Power outages caused by extreme weather events cost the economy of the United States billions of dollars every year and endanger the lives of the people affected by them. These types of events could be better managed if accurate predictions of storm impacts were available. While empirical power outage prediction models have been in development for many years, accurate operational predictions of the most extreme and impactful weather-related outage events have proven difficult to achieve for several reasons. In this paper, we describe a data intensive modeling approach specifically designed for forecasting the impacts of extreme weather events on power distribution grids. To that end, methods for developing datasets that include a large number of example storms and predictors are described. In addition, we test several methods of managing the extreme value distribution of the target variable via statistical transformation and balancing of the dataset. The best performing outage prediction model developed here is capable of predicting storm impacts across four orders of magnitude with R${}^2$ and Nash–Sutcliffe Efficiency scores of 0.82. Also, by investigating the model’s sensitivities and predictions for the highest impact events, we find that there is significant diversity in the meteorological factors that drive the predictions for the most severe events, suggesting that the weather hazards are more complex than they often treated in empirical analyses of their impacts. The accuracy of the outage model, together with the importance of various meteorological variables that contribute to that accuracy, validate the described methodology and suggest that future empirical analysis of the impacts of extreme weather should include multifaceted descriptions of the hazard to better represent the complex factors which contribute to the most impactful events.

极端天气事件造成的停电每年给美国经济造成数十亿美元的损失，并危及受其影响的人们的生命。如果能够准确预测风暴影响，则可以更好地管理这些类型的事件。虽然经验性停电预测模型已经开发多年，但由于多种原因，对最极端和最有影响的天气相关停电事件的准确运行预测已被证明难以实现。在本文中，我们描述了一种专门用于预测极端天气事件对配电网影响的数据密集型建模方法。为此，描述了用于开发包含大量示例风暴和预测变量的数据集的方法。此外，我们测试了几种通过数据集的统计转换和平衡来管理目标变量极值分布的方法。这里开发的性能最佳的中断预测模型能够预测风暴影响的四个数量级，R${}^2$Nash-Sutcliffe 效率得分为 0.82。此外，通过调查模型对最大影响事件的敏感性和预测，我们发现推动对最严重事件的预测的气象因素存在显着差异，这表明天气灾害比通常在经验中处理的更复杂。分析它们的影响。中断模型的准确性，以及有助于该准确性的各种气象变量的重要性，验证了所描述的方法，并建议未来对极端天气影响的经验分析应包括对危害的多方面描述，以更好地代表复杂因素这有助于最有影响力的事件。