CAREER: Optimal Control of Energy Systems via Structured Neural Networks: A Convex Approach

目前，电力能源系统正处于前所未有的变革时期。一方面，屋顶太阳能和智能建筑管理系统等技术的大规模部署有可能使电力系统更加高效、可持续和可靠。另一方面，实现这一承诺已被证明远非微不足道，因为许多功能仍未使用。两个主要的系统是商业和工业建筑的供暖、通风和空调（HVAC）系统，以及配电系统中的分布式能源。控制这些系统的一个基本挑战是它们的行为通常由具有未知参数的复杂动力学控制。例如，不同区域的温度设定值与暖通空调功耗之间的关系是由一组非线性高维偏微分方程控制的，其参数取决于具体的建筑特征，在实践中难以测量。类似地，配电系统是由非线性交流潮流方程控制的，但由于它们通常不受监测，因此它们的拓扑结构和线路参数要么是未知的，要么是严重过时的。这个CAREER建议通过利用现在可用的大量测量数据来解决这个挑战。从根本上不同于许多现有的人工智能应用，控制这些系统行为的物理定律——建筑热传递的热力学定律和功率流方程——得到了很好的研究，但系统参数尚不清楚，也不容易测量。该项目的目标是提供具有可证明保证的算法，将物理定律与数据相结合，以安全有效地运行这些能源系统。具体来说，我们提出了一个基于模型的框架，该框架使用结构化神经网络来实现模型的可追溯性和可表征性，通过将它们设计为从输入到输出的凸。该项目将通过与校园可持续发展办公室和当地公用事业公司合作，将研究和教育紧密结合起来，从而培养一代在电力系统和机器学习方面合格的专业人员。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The electric energy system is currently undergoing a period of unprecedented transformations. On the one hand, large-scale deployment of technologies such as rooftop solar and smart building management systems have the potential to make the power system more efficient, sustainable and reliable. On the other hand, achieving this promise has proven to be far from trivial, as many capabilities remain unused. Two primary systems of interest are the heating, ventilation, and air conditioning (HVAC) systems of commercial and industrial buildings, and distributed energy resources in the power distribution system. A fundamental challenge in controlling these systems is that their behaviors are often governed by complex dynamics with unknown parameters. For instance, the relationship between temperature setpoints in different zones and the HVAC power consumption is governed by a set of nonlinear high dimensional partial differential equations, whose parameters depend on detailed building characteristics that are difficult to measure in practice. Similarly, the distribution system is governed by nonlinear AC power flow equations, but since they are typically not monitored, their topology and line parameters are either not known or severely outdated.This CAREER proposal addresses this challenge by leveraging the significant amounts of measurement data that are now becoming available. Fundamentally different from many existing AI applications, the physical laws governing the behaviors of these systems---laws of thermodynamics for heat transfers in buildings and power flow equations---are well studied, but the system parameters are not known and cannot be easily measured. The goal of this project is to provide algorithms with provable guarantees that combine physical laws with data to safely and efficiently operate these energy systems. Specifically, we present a model-based framework that uses structured neural networks to achieve both model tractability and representability, by designing them to be convex from input to output. This project will tightly integrate research and education by working with the campus sustainability office and the local utility, thus training a generation of professionals qualified both in power systems and machine learning.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.