Methods for Dynamic Network Identification with Application to the Control of Smart Buildings

动态网络由相互作用的动态子系统组成。这样的网络出现在许多领域：活细胞、金融市场、互联网和电网就是一些例子。建筑中的供暖、通风和空调（HAVC）系统也可以通过动态网络进行建模，因为每个房间的气候取决于附近空间的气候。这种动态网络模型的知识对于设计和部署致力于提高能源效率和乘员舒适度的控制策略至关重要。然而，在实践中，这些网络的结构和动态要么是未知的，要么是不精确的。例如，由于所涉及的物理过程的复杂性，很难从物理定律中获得房间之间热相互作用的信息。该项目的目标是制定从测量数据中识别动态稀疏网络模型的算法。研究结果将支持HVAC系统的先进控制研究，以减少其能源使用并为电网提供需求侧灵活性。由于建筑消耗了全国75%的电力，通过智能建筑控制系统提高能源效率将有助于国家能源系统的可持续性。尽管“动态系统识别”是一个发展良好的领域，但动态网络识别领域却根本没有发展良好。传统的动态系统识别技术无法利用网络识别问题固有的稀疏性，而传统的机器学习技术大多只适用于静态网络。在这个项目中，我们结合了传统动态系统识别的思想，稀疏向量恢复的L1优化（来自压缩感知），以及机器学习的图形建模，以解决动态网络识别中的挑战。如果成功，该研究将(1)通过新算法为新兴的动态网络识别领域提供基础贡献，(2)使“智能建筑”技术能够在商业建筑中快速部署。此外，该项目将支持一些教育创新，以吸引来自代表性不足的群体的学生进入工程领域，并激发对工程的兴趣。

A dynamic network consists of interacting dynamic sub-systems. Such networks occur in many domains: living cells, financial markets, the Internet and the power grid are some examples. Heating, ventilation and air conditioning (HAVC) systems in buildings can also be modeled through dynamic networks since each room's climate depends on that of nearby spaces. Knowledge of such dynamic network models is essential to design and deploy control strategies devoted to the improvement of energy efficiency and occupant comfort. Yet, in practice the structure and dynamics of these networks are either unknown or imprecisely known. For instance, information on the thermal interaction among rooms is difficult to obtain from laws of physics due to the complexity of the physical processes involved. The goal of this project is to formulate algorithms for the identification of dynamic sparse network models from measured data. The research results will support the study of advanced controls for HVAC systems to reduce their energy use and to provide demand-side flexibility to the power grid. Since buildings consume 75% of the nation's electricity, improvement of energy efficiency through smart building control systems will contribute to the sustainability of the nation's energy system. Although 'dynamic system identification' is a well-developed field, the field of identification of dynamic networks is not at all well-developed. Traditional dynamic system identification techniques cannot exploit the inherent sparseness of the network identification problem, while traditional machine learning techniques are mostly applicable to only static networks. In this project we combine ideas from traditional dynamic system identification, L1 optimization for sparse vector recovery (from compressed sensing), and graphical modeling from machine learning to address the challenges in dynamic network identification. If successful, the research will (1) provide fundamental contribution to the nascent field of dynamic network identification through new algorithms, and (2) enable speedy deployment of 'smart building' technologies in commercial buildings. In addition, the project will support a number of educational innovations for attracting students from under-represented groups to engineering and generating excitement about engineering.