2.1. Methods
This study utilized the ABM detailed in Walsh et al. [
25] to derive restoration scenarios for the different Sandy2112 outage predictions presented in Wanik et al. [
11]. Several adaptations have been added to the ABM to better represent the restoration process. First, the work zones for crews was updated to a higher spatial resolution from state-wide to area work center (AWC). Crews are assigned an AWC during the model setup process. Crews stay in that AWC until all of the outages there have been repaired. Once their AWC is complete, if any outages remain, crews will then be assigned a new AWC based on the search strategy used. For example, if the strategy is set to nearest outage, a crew in a completed AWC will find the nearest outage in a different AWC and then set that as their AWC to complete restoration. Crews will remain in the newly assigned AWC and not “return” to their original AWC until statewide restoration is completed. Daily crew counts by AWC were provided from a local utility company that has been used as the input for this study, as previously detailed in
Table 1. The crew counts were from the actual crews allocated for the restoration of Sandy2012-caused outages in Connecticut but have been kept the same for the initial test of Sandy2112 scenarios. Furthermore, about 75% of crews are simulated to be working during the daytime hours and 25% of crews are working during the overnight hours.
The ABM was calibrated based on utility data from Sandy2012 to get a baseline for Sandy2112 experiments. The outage locations, crews working, and historic restoration curve were used. The individual repair time for each outage location was unknown. A series of repair time ranges from a minimum of one hour to a defined maximum was tested and compared to the historic restoration curve of Sandy2012. It was found that a range of repair times of one to eight hours in 15-minute increments with uniform distribution best fit Sandy2012. We have applied that same range to Sandy2112; outages are randomly assigned a repair time between one and eight hours during the model setup process. Repairs for different infrastructure require different time allocations. For example, on average, it takes 2 h to fix a wire, 4 h to fix a transformer, and 8 h to fix a downed pole. Applying a range of repair times implicitly assigns different fault types to the system.
During calibration, we used four different search strategies for outage assignment to crews: Nearest outage, most customers affected, most customers affected within a radius of the nearest outage (referred to as nearest within radius), and fastest repair time within radius of the nearest outage (referred to as fastest within radius). The same methods could not be applied directly to Sandy2112 scenarios because there are no data on predicted number of customers affected. The number of customers affected per outage depends on the infrastructure and the population density. The same infrastructure could have more customers affected in a city location compared to a rural location. Therefore, in this study, we are limited to running strategies that only include location or repair time. The use of different search strategies is what separates the ABM from other linear models. The ABM allows input of spatial data and models individual crew behavior, which adds complexity that linear models cannot capture. The Sandy2012 calibration showed that the shape of the restoration curve between strategies was different, but the overall estimated restoration time was not significantly different, as shown in
Figure 3. To account for all possible restoration curve scenarios, Sandy2112 simulations will be run with three search strategies: Nearest outage, fastest repair time, and fastest-within-radius. Search strategies can also be a way to optimize the restoration, which is another novel approach in modeling storm restoration. The restoration curves in
Figure 3 highlight how different strategies reduce the customers without power faster than others. In this scenario, although the final restoration time for all four strategies are close, the number of customers without power varies significantly at different time steps. However, in practice, safety is a limitation of this. Especially in the beginning of the storm, there are some outages that must be addressed before others due to safety concerns. Modeling different search strategies allows emergency managers to investigate different approaches to storm restoration. The ABM allows managers to see the direct impact their decisions can have on both the restoration curve and the overall restoration time.
One point to highlight regarding the historic restoration curve (black line in
Figure 3) is the periodic increases in customers remaining without power. Throughout the restoration, locations with power are sometimes shut off in order to make a safe repair. Once the power is restored, there is a greater decrease in customers without power—the repair returned all those who temporarily lost power along with those still waiting for restoration. The restoration curves of the ABM are smooth because the simulation does not account for more outages, whether intentional or unintentional, except for at the start.
Another way to compare the results of the ABM to the historic restoration is to look at the repair rates. Compared to
Table 3, the repair rates for the historic restoration of Sandy2012, the ABM repair rates for replicated Sandy2012 are higher, as seen in
Table 4. Part of this could be the repair times assigned to the outages. Although the location of each outage is known, the repair time of each outage is unknown, but the ABM assigns a repair time in model setup using uniform distribution. Therefore, the ABM may have more outages with faster repair times than in the historic restoration. The nearest outage strategies are consistent with the historic restoration, but the fastest repair time strategies result in higher rates early in restoration and lower rates later in restoration. The daily average across all strategies results in higher rates early and lower rates after day two or three. Additionally, the ABM gives a “best case” scenario. Crews are assumed to have all materials, knowledge, and skills to make a repair. Each outage is assumed to be ready for crews to begin work as soon as they arrive. Therefore, there is no lost time to acquire additional materials, change crew assignments based on fault type, or wait for a tree crew to clear the area. These three factors would also reduce the repair rate by introducing less efficiency. Although having more differentiation between crew types, materials on hand, and road clearing would make the model more realistic, it would also make the model much more complex. ABMs are designed to be a bottom-up approach [
27]. The developer continues to add complexity without over constraining the model. The goal of an ABM is to keep it as simple as possible to investigate the impacts individual parameters can have. Results from [
25] showed that the ABM was able to recreate the validation storms without the added levels of agent-type, and that simplification was carried over to this model version.
This study utilizes the outage predictions from Wanik et al.’s 2018 study. Wanik used six numerical weather simulations from Weather Research and Forecasting model [
11] based on different convective parameterization schemes (Goddard, Morris, NOTCFLX, WDM6, CNTRL, and ENS) and three machine-learning algorithms (random forest, gradient boosted trees, and Bayesian additive regression trees) to predict the number of outages per 2 km grid across the state of Connecticut for one utility company. Each town is comprised of multiple 2-km grid cells, so outages were aggregated for each town and the total outages are randomly placed in each town. Two sets of input variables were used to develop the Sandy2112 scenarios: One with full weather variables including wind and precipitation and a second using only the wind variables. The precipitation increases were substantial in the Sandy2112 predictions. In case the precipitation variable was overwhelming the machine learning prediction due to the extreme values, the outage prediction models were run again without the precipitation variables. Because there were differences in the number of outages predicted, both sets were used as inputs for the ABM to represent multiple extreme event outage scenarios. It is unknown which, if any, Sandy2112 scenario would happen. Using the range of predictions gives emergency managers a range of scenarios to be prepared for because each scenario would require different resources. In 2012, there were 16,460 outages for this utility company. Although the OPM results include the CNTRL weather simulations, CNTRL has been excluded as an input for the ABM, reducing the six weather simulations down to five. With five WRF simulation models, three machine-learning algorithms, and two sets of input variables, 30 different Sandy2112 scenarios were used, exhibiting total outages ranging from 13,372 to 34,630. Using three different crew search strategies, 90 different ABM scenarios were run. Using the ABM with 90 different parameter combinations is similar to experiencing 90 different storms. Each simulation will have different results, all which can be used to build knowledge on the impacts storm size and crew allocation have on the restoration time. The direct input of outage prediction model results is a novel approach that allows emergency managers to test different decisions and investigate the impacts those decisions have on the restoration.
Wanik et al. [
11] notes that each Sandy2112 scenario resulted in different storm paths and different magnitudes of weather variables. The different storm tracks are accounted for in the ABM by the number of outages predicted per town. The outage prediction model works at a more granular scale (2-km grids), but it is assumed that the most directly hit areas will experience more outages. Although the ABM assigns outages on a town level and crews on an area work center level, the final output is the statewide ETR. One key finding from Wanik et al. [
11] was that most Sandy2112 scenarios showed an increase in the number of outages, but there were four storm/machine learning model combinations that had a decrease in the number of outages predicted: (a) Goddard scheme with boosted trees and full weather variables, (b) Goddard scheme with boosted trees and wind variables only, (c) Goddard scheme with Bayesian additive regression trees and wind variables only, and (d) Morris scheme with boosted additive regression trees and full weather variables. These four scenarios with reduced outages show that changes in weather in the future may change the path of the storm or the storm intensity, resulting in less outages in the study area. Wanik found that random forest had the highest change in outages, followed by Bayesian additive regression trees and then by boosted trees. It is expected that the models with the most outages will have the longest ETRs and the least outages will have the shortest ETRs. Again, the use of the 30 different scenarios is intended to develop a range of estimated time to restoration based on the range of predicted outages from the different models used.
2.2. Results
In this part of the study, we explore how the changes in the number of outages will change the overall restoration time for the storm. The resources were held constant across the ABM simulations—the number of crews working does not increase. As expected, the estimated time to restoration increased with an increase in the number of outages and as crew counts were held constant. The only variables changing in each run is the number and location of outages. The storm track changes are represented by the changes in the location of the outages. Outages are summed per town and randomly located on roadways within that town. Crews are set to specific area work centers and the number of crews working corresponds to the historic restoration of Sandy2012, shown in
Table 1. Crew counts are forward filled (continuing indefinitely) from the last value per AWC in cases where ABM simulations extend beyond the 11-day crew counts from Sandy2012.
Panels A, B, and C in
Figure 4 show the variability of predicted outages by the different WRF convective parameterizations, machine-learning algorithms, and weather inputs used in outage prediction modeling, and the red horizontal line shows the Sandy2012 outages. Panels D, E, and F show the estimated ETR for each number of outages, grouped by the same WRF models, machine learning algorithms and weather parameters as panels A, B, and C. The boxplots for number of outages and the corresponding predicted ETR are similarly shaped and show the same pattern. Because of the corresponding shapes between the two boxplots, we hypothesize that the number of outages and ETR are linearly related.
Table A1 in
Appendix A summarizes the number of predicted outages and ETR for each Sandy2112 scenario.
To test our hypothesis of linearity, each number of predicted outages corresponds to one of the 30 Sandy2112 scenarios (summarized in
Table A1 in
Appendix A). Each scenario has three different predicted ETRs, one for each search strategy: Nearest outage, fastest repair time, and fastest repair time within a radius of the nearest outage.
Figure 5 highlights how the number of outages and the chosen search strategy affects the final ETR. The fastest repair time tends to have a longer ETR than nearest outage and fastest within radius. The Kendall Rank correlation coefficient for nearest outage is 0.9448, fastest repair time is 0.9206, and fastest within radius is 0.9700, and the regression equation for each strategy is shown below the legend of
Figure 5. Each of the Kendall’s correlation coefficients represent a high correlation and linear trend between the number of outages and the predicted restoration time. Kendall Rank correlation was used because the fastest within radius strategy indicated heteroscedasticity, requiring use of a nonparametric test. Additionally, the small sample size was another reason to use the Kendall’s Rank correlation. The black dot, which represents the Sandy2012 outages and restoration time, lies below the ABM ETR estimates indicating that the ABM slightly overestimates the predicted ETR.
Previous studies from Hines et al. [
9,
28] and Carreras et al. [
29] found insignificant correlation between blackout size and duration. However, previous studies are observations of different storms that have many different variables, including both the number of outages and the number of crews working. Using the ABM allows us to take away one of those dimensions, the number of crews, to find a simplification and now a linear relationship in the number of outages (representing size of the storm) and restoration duration. Our assumption of equal crew resources per scenario reduces one aspect of the variability seen in practice as previous studies compared only blackout durations and size without considering crew counts for each restoration. This linear relationship shows that the number of crews must change as the number of outages changes to prevent a change in the ETR. The simplification of the ABM allows us to study the impacts of one parameter before adding variability of multiple parameters.
In addition to comparing the final ETR, the restoration curve for each strategy can be investigated.
Figure 3 showed the curve in terms of thousands of customers remaining without power for Sandy2012 whereas
Figure 6 shows the number of outages remaining for the WDM6 WRF simulation and boosted tree machine-learning algorithm with precipitation and wind input variables for Sandy2112. The ABM does not include customer predictions, therefore the restoration curve in terms of customers effected cannot be compared.
Figure 6 shows the difference in the search strategies and the average of all three throughout restoration. Results prove our third hypothesis and shows differences among the search strategies. Specifically, crews going to the fastest repair time outage within a radius of the nearest outage (fastest-within-radius) and the nearest outage strategies consistently have the shortest ETRs. Crews going to the fastest repair time resulted in the longest overall ETR. By using the fastest repair time strategy, crews may spend more time travelling, but they also save the long duration outages until the end. When the repair time is not included in the prioritization, such as in the nearest outage strategy, there can be more overlap in the order. Some crews will have quick repairs and others will have longer. The differences in those repair times will prevent a step-like curve from forming, which can be seen in the fastest repair time strategy in
Figure 6. The fastest repair time search strategy tends to result in a long tail at the end of the restoration where there are only a few outages remaining, but those few outages have long repair times. The fastest within radius strategy has the shortest ETR and the steepest slope. The fastest strategy starts close to the fastest within radius strategy but has a longer tail at the end, resulting in a longer overall ETR. The nearest outage strategy reduces the number of outages the slowest at the beginning of a storm but begins to drop off about halfway through restoration and finishes with a restoration time close to the fastest within radius strategy.
Figure 6 shows the restoration curve for the Sandy2112 scenario with WDM6 WRF simulation and boosted tree machine learning algorithm with wind and precipitation input variables, but the overall shape of the three search strategies is consistent across all Sandy2112 scenarios. In practice, the restoration process does not have clearly defined search strategies like what was used in the ABM. The definition of search strategies in the ABM can be one way to study optimization of the restoration process and is what separates the ABM from other stochastic approaches. If there are significant differences between strategies, emergency managers can target that strategy once outages are made safe and priority locations have been restored.
Figure 7 compares the range of ETRs for each search strategy, along with the average of the three. A Shapiro–Wilkes test for each strategy determined the range of ETRs to be normally distributed, but due to the small sample size, a nonparametric test will be used. A Kruskal–Wallis test resulted in a
p value < 0.001, indicating differences between the medians of the search strategies and historic value (169 h). To determine which groups were different, a pairwise Wilcoxon Signed Rank test was used.
Table 5 presents the results. The small
p values (
p < 0.05) led to rejecting the null hypothesis and concluding that the medians of each group is different. In this experiment, the number of crews were held constant for all strategies and all storms. The significant difference between strategies exemplifies the value of using an ABM over a linear model and proves our third hypothesis correct. However, restoration in practice does not strictly follow one strategy vs. another. There is a mix of strategies along with several safety protocols in place. The use of the strategies in the ABM can guide emergency managers in the range of time to expect for restoration. For example, if crews were to follow more closely the fastest within radius strategy, the ETR would be the shortest. If crews went by the fastest outage strategy, the ETR would be the longest. Furthermore, if managers were to take an average of all three strategy ETRs, the modeled time would be somewhere in the middle. Beyond using the ABM as a predictive tool, the ABM can be utilized after a storm to test how the restoration could have been impacted by using different strategies. Implementing the ABM in this way allows for the use of the actual outage locations and emergency managers can determine if other decisions could have had better restoration results. Again, this use of the ABM can allow emergency managers to adapt to climate change as storm intensities and patterns change.