Collaborative Research: Lifelong Human-in-the-Loop Multiagent Learning for Decentralized Restoration of Distribution Systems (LifeGuard)

本研究开发了一种新颖的多智能体系统，为配电服务恢复提供实时解决方案，提高能源供应弹性。提出的多智能体框架将：(1)利用从图结构电力系统动态数据中获得的时空信息提供准确的故障识别和定位；2)提供一个分布式多智能体学习框架，考虑生成的解决方案的可扩展性和时间效率，寻找快速恢复系统的最优恢复策略；3)通过有效地将知识从源问题转移到电网拓扑和特性发生变化的目标问题，获得具有足够灵活性的终身电力恢复方案；4)为人类专家提供一个可解释的知识库，以评估修复方案并根据他们的先验知识和专业知识对其进行修改。学术和教育界将从这个项目产生的知识的迅速传播中受益。该研究计划鼓励研究生和本科生之间的创造性合作，寻找新的功能解决方案，以解决当前配电网运营中的挑战。该项目将为研究生和本科生制定新的课程，促进跨学科研究，并开展K-12外展活动。本研究的目的是开发一个分散的时空人工智能框架，用于大规模配电网络中分布式故障检测和识别，以及系统测量的高维性、稀疏性和部分可观测性。该项目开发了一个图胶囊网络来识别配电系统的时空动态模式，并检测故障的类型和位置。此外，我们设计了一种新颖的完全分散的多智能体系统，具有行为者-批评强化学习，以解决具有高维系统状态和动作的大规模恢复问题。此外，我们将多智能体系统的知识转移作为机器学习中的一个开放问题，并开发了一个能够适应电力系统拓扑和特征变化的终身恢复框架。该项目还将注意力模型整合到深度强化学习中，为拟议的恢复框架开发一个可解释的知识库，该知识库可用于使用人类专业知识进行恢复知识验证和修改。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。

This research develops a novel multiagent system to provide a real-time solution for distribution service restoration that enhances energy supply resilience. The proposed multiagent framework will 1) leverage the spatiotemporal information obtained from the graph-structured power system dynamic data to provide accurate fault identification and location; 2) provide a distributed multiagent learning framework to find optimal restoration policies for rapid system recovery considering the scalability and time efficiency of the generated solutions; 3) obtain lifelong power restoration schemes that are flexible enough to adapt to new restoration problems by efficiently transferring their knowledge from a source problem to a target problem where the topology and characteristics of the power network are changed; and 4) provide an interpretable knowledge base for the human experts to evaluate the restoration scheme and modify it based on their prior knowledge and expertise. The academic and educational communities will benefit from the rapid dissemination of the generated knowledge from this project. The research plan encourages inventive collaboration among graduate and undergraduate students to find novel functional solutions to address current challenges in the distribution network operation. The project will develop a new curriculum for graduate and undergraduate students, promote interdisciplinary research, and develop K-12 outreach activities. The objective of this research is to develop a decentralized spatiotemporal artificial intelligence framework for distributed fault detection and identification, and power system restoration in large-scale distribution power networks considering the high dimensionality, sparsity, and partial observability of the system measurements. The project develops a graph capsule network to recognize spatiotemporal dynamic patterns of distribution systems, and detect the type and location of faults. Moreover, we devise a novel fully decentralized multiagent system with actor-critic reinforcement learning to solve large-scale restoration problems with high-dimensional system states and actions. Furthermore, we address knowledge transfer of multiagent systems as an open problem in machine learning and develop a lifelong restoration framework capable of adapting to changes in the topology and characteristics of the power system. The project also incorporates attention models into deep reinforcement learning to develop an interpretable knowledge base for the proposed restoration framework that can be used for restoration knowledge verification and modification using human expertise.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.