Massive power failures are induced frequently by natural disasters. A fundamental challenge is how recovery can be resilient to the increasing severity of disruptions in a changing climate. We conduct a large-scale study on recovery from 169 failure events at two operational distribution grids in the states of New York and Massachusetts. Guided by unsupervised learning from non-stationary data, our analysis finds that under the widely adopted prioritization policy favoring large failures, recovery exhibits a scaling property where a majority ($\sim$90%) of customers recovers in a small fraction ($\sim$10%) of total downtime. However, recovery degrades with the severity of disruptions: large failures that cannot recover rapidly increase by $\sim$30% from the moderate to extreme events. Prolonged small failures dominate entire recovery processes. Further, our analysis demonstrates the promise of mitigating the degradation by enhancing recovery of a small fraction of large failures through distributed generation and storage.

自然灾害频繁引发大规模停电。一个根本性的挑战是复苏如何能够抵御气候变化中日益严重的破坏。我们对纽约州和马萨诸塞州两个运营配电网的 169 次故障事件的恢复进行了大规模研究。在非平稳数据的无监督学习的指导下，我们的分析发现，在广泛采用的有利于大故障的优先策略下，恢复表现出一种扩展特性，其中大多数（$～$90%) 的客户在一小部分 ($～$10%) 的总停机时间。但是，恢复会随着中断的严重程度而降低：无法快速恢复的大型故障会增加$～$30% 来自中等至极端事件。长时间的小故障在整个恢复过程中占主导地位。此外，我们的分析证明了通过分布式发电和存储增强一小部分大故障的恢复来减轻退化的前景。