CAREER: Faithful, Reducible, and Invertible Learning in Distribution System for Power Flow

随着配电网变得更加智能，新的服务也变得可用，如社区可再生集线器、电动汽车(EV)家庭能源管理和需求响应计划。然而，由于(1)系统不可观测性，(2)低分辨率和不良数据导致的数据质量问题，以及(3)太阳能、风能发电和电动汽车充电等高度间歇性能源，这些新服务对电网可靠性构成了重大挑战。如果不解决这些问题，停电和设备损坏可能会频繁发生，导致设备故障、昂贵的维护成本和不满意的客户。这项事业的目标是为在不可观测条件下建立配电网严格的潮流方程奠定理论基础。拟议研究的智力价值在于为学习理论、机制和体系结构生成新知识。主要进展是(1)设计了基于深度神经网络(DNN)的潮流方程，这些方程不仅在物理上是可简化的，而且在DNN中表示物理的中间层是凸形的；(2)创建了一个物理生成的对抗性网络，以增强在有限数据下重构潮流的稳健性；以及(3)推导了基于物理DNN的潮流方程的求逆解决方案，用于实时潮流分析。更广泛的影响包括对公用事业和公司的物理增强的人工智能框架，由于我们的模型的简单性、忠实性和可逆性，这也展示了一个更具弹性的配电网管理系统。这项拟议的工作将通过弥合配电网中物理和学习之间的差距，大大降低管理表后资源和排放的成本。这种对网格现代化的跨学科分析将在新的机器学习方法上创建一类新的开放源码，支持一个蓬勃发展的学者和行业合作者社区。综合教育部分还将为K12学生和少数族裔创建一个科学计划，让他们参与与电力系统人工智能相关的活动。为了在公用事业工程师中普及人工智能，PI计划部署拟议的平台。PI还将扩大他的网络研讨会系列，以促进学术界和行业之间的交流。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。

As the electric distribution grid becomes smarter, new services have become available, such as community renewable hubs, home energy management with electric vehicles (EV), and demand response programs. However, these new services pose significant challenges to grid reliability due to (1) system unobservability, (2) data quality issues because of low-resolution and bad data, and (3) highly intermittent energy sources like solar and wind generation and EV charging. Without solving these problems, power outages and equipment damage can happen frequently, leading to device malfunctions, expensive maintenance costs, and unsatisfied customers. The goal of this CAREER effort is to develop the theoretical foundation of building a rigorous power flow equation in the distribution grid under unobservability. The intellectual merit of the proposed research lies in the generation of new knowledge for learning theories, mechanisms, and architectures. Key advances are (1) designs of deep neural network (DNN)-based power flow equations that are not only physically reducible but also convex to the intermediate layer in DNN that represents physics, (2) creation of a physical-generative adversarial network to boost the robustness of reconstructing power flow against limited data, and (3) derivation of the solution for inverting the DNN-based power flow equation based on the physical DNN for real-time power flow analysis.The Broader Impacts include a physics-enhanced AI framework to utilities and companies that also demonstrate a more resilient distribution grid management system due to the reducibility, faithfulness, and invertibility of our models. The proposed work will help reduce the cost of managing behind-the-meter resources and emissions greatly by bridging the gap between physics and learning in distribution grids. Such interdisciplinary analysis for grid modernization will create a new class of open-source code on new machine learning methods, supporting a thriving community of academics and industry collaborators. The integrated education component will also create a scientific program for K12 students and minorities to engage in activities related to AI for power systems. To popularize AI among the utility engineers, the PI plans to deploy the proposed platforms. The PI will also expand his webinar series to promote communications between academia and industry.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.