Machine Learning for Communication-Cognizant Smart Inverter Control

智能逆变器将太阳能电池板、储能单元和电动汽车接入主电网，是一种有效的电网控制机制，可以应对住宅电网中太阳能发电引起的电压波动。然而，实时为数百台逆变器选择功率注入是一项艰巨的任务。为了在网络要求优化解决方案和低于标准的局部控制之间找到平衡点，本文提出的研究开发了新的算法方案，用于基于通过机器学习（ML）训练的控制规则或策略来协调逆变器。该计划促进了对住宅电网的通信限制分散控制的理解，并发现了智能电表数据的新用途，可立即适用于配电系统并与电力公用事业实践相关。对行业的预期好处是尖端的物理感知ML解决方案，先进的资产管理，以及逆变器，传感器和电网通信的增值。通过实现分布式太阳能发电、电动汽车和需求响应计划等智能电网技术的可靠集成，增强电网的可控性将使社会受益。处理逆变器的电压漂移延长了稳压器和电容器的使用寿命，并推迟了昂贵的电网升级。拟议的研究活动与教育和外联目标相结合。该项目通过可再生能源的实践学习活动向大学预科女生提供服务，并继续通过访问住宅社区和研讨会向本科生提供服务。研究结果将被整合到配电电网和能源系统机器学习的研究生课程中，从而教育即将到来的电力工程劳动力及时使用机器学习工具。该项目旨在将机器学习（ML）模型集成到最优潮流（OPF）公式中，以开发配电电网中数据自适应和通信限制的智能逆变器控制方案。这是通过完成以下任务来实现的：i)设计算法解决方案，以确保ML模块符合逆变器规格和馈线约束；ii)在基于模型和无模型设置下，通过近似和精确的交流功率流方程捕获ML模块之间的物理耦合；iii)针对太阳能发电和负荷的不确定性，提出了基于随机和机会约束ml的OPF方案；Iv)设计集中训练的学习架构，并在紧张的通信预算下以分布式方式实施，以利用大的历史数据和小的实时数据馈送；v)确保ML模块由状态驱动闭环时的动态稳定性和安全性。结果将是一套全面的基于ml的逆变器优化解决方案，使用基准馈线的实际数据进行验证。提出的研究是在电力系统、机器学习、鲁棒控制和随机优化的关系上设计数据驱动的非线性逆变器控制策略的先驱。通过随机约束OPF学习逆变器控制规则的原始和潜在的变革思想是这个项目的贯穿主题。当OPF和局部控制方案无法适应时变的网格条件时，所提出的ML-OPF解决方案利用创造性设计的神经网络拓扑，以及基于核和强化学习的进步，在训练-大/操作-小数据范式下有效运行。建立通信感知的逆变器控制策略将提高我们在通信预算紧张的物理或工程系统上的约束学习知识。研究任务探索不同的机器学习工具；考虑基于模型和无模型的设置；处理近似和精确网格模型；并考虑公用事业和逆变器之间不同的通信能力。该项目独特地结合了配电网、统计学习、鲁棒控制和强化学习方面的专业知识，超越了通信受限的逆变器控制，解决了带宽受限的资源分配问题和网络物理系统控制。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。

Smart inverters interfacing solar panels, energy storage units, and electric vehicles to the main electric grid are an effective grid control mechanism to cope with the voltage fluctuations induced by solar generation in residential grids. However, selecting the power injections for hundreds of inverters in real time is a formidable task. To hit the sweet spot between cyber demanding optimization solutions and subpar local control, the proposed research develops novel algorithmic schemes for coordinating inverters based on control rules or policies trained through machine learning (ML). This program advances understanding of communication-limited decentralized control of residential electric grids and discovers novel uses of smart meter data, with immediate applicability to distribution systems and relevance to electric power utility practice. The expected benefits to industry are cutting-edge physics-aware ML solutions, advanced asset management, and value added for inverters, sensors, and grid communications. Enhancing grid controllability benefits society by enabling reliable integration of smart grid technologies, such as distributed solar generation, electric vehicles, and demand-response programs. Handling voltage excursions by inverters extends the lifetime of voltage regulators and capacitors, and postpones costly grid upgrades. The proposed research activities are integrated with educational and outreach objectives. The program reaches pre-college female students through hands-on learning activities on renewable energy sources and continues the PIs' outreach to undergraduate students through visits to residential communities and seminars. The findings will be integrated into the curricula of (under)-graduate courses on distribution grids and ML for energy systems, thus educating the upcoming power-engineering workforce on timely ML tools. This project aspires to integrate machine learning (ML) models into optimal power flow (OPF) formulations to develop data-adaptive and communication-limited smart inverter control schemes in power distribution grids. This is achieved by accomplishing the ensuing tasks: i) devise algorithmic solutions to ensure ML modules comply with inverter specifications and feeder constraints; ii) capture the physical coupling across ML modules through approximate and exact AC power flow equations under both model-based and model-free setups; iii) put forth stochastic and chance-constrained ML-based OPF schemes to deal with the uncertainty of solar generation and loads; iv) design learning architectures that are centrally trained and implemented in a distributed manner on a tight communication budget to exploit big historical data but small real-time data feeds; and v) ensure dynamic stability and safety when ML modules are driven by states to close the loop. The outcome will be a comprehensive suite of ML-based inverter optimization solutions validated using real-world data on benchmark feeders. The proposed research pioneers the design of data-driven nonlinear inverter control policies at the nexus of power systems, ML, robust control, and stochastic optimization. The original and potentially transformative idea of learning inverter control rules through a stochastic constrained OPF is the permeating theme of this project. When OPF and local control schemes fail to adjust to time-varying grid conditions, the proposed ML-OPF solutions exploit creatively designed neural network topologies, as well as advances in kernel-based and reinforcement learning to operate effectively under the train-big/operate-small data paradigm. Building communication-aware inverter control policies will advance our knowledge on constrained learning over a physical or engineered system under a tight communication budget. Research tasks explore different ML tools; account for model-based and model-free setups; handle approximate and exact grid models; and consider varying communication capacities between the utility and inverters. Uniquely combining expertise on distribution grids, statistical learning, robust control, and reinforcement learning, this project goes beyond communication-limited inverter control to bandwidth-limited resource allocation problems and cyber-physical systems control.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.