EAGER: Real-Time: Decision and Control of Complex Engineered Systems Enabled by Machine Learning and High-performance Computing

近年来，机器学习取得了重大进展。机器学习是一种统计技术，它使计算机能够利用现有数据“学习”。机器学习方法在图像识别、语言翻译、语音处理和其他消费应用中取得了巨大的成功。这引起了全球学术界、工业界和政府的极大兴趣。纯机器学习方法的缺点是它没有使用特定系统的物理属性知识，这可以显着提高这些方法的性能。这个早期概念探索性研究资助(AGER)项目将产生基本的结果和方法，这些结果和方法结合了机器学习技术的优势和系统的物理属性知识，使复杂工程系统的决策和控制成为可能。这项研究将在控制大型风能发电厂的背景下进行。尽管大型风电场的运行条件多变和不确定，但最大限度地提高发电量是一个尚未解决的问题，变革方法和创新的时机已经成熟。随着风电场所有者和运营商不断寻求新的方法来提高年发电量和降低风电成本，该项目的研究可能会通过利用与NSF I-UCRC for Wind Energy Science，Research and Technology(Windstar)的联系而过渡到工业。这个渴望项目的主要想法是利用深度学习和高性能计算模拟方面的进步来控制复杂的工程系统。我们的假设是，可以定制(半监督)机器学习技术来从高性能仿真数据中提取信息，以处理复杂工程系统中用于实时决策的控制系统体系结构和控制算法的联合识别问题。该项目的研究目标具有巨大的潜力，有助于将高性能计算模拟和数据、机器学习和控制相结合，以改进用于控制复杂工程系统的最先进工具。该项目的试验台是一座风力发电厂。随着涡轮机变得更大，并且彼此放置得更近，涡轮机之间的空气动力耦合将会增加，从而产生一个真正大规模的复杂工程系统，尽管环境不确定和涡轮机组件的多变性，该系统仍必须运行。该项目的具体目标包括：高级学习算法，用于从风电场的大型涡流模拟数据中提取控制系统架构和训练算法；实时决策算法，用于从第一个目标中发现的特定地点库中选择架构和算法；以及实时算法，用于调整控制解决方案的关键参数，以进一步改善整体能源生产。该奖项反映了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。

In recent years, there have been significant advances in machine learning - statistical techniques that enable computers to "learn" using available data. Machine learning methods have demonstrated great success in image recognition, language translation, speech processing, and other consumer applications. This has led to great interest globally in academia, industry, and government. The drawback in purely machine learning methods is that it does not use the knowledge of physical properties of specific system which could significantly improve the performance of these methods. This EArly-concept Grant for Exploratory Research (EAGER) project will lead to fundamental results and methods that combine the advantages of machine learning techniques and knowledge of physical attributes of the system to enable decision making and control of complex engineered systems. The research will be conducted in the context of control of large wind energy plants. Maximizing power production despite variable and uncertain operating conditions in large wind plants is an unsolved problem that is ripe for transformative approaches and innovation. The research from this project is likely to transition to industry by leveraging connections with the NSF I-UCRC for Wind Energy Science, Research and Technology (WindSTAR) as wind plant owners and operators constantly seek new ways to improve annual energy production and reduce the cost of electricity from wind.The main idea of this EAGER project is to leverage advances in deep learning and high performance computing simulations for the control of complex engineered systems. Our hypothesis is that techniques from (semi-supervised) machine learning can be tailored to extract information from high performance simulation data to deal with the joint problem of identifying control system architectures and control algorithms for real-time decision making in complex engineered systems. The research goals of this project have great potential to contribute to the convergence of high performance computing simulations and data, machine learning, and controls to advance the state-of-art tools for controlling complex engineered systems. The testbed for the project is a wind plant. As turbines become larger, and are placed closer to one another, the aerodynamic coupling amongst turbines will increase resulting in a truly large-scale complex engineered system that must perform despite environmental uncertainty and variability of turbine components. Specific goals of this project include: Advanced learning algorithms for extracting control system architecture and training algorithms from large eddy simulation data of the wind farms; Real-time decision algorithms to select architecture and algorithms from site-specific libraries discovered in the first goal; and Real-time algorithms for tuning key parameters of the control solutions for additional improvements in the overall energy production.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.