CAREER: Learning-Assisted Optimal Power Flow with Confidence

NSF CAREER项目旨在通过将数据驱动技术集成到电网运营中来彻底改变电网运营方式，以便在更快的时间尺度上可靠地运营资产。随着越来越多的可再生能源被引入电网，电网运营变得越来越复杂，电网组件（如发电厂和储能）必须在更快的时间尺度上运行，以平衡功率波动并保持电网可靠运行。该项目通过使用历史数据和机器学习加快解决时间来改变网格运营商解决的传统优化问题。该项目的智力优势集中在使用混合方法将基于学习的解决方案嵌入到现有的网格操作中，将传统的优化与机器学习相结合。电网运营商不必依赖纯粹的数据驱动解决方案，他们可以更有信心地相信，电网可以在高度可再生能源的未来继续可靠地运行。该项目更广泛的影响包括一个设有奖学金的“学习能源”计划，以及公开发布的该研究领域新兴算法的数据集和基准。与通常用于解决收敛速度和成功的这些困难问题的传统近似相反，该项目提议利用数据来保留更复杂的关系，同时保持极快的计算速度。多个“学习辅助”算法将被开发，以提高电力系统运行的各个方面，包括随机，分布式，交流最佳潮流。特别是神经网络可以很好地近似复杂的函数，并在毫秒内执行推理，但单独使用，是黑盒解决方案，可能无法让网格操作员有信心在真实的系统操作中使用它们。在这里，采用联合收割机优化和机器学习的混合方法来确定非凸求解器的更好的热起点，帮助分布式优化技术的收敛时间，提高难以求解的最优潮流场景中的收敛成功率，该奖项反映了NSF的法定使命，并通过使用基金会的学术价值和更广泛的影响审查标准。

This NSF CAREER project aims to revolutionize the way electric power grids operate by integrating data-driven techniques into grid operations in order to reliably operate assets on faster time-scales. As more renewable energy is introduced into the grid and as grid operations grow increasingly complex, grid components such as power plants and energy storage must be operated on faster time-scales to balance fluctuations in power and keep the grid operating reliably. This project transforms the traditional optimization problems that grid operators solve by expediting their solution time using historical data and machine learning. The intellectual merits of this project focus on embedding the learning-based solutions into existing grid operations using a hybrid approach, by combining conventional optimization with machine learning. Rather than relying on purely data-driven solutions, grid operators can have increased confidence that the grid can continue to operate reliably under highly renewable energy futures. The broader impacts of the project include a "learning-for-energy" program with scholarship prizes, and publicly released datasets and benchmarks for algorithms emerging in this area of research.In contrast to traditional approximations of these difficult problems that are typically used to address convergence speed and success, this project proposes to leverage data to preserve more complex relationships while maintaining extremely fast computational speeds. Multiple "learning-assisted" algorithms will be developed to enhance various facets of power system operations including stochastic, distributed, and AC optimal power flow. Neural networks in particular can approximate complex functions well and perform inference in milliseconds, but alone, are black-box solutions that may not give grid operators the confidence to use them in real system operation. Here, hybrid approaches that combine optimization and machine learning are taken to determine better warm-starting points for nonconvex solvers, assist the convergence time of distributed optimization techniques, improve convergence success in difficult-to-solve optimal power flow scenarios, and reduce problem complexity in stochastic grid optimization.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.