Collaborative Research: Learning for Faster Computations to Enhance Efficiency and Security of Power System Operations

电网是一个复杂的关键基础设施系统，支撑着美国的所有经济和社会活动。因此，始终保持其高效、可靠和安全的运作至关重要。然而，随着可再生能源和电动汽车的快速增长，以及对网络安全的日益担忧，该系统正在经历前所未有的转型时期。因此，不仅对电网的高效和安全运行提出了更高的要求，而且实现它变得更具挑战性。从计算的角度来看，这个问题尤其严重，因为需要更频繁地解决更复杂的问题。因此，用于解决安全电力系统操作问题的常规方法在快速发展的电网中维持其功效方面面临重大且紧迫的挑战。为了克服这些挑战，该项目将开发新的基于机器学习的方法，以大大加快解决关键和大规模安全电力系统运行问题。值得注意的是，开发的方法将数据驱动的方法与电力系统的物理模型相结合。该项目的影响扩展到所有工程系统中的机器学习算法设计，其中可以结合物理系统模型的知识和算法设计中的传统智慧。开发的算法将大大提高电力系统的效率，可靠性和安全性，在可再生能源的高渗透率的情况下，不需要建造更多的输电线路或获得更高的备用容量，为消费者带来巨大的经济节省。该项目还将通过培训下一代劳动力来掌握机器学习和电力系统的跨学科专业知识，为行业迫切需要的教育需求做出贡献。研究所致力于通过该项目促进研究和教育的多样性，让少数民族和贫困背景的学生参与进来。该项目将开发新的机器学习算法，利用现有的计算工具并与之集成，以大大提高解决具有挑战性的电力系统运行问题的计算效率。我们通过设计算法来实现这一点，这些算法使用数据来取代一些基于人类经验的现有算法。我们使用自下而上的方法，通过仔细制定问题来确定物理系统和机器学习之间的最佳接口。这使我们能够设计出了解问题物理特性的算法，并补充该领域现有的工具。具体来说，我们追求三个研究方向：i）通过将数据驱动组件引入现有算法来解决最佳发电机调度水平; ii）使用强化学习快速识别和量化有问题的突发事件; iii）利用以前的结果找到最安全和最有效的发电机组承诺计划。这些算法可以直接集成到当前的求解器，并有可能提供数量级的加速比现有的方法。因此，该项目提供了a）新的机器学习范例和算法，B）将机器学习方法与基于物理模型的优化算法相结合的创新方法，以及c）潜在的变革性工具，以更快的速度在整体框架内解决关键的电力系统运行问题。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的知识产权进行评估来支持。优点和更广泛的影响审查标准。

The electric grid is a complex critical infrastructure system that underpins all economic and social activities in the US. It is thus of utmost importance to maintain its efficient, reliable and secure operation at all times. The system, however, is undergoing an unprecedented period of transformation with rapid growths in renewable energy and electric vehicles, as well as increasing concerns of cyber security. Consequently, not only there is a higher requirement for efficient and secure operation of the grid, but also achieving it becomes much more challenging. The issue is especially acute from a computational perspective, as problems of much greater complexity need to be solved more frequently. As such, conventional approaches for solving secure power system operation problems face major and pressing challenges in maintaining their efficacy in the rapidly evolving power grids. To overcome these challenges, this project will develop novel machine-learning-based methods to greatly accelerate solving key and large-scale secure power system operation problems. Notably, the developed methods will integrate data-driven methods with the physical models of power systems. The impact of the project extends to machine learning algorithm design in all engineering systems where knowledge from physical system models and conventional wisdom in algorithm design can be incorporated. The developed algorithms will lead to greatly enhanced efficiency, reliability and security of power systems in the presence of high penetration of renewable energy and without the need of building more transmission lines or procuring much higher reserve capacity, resulting in tremendous economic savings for consumers. The project will also contribute to the much-demanded educational needs in the industry by training the next generation workforce to master interdisciplinary expertise of machine learning and power systems. The PIs are committed to promote diversity in research and education through the project by engaging students of minorities and from under-privileged backgrounds. This project will develop new machine learning algorithms, both leveraging and integrated with existing computational tools, to greatly improve the computational efficiency of solving challenging power system operation problems. We accomplish this by designing algorithms that use data to replace some of the existing heuristics based on human experience. We use a bottom-up approach by carefully formulating the problems to determine the best interface between the physical system and machine learning. This allows us to design algorithms that are aware of the physics of the problems and complement existing tools in the field. Specifically, we pursue three research thrusts: i) solving for optimal generator dispatch levels by introducing a data-driven component to the existing algorithms; ii) enabling fast identification and quantification of problematic contingencies using reinforcement learning; and iii) finding the most secure and efficient generation unit commitment schedule utilizing the results from the previous thrusts. These algorithms can be directly integrated into current solvers and have the potential of providing orders of magnitude speedup over existing methods. As such, this project offers a) new machine learning paradigms and algorithms, b) innovative ways of integrating machine learning methods with physical model-based optimization algorithms, and c) potentially transformative tools that solve key power system operation problems in a holistic framework with much faster speeds.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.