Study overview

In the present analysis, we use locality-level (n = 1,764) power outage, weather, urbanicity, and social vulnerability data from January 1, 2017-December 31, 2020 to assess the impact of extreme weather on power outage distributions and durations, while considering vulnerability factors.

Power outage ascertainment

We obtained information on customers without power in 30-minute increments within localities from the NYS Department of Public Service from 2017–2020 [23]. We excluded localities with <30 customers or >5% temporal missingness over the study period, resulting in 1,764 (94.6%) included localities. The dataset also included locality boundaries in a shapefile format and the number of customers served, and the electrical utilities operating in each locality. A power operating locality is the smallest level at which outage data is reported to the state and is comparable in size to zip code tabulation areas; the localities serve ~11,000 customers, on average. These customers include residential, commercial, and electrical meters. We aggregated the 30-minute data to the hourly level to match our weather metric data.

Weather metric ascertainment

We primarily sourced weather data from land-surface model estimates. We pulled data on average temperature, windspeed, and precipitation at the hourly-level from forcing data for Phase 2 of the North American Land Data Assimilation System (NLDAS-2) [24]. NLDAS-2 provides gridded estimates of each of these variables with ~14km² resolution. We obtained hourly snowfall data from the ERA5-Land reanalysis dataset, which is available hourly with ~11km² resolution [25]. We aggregated the gridded datasets to locality boundaries via areal weighting using Google Earth Engine [26]. We collected lightning strike data from the International Space Station Lightning Imaging Sensor, which records the time and location of lightning strikes, starting in March 2017 [27]. To match the spatial and temporal resolution of the other data, we calculated the hourly number of lightning strikes in each locality. Since lightning data was only available beginning March 2017, to preserve as much data as possible, we assumed no lightning strikes occurred during the first two months of 2017. Lightning strikes were most common from April to August (n = 1624 total strikes), and only 4 total strikes occurred in January and February 2018–2020 combined.

Urbanicity ascertainment

Because of unique population, outage, and weather profiles, we ran all analyses separately for NYC, non-NYC urban, and rural regions of NYS. To classify localities into their respective regions, we first used 2010 US Census on the percent of the total population classified as urban/rural at the block group. We interpolated this to the locality-level with areal weighting. When localities had >50% of inhabitants designated as rural, we assigned the locality a rural classification [28]. We distinguished between NYC and non-NYC urban, using the county indicators included in the power outage data. We assigned localities listed in the New York, Bronx, Kings, Queens, and Richmond counties as NYC. The final classification of each locality used in all further analysis is available in Fig 1 and S1 Fig.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The spatial boundaries of each power operating locality in NYC, non-NYC urban, and rural regions of NYS.

Republished from The New York State Department of Public Service under a CC BY license, with permission from The New York State Department of Public Service original copyright 2020.

https://doi.org/10.1371/journal.pclm.0000364.g001

The lagged effect of severe weather on power outages, by severe weather metric and by region

To assess the lagged and non-linear effects of weather on the proportion of customers without power in a locality, we used negative binomial generalized additive models, an extension of generalized linear models that allow for smoothed or nonlinear fits [29]. We selected the negative binomial fit after assessing regression diagnostics using the DHARMa package across a range of model types including Poisson, negative binomial, Tweedie, and zero-inflated Poisson [30]. We modelled weather using distributed lag nonlinear models (DLNMs), which simultaneously allow for the modelling of nonlinear exposure-outcome relationships and delayed events [31]. DLNMs are also advantageous because they provide constrained lag terms which account for high temporal autocorrelation, a feature of the hourly weather data. We adjusted for seasonal and temporal trends by including a natural spline term with 6 knots per year for the date [32]. We also included fixed effects for the utility that serves each locality to account for a lack of spatial independence. We used the Akaike information criterion to select the appropriate degrees of freedom for both the exposure and the lag out of a range of 2–5. We used the Moran’s I to assess spatial autocorrelation in the final models (S15–S17 Figs) [33].

Outage, severe weather, and severe weather-driven outage classifications

For the remainder of our analyses, we conceptualized outages, severe weather events, and a severe weather-driven outage each as binary variables. We relied on the severe weather and power outage literature, our weather and outage data’s distribution, and the results from our first objective, which assessed the lagged effect of severe weather on outages.

In previous literature, outages have been defined as hours where the proportion of customers without power exceeds the 90^th percentile of the hourly proportion without power statewide [34, 35]. We adapted a similar definition for our analyses, but instead defined the 90^th percentile separately for each region (NYC, urban non-NYC, and rural) to prevent the undercounting of outages in NYC, where the population in a locality was much larger than in rural regions. The 90^th percentile of customers without power was 0.04%, 3.4%, and 15.2%, for NYC, non-NYC urban, and rural, respectively. As an example, for our binary outage metric, any hour where the percentage of customers without power exceeded 3.4% in a non-NYC urban power operating locality was classified as an outage.

We similarly used percentile classifications from our data distribution to determine the presence of a severe weather event. We hoped to identify pertinent thresholds in our first objective, but we observed the general pattern of outages increasing at the extreme of weather metric distributions rather than distinct thresholds. Thus, we used a statewide 97.5^th percentile to define severe weather events. We chose to keep this metric statewide for interpretability. For most of the continuous metrics (precipitation, wind speed, snowfall) we identified the 97.5^th percentile of each metric from 2017–2020, and then any hour above that threshold we defined as a severe-weather event. For temperature, the definition slightly deviated. To define extremely hot hours, we calculated the 97.5^th percentile of temperatures during the during the hot months (May-Sept) and to define extremely cold hours, we calculated the 2.5th percentile during cold months (Oct-April). The presence of lighting was collapsed to a binary depending on whether a locality experienced any lightning strikes during that hour.

Finally, to create a severe weather-driven outage definition, we used the two previously described definitions and a lag component. Using the results from our first objective, it appeared that much of the effect of the extreme events was immediate (within 8–12 hours of exposure). Therefore, we decided to use 8 hours as our window of interest. Our final definition for a severe weather-driven outage was an outage that started either within the same hour of a severe weather event or within 8 hours following a severe weather event.

The most prevalent and hazardous combinations of severe weather types that contribute to outages

Once we defined outages and severe weather-driven outages, we used these definitions to achieve study objectives 2–4. We wanted to identify the severe weather events (or combination of severe weather events) that lead to the most significant power outages in frequency, duration, or proportion of customers impacted. To do so, we first calculated the frequency, average duration, and average proportion of customers impacted for each classification of severe weather driven outage (e.g., wind + precipitation-driven, wind-driven). We omitted snow-driven outages from this analysis to reduce the number of classes and redundancies between groups, as snow is part of the precipitation estimates.

We then calculated a severe/non-severe outage ratio for each severe weather combination. We calculated this using the following formula: where, for each severe weather combination, we divided the number of outages from a severe weather combination, by the number of hours with the severe weather combination. Then to standardize this across non-severe outages, we divided the numerator by the number of outages without any severe weather event divided by the hours without any severe weather events. A severe/non-severe ratio > 1 indicates that outages are more likely due to that severe weather combination, i, than times without any severe weather events. The ratios can then be compared across severe weather types to quantify which combinations are the most likely to cause outages.

The distribution of the severe weather driven outages across the state

With outages, severe weather driven outages, and pertinent combinations of severe weather types defined, we performed descriptive analyses by mapping the frequency, average number of customers without power, and the duration of outages during severe weather driven and non-severe weather driven outages for each region. We also calculated the percentage of outages that were due to severe weather events in each locality. Finally, we presented the frequency of severe weather driven outages by type. For this mapping and for subsequent analyses, to reduce the number of analyses, we present outages driven by each of the six weather metrics of interest (cold, heat, lightning, precipitation, snow, and wind). Thus, outages caused by both wind and precipitation would be counted in both the wind and precipitation panels of Fig 4.

Disparities in the frequency and duration of severe weather-driven outages

We then aimed to assess the association between social vulnerability and outage exposure.

Social vulnerability classification

We used the 2020 Social Vulnerability Index (SVI) created by the CDC/ATSDR to determine the social vulnerability of each power operating locality [36]. The CDC/ATSDR designed the index to identify communities that may need support during disasters like those driven by climate change. It incorporates 16 social factors from the 2016–2020 American Community Survey that capture several aspects of outage-related vulnerability (e.g., poverty, disability, housing type, age, English-language proficiency). The final index score ranges from 0–100 where higher values indicate increased social vulnerability. We downloaded the 2020 SVI at the census tract level, and used areal interpolation to determine the scores using to the power operating locality using a target-density weighting approach [37]. Finally, for each urbanicity region, we grouped each locality into their respective SVI quartile resulting in the final distribution displayed in S2 Fig.

Statistical analysis–the frequency and duration of severe weather-driven outages

To estimate disparities in the distribution of outages, we used targeted maximum likelihood estimation (TMLE) [38], a doubly robust maximum-likelihood–based approach. We estimate the average treatment effect of being in the highest quartile of SVI versus all others on risk of a severe weather-driven outage. We performed these analyses separately for each of the rural, non-NYC urban, and NYC regions. We implemented TMLE using the ltmle [39] and SuperLearner [40] packages in R. We ran three sets of sensitivity analyses. In the first, we stratified the analyses by year. In the second, we re-ran the analyses raising the thresholds for the number of severe-driven outages from n = 1+ to n = 3+ and then n = 5+. In the third, we included latitude and longitude to account for possible spatial confounding.

To understand disparities in the duration of outages, we present the duration of each outage type by SVI quartile along with results from Kruskal−Wallis tests [41]. All code for the conducted analyses are available on GitHub: https://github.com/nina-flores/nys_severe_weather_outages.

The lagged effect of severe weather on power outages, by type and by region

Using DLNMs, we examined both nonlinear exposure-outcome relationships and delayed events, for each weather metric of interest. An example of the output from these analyses showing the lagged relationship between hourly temperatures and the proportion of customers without power for non-NYC urban localities is displayed in Fig 2. Here, we visualize the relative rate of customers without power as temperatures increase or decrease away from the median of 9.8°C across 24 hours of lags. By focusing on the same hour of exposure, lag 0, we found that an increase in temperature to 30°C in non-NYC urban localities leads to 3.8 (95% CI: 3.6–4.1) times the rate of customers without power, during that same hour of the temperature increase, compared to the median temperature. However, we observed a 5–15-hour lag between extreme cold temperatures and an increased rate of outages. We created a Shiny dashboard so that readers could view the 3D plots (Fig 2A) and 2D plots across any number of lags up to 24 (Fig 2B) for each of the weather metrics (precipitation, snowfall, temperature, and windspeed) and each of the regions of analysis (NYC, non-NYC urban, rural; https://oyb6ek-nina-flores.shinyapps.io/severe-weather-app/).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

The lagged relationship of hourly temperature and the proportion of customers without power for non-NYC urban localities at all lags 0–24 (a), and at lag 0 (b). Panel b is constructed by slicing panel a where the lag = 0, as shown by the light blue plane at lag 0. All rates are relative to the overall median temperature of 9.8°C, shown by the dotted blue vertical line (b). To view these figures for all other weather metrics (precipitation, snowfall, temperature, and windspeed) and regions of analysis (NYC, non-NYC urban, rural), please visit our shiny dashboard: https://oyb6ek-nina-flores.shinyapps.io/severe-weather-app/.

https://doi.org/10.1371/journal.pclm.0000364.g002

We found that the weather metrics most strongly associated with power outages varied by region: in NYC and non-NYC urban areas, precipitation led to the largest rate ratios whereas in rural NYS, extreme wind led to the largest rate ratios. However, no region had a clear threshold at which any of the weather metrics distinctly increased outages. Rather, there was a smooth trend that outages increased as each weather metric became more extreme.

Lagged effects differed by weather metric and region. For instance, in non-NYC urban areas, extremely hot temperatures (>30°C) were most strongly associated with immediate increases in the rate of outages whereas extremely cold temperatures (<-10°C) were most strongly associated with increases in the rate of outages after 6–8 lagged hours. However, in rural regions of NYS, the impacts of extremely hot and extremely cold temperatures were most observable at lag 0. Though there was heterogeneity in the lagged effects, we observed that the effect peaked across most weather metrics and regions after 8–12 lagged hours, which influenced our final definition of weather-driven outages.

The most prevalent and hazardous combinations of severe weather types that contribute to outages

We identified outages that exceeded the region (NYC, non-NYC urban, and rural) specific 90^th percentile of customers without power and described their summary statistics (Table 1). By calculating the frequency of outages, their average duration, and the average proportion of customers without power during outages, by severe weather cause, we found that wind was the most frequent and the strongest single predictor of prolonged outages across all 3 regions (Table 1). Following wind, precipitation and heat alone also had consistently high frequency and durations across regions. Extreme cold alone had varying impacts across regions. For instance, in rural regions of NYS, outages driven by extreme cold were less likely to induce prolonged outages than non-severe weather conditions; however, in NYC, extreme cold conditions were more likely to cause prolonged outages than non-severe weather conditions.

When comparing multiple severe weather metrics simultaneously, we found that some combinations of multiple severe weather events led to longer or more widespread outages than single causes alone. The combination of extreme precipitation + wind led to the longest average duration for any severe weather type across all three regions, with an average duration of 20.2 hours in NYC, 18.5 hours in non-NYC urban, and 12.5 hours in rural NYS.

Using a severe/non-severe weather ratio to understand the impact of each weather metric on outages, we found that, though outages associated with lightning were relatively infrequent, they were the most likely single severe weather event to co-occur with outages across all 3 regions (Table 2). Following lightning, precipitation and wind also had consistently high severe/non-severe weather ratios across regions. Extreme cold alone had varying impacts across regions. For instance, in rural regions of NYS, outages driven by extreme cold were less likely to induce outages than non-severe weather conditions, however, in NYC, extreme cold conditions were more likely to cause outages than non-severe weather conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The frequency and severe/non-severe weather ratios for each weather driven outage.

https://doi.org/10.1371/journal.pclm.0000364.t002

When comparing multiple severe weather metrics simultaneously, we found that, generally, the combination of multiple severe weather events had higher ratios than single events alone. Overall, the top 5 combinations driving outages of multiple types, heat + lightning + precipitation, lightning + precipitation, lightning + wind, heat + precipitation, and precipitation + wind, all had ratios greater than lightning’s 105. Of note, heat + precipitation and precipitation + wind were both frequent (caused 718 and 1,878 outages, respectively) and had high severe/non-severe weather ratios.

Across NYS, 39.9% of all outages co-occurred with severe weather (Table 1). However, there was heterogeneity in this percentage across regions and localities. Severe weather contributed to over 50%, 85%, and 87% of all outages in some NYC, non-NYC urban, and rural localities, respectively. Across all 3 regions, generally, severe weather-driven outages impacted larger percentages of electrical customers and had longer durations than non-severe weather-driven outages (S3–S6 Figs and Table 2). In maps of the frequency of severe weather-driven outages overall and by weather metric, we found that certain localities were vulnerable to outages across all severe weather metrics. In NYC, localities in Queens, the Bronx, and Staten Island experienced the most severe weather-driven outages, with localities in Queens experiencing outages driven by each weather metric (cold, heat, lightning, precipitation, snow, and wind; Fig 3). In non-NYC urban regions, the most frequent severe weather outages occurred on Long Island and in the Hudson Valley and in rural regions, the most frequent severe weather outages occurred in North and Central NYS (S7 and S8 Figs).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3.

The frequency of (a) any severe weather driven outage and the frequency of outages co-occurring with extreme (b) cold, (c), heat, (d) lightning, (e) precipitation, (f) snow, and (g) wind in NYC, from 2017–2020. Republished from The New York State Department of Public Service under a CC BY license, with permission from The New York State Department of Public Service original copyright 2020.

https://doi.org/10.1371/journal.pclm.0000364.g003

In analyses assessing whether differences in the distribution of outages were due to social vulnerability, we found different effects across urbanicity (Fig 4). We estimated that in NYC, had all regions been in the 4^th quartile of SVI the number of heat-, precipitation-, and wind-driven outages would have been 12.1% (3.3%, 21.0%), 14.8% (-0.5%, 30.2%), and 19.1% (8.5%, 29.8%) higher, respectively, versus if all localities had been in quartiles 1–3. We estimated that in non-NYC urban NYS, had all regions been in the 4^th quartile of SVI the number of heat-driven outages would have been 7.5% lower (-16.3%, 1.3%), versus if all localities had been in quartiles 1–3. Finally, in rural NYS, we estimated that if all regions had been in the 4^th quartile of SVI the number of snow-driven outages would have been 5.6% lower (-13.2%, 2.0%), versus if all localities had been in quartiles 1–3. Otherwise, SVI did not seem to have an impact on outage frequency for non-NYC urban and rural localities. These results were consistent with a sensitivity analysis that used higher counts of severe weather outages (n = 3+ and 5+, rather than n = 1+) as the outcome (S9 Fig). When stratifying analyses by year, in NYC we found significantly positive associations between SVI and snow-driven outages for the year 2018–2019 as well (S10 Fig). When adding latitude and longitude to the models to account for possible spatial confounding, most estimates remained the same. However, in NYC there was no longer a relationship between vulnerability and precipitation-driven outages (S11 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. The percent difference in the average treatment effect comparing the highest quartile of SVI (most vulnerable) to all others, presented for each outage cause and by urbanicity.

The percentages can be interpreted as the percent difference in the probability of an outage of each severe weather type had all localities been in the highest quartile of SVI (most vulnerable) versus if all localities had been in quartiles 1–3, presented for each outage cause and by urbanicity.

https://doi.org/10.1371/journal.pclm.0000364.g004

The duration of some outage types was also longer in regions with higher SVI; for example, outages driven by wind and precipitation lasted the longest in regions in the 4^th quartile of SVI (SVI Q1 = 15.0, SVI Q2 = 17.2, SVI Q3 = 15.8, and SVI Q4 = 18.0 hours, S1 and S2 Tables). However, this also varies by region (S12–S14 Figs). In NYC, outages with the longest durations in high vulnerability regions co-occurred with precipitation whereas in rural regions, these co-occurred with precipitation and snow.