Scientists from Princeton University have developed a highly accurate model for predicting power outages during tropical cyclones. The model can assess the impacts of adverse weather on specific areas of up to 1 square kilometer, making it much more accurate than models that only forecast at the county level. The system determines where outages are most likely to occur and how many customers could lose power, which will ultimately allow energy companies to dispatch repair crews and other resources to the most vulnerable areas in advance.
For the new model, the researchers combined outage data from ten energy companies in nine U.S. states serving more than 14 million customers, as well as information on weather, distribution grids, terrain, vegetation and population. Outage data were collected from publicly available online maps of energy companies approximately every ten minutes. This made it possible to obtain the coordinates of individual outages, as well as to determine the time of their occurrence and restoration of power. If detailed grid maps were unavailable, the length of distribution lines was estimated based on road locations, population density and urban development.
Unlike models that primarily look for statistical relationships in large data sets, the new system also took into account the physical structure of the power grids. For each area, it estimated the total length of distribution lines and the proportion of cables buried underground, after which it compared these data with the speed and duration of strong winds, precipitation, storm surges, vegetation and other factors. On this basis, it calculated the probability of damage to overhead lines, the number of possible outages and the number of customers that could lose power.
The model was tested on tropical cyclones from 2020, which had not been used in its training. When estimating the total number of customers without power, the goodness-of-fit indicator reached 0.824, with a maximum value of 1. As a point of comparison, a previously published model for the continental United States had a goodness-of-fit indicator of 0.46.
Wind turned out to be the chief factor in outages. Removing the associated metrics virtually eliminates the accuracy of the calculations. Precipitation, on the other hand, had essentially no impact on accuracy. Accounting for storm surges proved significantly more important: without it, forecast quality declined. This shows that, when assessing the impact of tropical cyclones, it is important to consider not only the impact of wind on power lines and trees, but also the flooding of coastal infrastructure.
Using the new system, the scientists also built a detailed picture of outages for 53 tropical cyclones since 2014. The analysis revealed major differences between areas with differing levels of social vulnerability. With wind gusts of 20–30 m/s, the probability of at least one outage stood at 13.7% in the least vulnerable areas and 19% in the most vulnerable areas. This means that, with comparable wind loads, the latter experienced outages approximately 39% more often. However, the magnitude of an outage after its occurrence varied significantly less between these groups.
One reason for this could be the design of the power grids. In more socially vulnerable areas, the researchers found a lower proportion of underground lines and a greater extent of overhead infrastructure exposed to wind and falling trees. The difference is especially noticeable outside of cities: among the modeled non-urban areas, the share of areas where at least 30% of distribution lines were buried underground was 47.9% for the least vulnerable communities and only 6.3% for the most vulnerable. In cities, the corresponding figures were 55.7% and 28.7%.
The developers suggest that their model could be used both to prepare energy companies for approaching hurricanes and to improve the long-term resilience of power grids. It could be helpful in identifying areas where infrastructure upgrades will have the greatest impact and modeling future risks from climate change. Since the authors have made the source code, model parameters and processed data publicly available, this development could be used for further research.
