ERI: Towards Data-driven Learning and Control of Building HVAC Systems

在美国，建筑物每年的能耗约占总能耗的40%，其中约44%用于供暖、通风和空调(HVAC)系统。在美国，建筑物通过其能源使用间接贡献了每年能源消耗产生的二氧化碳排放总量的约35%。通过使用先进的暖通空调系统控制，有很大的潜力减少建筑物的能源消耗及其相关的环境影响。此外，美国能源部已经启动了一项关于电网互动节能建筑的国家战略，这将有助于将建筑的能源效率和需求灵活性提高两倍，并提高电网的效率和可靠性。模型预测控制(Model Predictive Control，MPC)作为一种很有潜力的先进建筑控制技术应运而生。然而，它向实践的过渡受到根本性挑战的阻碍，包括开发用于控制的准确建筑模型的难度和高昂的成本，以及在建筑物中实施MPC的高度工程努力。该项目将为克服建筑预测控制的这些基本挑战，整合机器学习、控制理论、优化理论和建筑科学奠定科学基础。它将为暖通空调系统的数据驱动学习和控制开发新的方法和算法，并在真实建筑的实验中进行演示。更广泛地说，这项研究将促进学习和控制复杂物理系统的科学知识，这将在许多其他应用中产生深远影响。它将把研究工作融入教育和推广活动中，包括为本科生提供新的研究机会，以及向K-12学校的学生和公众开展推广活动，以加深公众对建筑能效及其技术的了解。该项目的目标是开发一种新的框架、理论和方法，通过将机器学习、动力学、控制和优化联系起来，有效和高效地进行数据驱动的建模、学习和控制，以构建暖通空调系统。为此，本项目的具体目标是开发(1)物理信息数据驱动的建模方法，用于构建暖通空调系统，有效地将适当的领域见解融入到机器学习模型中；(2)主动学习方法，以获得最丰富的实验数据，以提高模型的精度和样本效率；以及(3)基于物理信息的数据驱动模型的有效公式和高效的优化算法，用于基于学习的预测控制(LB-MPC)。这些方法的可行性和优点将通过对各种真实建筑的广泛实验验证来验证。该项目通过建立全面的物理信息数据驱动的建模基础和一套用于建筑暖通空调系统的学习、控制和优化方法，为建筑提供了一条走向自主、高效和实用的LB-MPC的道路。它将弥合黑盒和灰盒建模方法之间的差距，通过将适当的领域见解有效地整合到数据驱动模型中，实现可靠、样本高效和准确的数据驱动模型，从而推动面向控制的建筑建模技术的发展。它还有可能通过为学习找到最佳激励轨迹的主动学习方法来转换用于数据驱动的建筑建模的训练数据集合。最后，它将通过制定有效和易处理的LB-MPC优化问题并为高效地解决这些问题定制算法来克服数据驱动控制的计算挑战。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。

Buildings account for about 40% of the total annual energy consumption in the U.S., of which about 44% is for heating, ventilation, and air conditioning (HVAC) systems. Indirectly through their energy use, buildings contribute about 35% of the total annual carbon dioxide emissions from energy consumption in the U.S. There is a significant potential for reducing energy use of buildings and their associated environmental impact by using advanced control of HVAC systems. Moreover, the U.S. Department of Energy has initiated a national strategy on grid-interactive efficient buildings, that will help triple the energy efficiency and demand flexibility of buildings and improve the power grid efficiency and reliability. Model Predictive Control (MPC) has emerged as a potential advanced building control technology to attain these goals. However, its transition to practice has been hampered by fundamental challenges, including the difficulty and high cost of developing accurate building models for control and the high engineering effort to implement MPC in buildings. This project will lay the scientific foundation for overcoming these fundamental challenges of MPC for buildings, integrating machine learning, control theory, optimization theory, and building science. It will develop novel methods and algorithms for data-driven learning and control of HVAC systems, and demonstrate them in experiments with real buildings. More broadly, this research will advance scientific knowledge in learning and control of complex physical systems, which will have far-reaching impacts in many other applications. It will integrate research efforts into education and outreach, including new research opportunities for undergraduate students and outreach activities to K-12 school students and the public to enrich public understanding of building energy efficiency and its technologies. These efforts are complemented by extensive recruitment and mentorship of underrepresented minorities in STEM.The goal of this project is to develop a new framework, theory, and methods for effective and efficient data-driven modeling, learning, and control of building HVAC systems by bridging machine learning, dynamics, control, and optimization. To this end, the specific objectives of this project are to develop (1) a physics-informed data-driven modeling approach for building HVAC systems that effectively incorporates appropriate domain insights into machine learning models; (2) active learning methods to obtain the most informative experimental data for improving model accuracy and sample efficiency; and (3) effective formulations and efficient optimization algorithms for learning-based MPC (LB-MPC) with the physics-informed data-driven models. The feasibility and merits of these methods will be validated through extensive experimental verification on a variety of real buildings. This project provides a path towards autonomous, performant, and practical LB-MPC for buildings by establishing a holistic physics-informed data-driven modeling foundation and a suite of learning, control, and optimization methods for building HVAC systems. It will bridge the gap between black-box and gray-box modeling approaches to advance the state of the art on control-oriented building modeling by effectively incorporating appropriate domain insights into data-driven models, enabling reliable, sample-efficient, and accurate data-driven models. It also has the potential to transform the collection of training data for data-driven building modeling through active learning methods that find the optimal excitation trajectory for learning. Finally, it will overcome the computational challenges of data-driven control by formulating effective and tractable LB-MPC optimization problems and tailoring algorithms for solving these problems efficiently.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.