CAREER: Deciphering Large-Scale Real Outage Data for Cascading Failure Analysis, Prevention, and Intervention

级联故障是复杂工程系统中的常见现象，如电网、天然气系统、交通网络、互联网和相互依赖的关键基础设施系统。例如，2003年的美国-加拿大大停电和2012年的印度大停电等以往的重大级联停电造成了许多部件故障，重大的经济损失和严重的社会影响。因此，大幅提高复杂系统的弹性，保证系统的正常运行，对于最大限度地减少服务中断，保持经济增长和繁荣至关重要。然而，现有的级联失效研究严重依赖于仿真模型，这些模型要么过于笼统，无法提供可实施的预防/干预策略，要么难以进行基准测试或验证，无法得出可靠的结论。在这个CAREER项目中，为了更好地分析、预防和干预级联故障，将通过开发一个全新的真实中断数据驱动的研究框架来解决现有方法的重大局限性。本项目所开展的级联故障分析、预防和干预研究，将有助于大大降低灾难性停电的风险，增强电力系统的恢复能力，为美国及世界各国带来巨大的经济效益和社会效益。提出了几项教育活动，包括在本科和研究生阶段开发课程，让本科生参与研究，以及让西班牙裔和女性学生参与项目。对于K-12和社区教育，将建立一个基于级联停电场景的示范。学术团体将参与网络研讨会、研讨会和会议小组会议。CAREER项目的目标是为级联故障和电力系统弹性的研究开辟一个新的数据驱动的研究方向，为数据驱动的级联故障分析、预防和干预提供系统的、变革性的理论基础和算法技术，解决现有方法的固有局限性，刷新对现实级联故障的理解，并提供分析、预防、最后干预级联失败。网络科学、统计推断、数据科学和深度学习将与电力系统领域知识无缝集成，以获得解决极具挑战性问题的独特解决方案。具体而言，以下三个相互关联的项目将重点突出，以显著推进研究议程：1)通过有效和准确地估计故障交互，将公用事业中断数据转换为信息，为数据驱动的级联故障方法奠定坚实的基础，解决级联阶段传播模式演变，级联之间的高异质性和数据稀缺性的挑战。将开发深度学习方法来揭示交互的结构特征并恢复缺失的组件交互。2)利用实际电力中断数据，基于故障相互作用的估计，研究级联故障传播的复杂性和普遍性，将信息转化为知识，揭示级联行为的时间演化、空间传播模式以及时间演化和空间传播的临界性。3)通过将知识转化为可操作的智慧，将推进极具挑战性的级联故障预防和干预研究。级联故障预防将基于使用预期停机次数确定的关键组件来开发。关键部件的结构或操作特征将使用故障交互网络和原始系统拓扑来识别。此外，将制定一个新的级联故障干预问题，并通过一种深度、安全的强化学习方法来解决，该方法利用了从实际停电数据和电力系统领域知识中学习到的信息/知识。通过解释学习到的最优策略的高级抽象，将提供有关级联故障干预的有用见解。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Cascading failure is a common phenomenon in complex engineered systems, such as electric power grids, natural gas systems, transportation networks, Internet, and interdependent critical infrastructure systems. For example, previous major cascading blackouts such as the U.S.-Canadian blackout in 2003 and the Indian blackout in 2012 have caused many component failures, significant economic losses, and severe social impacts. Therefore, greatly enhancing the complex system resilience and keeping the lights on are thus very critical for minimizing the service disruptions and maintaining the economic growth and prosperity. However, the existing cascading failure study heavily relies on simulation models, which are either too general to be able to provide implementable prevention/intervention strategies or too difficult to benchmark or validate for drawing reliable conclusions. In this CAREER project, the significant limitations of the existing approaches will be addressed by developing a fresh-new real outage data driven research framework in order to better analyze, prevent, and intervene cascading failures. The research on cascading failure analysis, prevention, and intervention to be conducted in this project will help greatly reduce the risk of catastrophic blackouts and enhance the resilience of power systems, bringing tremendous economic and social benefits to U.S. and other countries around the world. Several educational activities including curriculum development at undergraduate and graduate level, engaging undergraduate students in research, and involving Hispanic and women students in the project are proposed. For K-12 and community education, a demonstration based on a cascading blackout scenario will be built. Academic community will be engaged with webinars, seminars, and panel sessions at conferences.The goal of this CAREER project is to initiate a new data-driven research direction for studying cascading failure and power system resilience and develop systematic, transformative theoretical foundations and algorithmic techniques for data-driven cascading failure analysis, prevention, and intervention, addressing the inherent limitations of existing approaches, refreshing the understanding of real-world cascading failure, and providing tools to analyze, prevent, and finally intervene cascading failures. Network science, statistical inference, data science, and deep learning will be seamlessly integrated with power system domain knowledge in order to obtain a unique solution to the very challenging problem. Specifically, the following three inter-related projects will be highlighted to significantly advance the research agenda: 1) A solid foundation of the data-driven cascading failure approach will be built through an efficient and accurate estimation of failure interactions to convert utility outage data to information, addressing the challenges of propagation pattern evolution over cascading stages, high heterogeneity among cascades, and data scarcity. Deep learning approaches will be developed to reveal the structural features of the interactions and recover the missing component interactions. 2) Utilizing real utility outage data, the complex and universal temporal-spatial properties of cascading failure propagation will be investigated based on the estimated failure interactions to convert information to knowledge, revealing the temporal evolution of cascading behaviors, spatial propagation patterns, and criticality in both temporal evolution and spatial propagation. 3) The very challenging cascading failure prevention and intervention study will be advanced by converting knowledge to actionable wisdom. Cascading failure prevention is to be developed based on the identified critical components using the expected number of outages. Structural or operational signatures of critical components will be identified using both the failure interaction network and the original system topology. Further, a novel cascading failure intervention problem will be formulated and solved by a deep, safe reinforcement learning approach that utilizes both the learned information/knowledge from real outage data and the power system domain knowledge. By interpreting the high-level abstraction of the learned optimal policies, useful insights about cascading failure intervention will be provided.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.