical power outage data [3]. More specifically, we used power outage data from 23 historical hurricanes that affected the continental United States and Puerto Rico between 2016 and 2023 with data obtained through EAGLE-I, a system that compiles data scraped from utility company websites in real time [4]. RAFT was used to create spatially detailed wind speed and precipitation data along the known track of each historical storm [5, 6]. This data was used alongside the outage data, population statistics, and geographic information at the county level to train and refine the power outage prediction model used for this case study. The resulting model is able to capture spatial trends of outages with reasonable estimates of magnitude.
Read More
Hurricanes in a Changing Environment
To model how hurricane risk will change over time, we gathered climate conditions from nine state-of-the-art fully-coupled climate models belonging to the Coupled Model Intercomparison Project phase 6 (CMIP6). Data are obtained for both the historical (1980-2014) and future (2066-2100) periods under the SSP5-8.5 warming scenario, representing a scenario with very high carbon emissions. Under this scenario, the radiative forcing of greenhouse gases is projected to reach 8.5 W m-2 by the year 2100. These climate conditions are then fed into RAFT’s synthetic hurricane model to generate nearly 50,000 storms under each climate scenario and CMIP environment, resulting in a sample of 900,000 synthetic hurricane events representative of the current and future climate periods. We apply bias correction to the modeled intensities to eliminate systematic biases that may be present in the CMIP6 environmental conditions [8].
The Changing Power Outage Risk Landscape
By coupling synthetic tropical cyclones with predicted power outages from each storm, we provide a picture of how the risks of hurricane-induced power outages are projected to change based on climate-driven changes to hurricane behavior. With such a large sample size of synthetic storms, we can pick up on changing patterns of regional risk, capturing expectations as well as the tails of the distributions to better capture a broader view of risks. In addition, our outage model is built to predict the fraction of customers expected to be without power from a storm, and this allows us the flexibility to adjust to changing populations.
Our results show a dramatic increase in risk along the entirety of the US coastline, particularly in the Gulf Coast, Florida, and Puerto Rico, driven by an increase in the number of tropical cyclones and major hurricanes projected in a future climate.
What Might Future Storms Look Like?
Understanding the Risk for Key Metro Areas
Limitations of this Case Study
While this case study offers useful insights into how the risk from hurricanes may change in the future, there are a number of limitations that should be taken into account when interpreting results. Most importantly, the outage predictions are based on a data-driven model, with data at the county level for storms that have occurred in the past. The model resolution is limited to the resolution (and quality) of the data used in training, and this research scope was designed to provide high-level information on changing risk, not to enable detailed storm preparation or response strategies. In addition, disruptions in the power system cannot be linked to specific parts of the network, service territory, type of infrastructure, or failure mode causing the outage. Thus, the outage model cannot capture any change in expected performance due to resilience upgrades on the grid, such as undergrounding lines or enhanced vegetation management.
Results shown here also represent only one future time period (2066-2100) and one emissions pathway (SSP5-8.5). There is substantial uncertainty regarding how the climate will evolve in the coming decades, and the SSP5-8.5 emissions pathway represents a higher emissions scenario and thus an upper bound for how drastic changes may be.
References
Balaguru, Karthik, Wenwei Xu, Chuan-Chieh Chang, L. Ruby Leung, David R. Judi, Samson M. Hagos, Michael F. Wehner, James P. Kossin, and Mingfang Ting. "Increased US coastal hurricane risk under climate change." Science advances 9, no. 14 (2023): eadf0259.].
Xu, Wenwei, Karthik Balaguru, David R. Judi, Julian Rice, L. Ruby Leung, and Serena Lipari. "A North Atlantic synthetic tropical cyclone track, intensity, and rainfall dataset." Scientific Data 11, no. 1 (2024): 130.
Guikema, Seth David, Roshanak Nateghi, Steven M. Quiring, Andrea Staid, Allison C. Reilly, and Michael Gao. "Predicting hurricane power outages to support storm response planning." Ieee Access 2 (2014): 1364-1373.
EAGLE-I, https://EAGLE-I.doe.gov/login
Holland, Greg J., James I. Belanger, and Angela Fritz. "A revised model for radial profiles of hurricane winds." Monthly weather review 138, no. 12 (2010): 4393-4401.
Lu, Ping, Ning Lin, Kerry Emanuel, Daniel Chavas, and James Smith. "Assessing hurricane rainfall mechanisms using a physics-based model: Hurricanes Isabel (2003) and Irene (2011)." Journal of the Atmospheric Sciences 75, no. 7 (2018): 2337-2358.
Pedregosa, Fabian, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel et al. "Scikit-learn: Machine learning in Python." the Journal of machine Learning research 12 (2011): 2825-2830.
Cannon, Alex J., Stephen R. Sobie, and Trevor Q. Murdock. "Bias correction of GCM precipitation by quantile mapping: how well do methods preserve changes in quantiles and extremes?." Journal of Climate 28, no. 17 (2015): 6938-6959.
Balaguru, Karthik, Gregory R. Foltz, L. Ruby Leung, Samson M. Hagos, and David R. Judi. "On the use of ocean dynamic temperature for hurricane intensity forecasting." Weather and Forecasting 33, no. 2 (2018): 411-418.
Projected changes in hurricane-induced power outages in a future climate
The number of hurricanes (i.e., exposure to hurricane-force winds) occurring per decade, with a map layer showing simulated values for the current climate, the projected values for the future climate, and the change between the two
The number of outage events experienced per person per decade, with a map layer showing simulated values for the current climate, the projected values for the future climate, and the change between the two
The number of severe outage events per decade (here defined as events with at least 50% of the county population without power), with a map layer showing simulated values for the current climate, the projected values for the future climate, and the change between the two
The 20-year return period outage magnitude (here defined as the percentage of population without power), with a map layer showing simulated values for the current climate, the projected values for the future climate, and the change between the two
The power outage model is based on decision trees, specifically the state-of-the-art ensemble method called Histogram Gradient-Boosted Regression, which chains together many decision trees by having each tree correct for the errors of the previous one [7]. We perform feature selection and hyperparameter tuning, which yielded a model that selected the following features:
Maximum wind speed
Population
Maximum rainfall rate (averaged over the county’s area)
Duration of hurricane-force winds (≥ 33 m/s)
Poverty rate
Percentage of population living in coastal counties at 10 meters elevation or below
For the final set of variables, we also performed a permutation test to determine how much the model relies on each feature [7]. Unsurprisingly, maximum wind speed is a dominant predictor. Notably, our model selects rainfall rate as the third-most important feature, enabling this model to distinguish among storms with varying degrees of local precipitation.
A comparison of actual outage fraction, as collected through EAGLE-I (left), and the predicted outage fraction (right) for Hurricane Idalia (2023). A leave-one-out scheme, where the model is trained on all storms except this one, is used to get an unbiased view of the model’s ability to generalize.
Peak fraction of population without power
Peak fraction of population without power
If your browser has trouble loading this map click here.
Even with a focus on aggregated, regional changes in risk, it can be useful to dive into specific locations to understand the changing landscape. For example, the figures below each show nine tracks generated within the climate conditions of the nine different CMIP6 models, with the same randomly-selected starting location and time of year. As an initial takeaway, these figures illustrate how much variation there is between the modeled climate conditions, as the storms have wildly different tracks and intensities. Second, we can see the influence of climate change in the difference between the two, with the future climate showing consistently stronger storms that make landfall in the United States more frequently and penetrate farther inland when they do.
Sample storm tracks with associated intensity created under the current climate and future climate initiated at the same location and time of year. Future storms show higher intensities and more storm tracks making landfall at high intensities.
We can also get a sense of the wide range of storm tracks that might affect a specific location. Below are figures showing samples of 100 storms that caused an outage of 25% or more of the population in the counties containing Houston, Texas; Miami, Florida; New Orleans, Louisiana; New York, New York; and San Juan, Puerto Rico. We can see that storms affecting some locations like Houston have relatively consistent shapes and development patterns, while places like Miami and New York are affected by a much more diverse set of storm tracks.
Maximum wind speed (knots)
Maximum wind speed (knots)
Maximum wind speed (knots)
Maximum wind speed (knots)
Sample storm tracks and intensities affecting the Houston, Texas area
Sample storm tracks and intensities affecting the New Orleans, Louisiana area
Maximum wind speed (knots)
Sample storm tracks and intensities affecting the Miami, Florida area
Sample storm tracks and intensities affecting the New York, New York area
Maximum wind speed (knots)
Maximum wind speed (knots)
Sample storm tracks and intensities affecting the San Juan, Puerto Rico area
Read More
Houston, TX
New Orleans, LA
Tampa, FL
Miami, FL
Charleston, SC
Norfolk, VA
Washington, DC
Philadelphia, PA
Manhattan, NY
Boston, MA
San Juan, PR
Harris County Orleans Parish Hillsborough County Miami-Dade County Charleston County Norfolk City District of Columbia Philadelphia County New York County Suffolk County San Juan Region
0.49 1.06 1.05 0.96 0.92 0.74 0.29 0.30 0.32 0.85 2.16
0.84
1.81 1.98 2.11 1.29 1.07 0.35 0.44 0.47 1.49 3.49
+72% +70% +89% +119% +40% +44% +21% +49% +47% +76% +62%
Metro Area
County
Current Climate
Future Climate
Percent Change
Projected Outage Events per Person per Decade
Although the RAFT model represents the Atlantic hurricane climatology reasonably, there are a few limitations associated with the methodology employed. For instance, cyclogenesis in RAFT is based on a Gaussian spatio-temporal distribution derived statistically based on observations. In other words, for each month, we take a certain number of seeds and randomly distribute them over the climatological locations of storm genesis. Subsequently, we allow the environment to determine whether the seed grows into a hurricane or not. While this method accounts for the effects of climate change on cyclogenesis to an extent, it can potentially underrepresent the number of storms that form over the traditional regions of genesis. Also, it does not account for new regions that may become favorable for cyclogenesis in future. This approach was chosen in part due to the current lack of consensus with regards to how storm genesis patterns and rates may change in the future. Similarly, another potential limitation of the methodology relates to the simplistic representation of oceanic processes within RAFT [9]. Currently, the sea surface temperature is used to represent the role of the ocean in storm intensification. However, studies have shown that upper-ocean heat content and strong salinity stratification may also modulate hurricane intensity significantly in certain regions.
Beyond the RAFT methodology, a few limitations are also associated with the outage model, and climate model simulations used to project storm climatology into the future. To train the machine learning-based outage model, we used data provided by EAGLE-I for the 8-year period 2014-2022. Although the trained model parameters are statistically significant, it is possible that a longer record of outage data can enhance the model skill. Finally, the CMIP6 climate model simulations used to generate the environmental conditions that are fed into RAFT are known to have systematic biases. To mitigate the influence of these errors on hurricane projections, we bias correct all historical and future synthetic tropical cyclone intensities using the statistical method of quantile delta mapping to ensure that intensity distributions remain realistic [8].
Read More
current climate
future climate
With results at the county level, we can dive into localized changes in projected risk for a range of locations. This table presents the model results for the number of outage events experienced per person per decade for eleven metropolitan areas across the Atlantic coast of the U.S. and Puerto Rico, showing the simulated number for the current climate, the projected number for the future climate, and the percent change between the two. Results show increasing risk across the board, although the magnitude of the increase varies significantly with location.
3002030610
July 2024
