ERI: Learning to Operate Distribution Grids with Extreme Penetration of Inverter-based Resources (L2ODG)

随着电网从中央控制系统转向越来越依赖基于逆变器的资源（ibr），如太阳能和风能，电网正在经历一场重大变革。这一变化为电力系统的运行和管理带来了新的挑战和机遇。拟议的研究旨在开发基于机器学习的创新解决方案，以解决将ibr集成到配电网中所带来的实时操作挑战。通过创建一套强大的、可扩展的、安全关键的机器学习工具，该项目将使清洁能源的广泛采用和配电电网的转型成为可能。这项研究更广泛的影响包括通过跨学科研究和课程设计，将电力工程、机器学习和物联网结合起来，培养下一代科学家和工程师。此外，该项目将开发开源数据集和模型，同时通过互动演示在代表性不足的大学预科学生中推广STEM教育。这项研究的成功完成将加速清洁能源的采用，增强电力系统的弹性，并确保公正和公平的能源转型。拟议的项目旨在开发一个基于学习的控制和优化框架，以解决与配电网和基于逆变器的资源（IBRs）集成相关的实时操作挑战。研究目标包括：1)基于不准确系统模型的样本高效混合学习，将深度强化学习（DRL）与简化网格模型和现实世界试验相结合，以实现快速学习和性能提高；2)用于实时网络重构的图强化学习，将DRL与图神经网络（GNN）集成在一起，创建一种新的学习架构，能够适应并推广到任意网络拓扑，解决电压违规、无功分配和系统损耗最小化等问题；3)网格边缘的分布式学习和控制，通过将联邦多智能体学习与深度压缩技术相结合，建立一个可扩展的分布式学习框架，运行在资源受限的边缘设备上。所提出的框架将被设计成鲁棒、自适应、安全和轻量级，在考虑网络约束以确保系统安全的同时，提供网格拓扑的实时优化和控制。该项目将有助于为未来具有大规模ibr和分散测量的配电系统开发可扩展的、基于学习的控制解决方案，从而培养更具弹性和可持续性的电网。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。

The power grid is undergoing a significant transformation as it shifts from centrally controlled systems to increasingly relying on inverter-based resources (IBRs), such as solar and wind power. This change brings new challenges and opportunities for how power systems are operated and managed. The proposed research aims to develop innovative machine learning-based solutions to address real-time operational challenges arising from integrating IBRs in power distribution grids. By creating a suite of robust, scalable, and safety-critical machine learning tools, this project will enable widespread adoption of clean energy sources and transform distribution grids. The broader impacts of this research include training the next generation of scientists and engineers through interdisciplinary research and curriculum design that integrates power engineering, machine learning, and the Internet-of-Things. Furthermore, the project will develop open-source datasets and models while promoting STEM education among underrepresented pre-college students through interactive demonstrations. Successful completion of this research will accelerate the adoption of clean energy resources, enhance power system resilience, and ensure a just and equitable energy transition.The proposed project seeks to develop a learning-based control and optimization framework addressing real-time operational challenges associated with power distribution grids and increasing integration of inverter-based resources (IBRs). The research objectives include: 1) Sample-efficient Hybrid Learning with Inaccurate System Models, combining deep reinforcement learning (DRL) with simplified grid models and real-world trials for rapid learning and performance improvement; 2) Graph Reinforcement Learning for Real-time Network Reconfiguration, integrating DRL with Graph Neural Networks (GNN) to create a novel learning architecture capable of adapting to and generalizing to arbitrary network topologies, addressing issues like voltage violations, reactive power distribution, and system loss minimization; and 3) Distributed Learning and Control over the Grid Edge, establishing a scalable, distributed learning framework that runs on resource-restricted edge devices by combining federated multi-agent learning with deep compression techniques. The proposed framework will be designed to be robust, adaptive, safe, and lightweight, offering real-time optimization and control of grid topology while considering network constraints to ensure system safety. This project will contribute to developing scalable, learning-based control solutions for future distribution systems with massive IBRs and dispersed measurements, fostering a more resilient and sustainable power grid.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.