Collaborative Research: SLES: Safety under Distributional Shift in Learning-Enabled Power Systems

该NSF项目旨在使支持学习的安全关键系统的设计发生革命性变化，特别关注电力系统。由于环境和技术的加速变化，这些系统面临着越来越多的挑战。该项目将通过引入反脆弱性的概念，为这类系统的运作带来变革性的变化，从而促进变革，使之成为加强系统的机会。这一创新观点对于在我们快速变化的环境中有效管理分配转变至关重要。这一转变将通过开创主动的、基于内存的反脆弱系统、探索用于协作决策的多代理系统以及应用先进的验证和严格压力测试技术来实现。该项目的学术价值包括一种突破性的方法，以拥抱变化和不确定性。该项目没有将这些因素视为有害因素，而是将它们用作自我改进的催化剂，为以一种弹性和适应性的方式运营安全关键系统奠定了基础。该项目的更广泛影响包括提高电力系统等关键基础设施的复原力和可靠性，以确保不间断地获得重要服务。该项目还寻求成为关于安全决策和公共宣传活动的跨学科对话的中心，以促进STEM社区的科学素养和多样性。在电力系统运行中，安全至关重要，需要遵守描述各种参数(如电压、频率或设备健康状况)动态的严格数学模型。在环境日益复杂和不可预测的驱动下，在分配变化中，维护端到端安全的任务正变得极其复杂。我们的项目通过三个相互关联的研究推进来应对这些挑战。第一个推力的目标是利用元安全学习和离线强化学习等先进技术，创建主动的、反脆弱的系统，以预测和适应变化。第二个推力通过多智能体系统支持系统的抗脆弱性，鼓励探索、合作和分布式控制，以确保即使在重大干扰下也能恢复和安全。第三个重点是验证和压力测试，采用多目标对抗性学习和真实世界案例研究，以更好地处理罕见或意外事件。这些研究成果提供了对系统脆弱性、分布式决策和合作行为的全面理解。在经验分析和数学保证的支持下，建议的方法提供了一种稳健的方法，在不断变化的挑战中确保学习型系统的安全，标志着该领域的重大进步。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This NSF project aims to revolutionize the design of learning-enabled, safety-critical systems, with a special focus on power systems. These systems face increasing challenges due to accelerated environmental and technological changes. The project will bring transformative change to the operation of such systems by introducing the concept of antifragility, which promotes change as an opportunity for system enhancement. This innovative viewpoint is crucial for effectively managing distributional shifts in our rapidly changing environment. This transformation will be achieved by pioneering proactive, memory-based antifragile systems, exploring multi-agent systems for cooperative decision-making, and applying advanced techniques for validation and rigorous stress testing. The intellectual merits of the project include a groundbreaking approach towards embracing change and uncertainty. Rather than perceiving these factors as detriments, the project uses them as catalysts for self-improvement, setting the stage for a resilient and adaptive way to operate safety-critical systems. The broader impacts of the project include enhancing the resilience and reliability of crucial infrastructures such as power systems to ensure uninterrupted access to vital services. The project also seeks to serve as a hub for cross-disciplinary dialogue on safe decision-making and public outreach activities to foster scientific literacy and diversity within the STEM community.In power system operation, safety is crucial and requires adherence to rigorous mathematical models that describe the dynamics of various parameters such as voltage, frequency, or the health of an equipment. The task of preserving end-to-end safety is becoming prohibitively complex amidst distributional shifts, driven by the growing complexity and unpredictability of the environment. Our project addresses these challenges through three interconnected research thrusts. The first thrust targets the creation of proactive, antifragile systems that anticipate and adapt to changes, using advanced techniques such as meta-safe learning and offline reinforcement learning. The second thrust bolsters system antifragility through multi-agent systems, encouraging exploration, cooperation, and distributed control to ensure resilience and safety, even under significant disturbances. The third thrust is devoted to validation and stress testing, employing multi-objective adversarial learning and real-world case studies to better handle rare or unexpected events. These research thrusts provide a comprehensive understanding of system fragility, distributed decision-making, and cooperative behavior. Supported by empirical analysis and mathematical guarantees, the proposed methodologies offer a robust approach to ensuring the safety of learning-enabled systems amidst evolving challenges, marking a significant advancement in the field.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.