Computationally efficient integrated design and predictive control of flexible energy systems using multi‐fidelity simulation‐based Bayesian optimization
Computationally efficient integrated design and predictive control of flexible energy systems using multi‐fidelity simulation‐based Bayesian optimization
复制标题
DOI：
10.1002/oca.2817
复制
发表时间：
2021-11
影响因子：
1.8
通讯作者：
Farshud Sorourifar;Naitik A. Choksi;J. Paulson
中科院分区：
文献类型：
--
作者：
Farshud Sorourifar;Naitik A. Choksi;J. Paulson
我们提出了一种多保真黑盒优化方法，用于在存在不确定性的情况下约束非线性系统的集成设计和控制（IDC）。 IDC 框架对于下一代（灵活）制造和能源系统的系统设计变得越来越重要。然而，当（i）动态发生在比系统生命周期短得多的时间尺度上，（ii）不确定性由具有高方差的连续随机变量描述，以及（iii）操作决策涉及离散变量和连续变量的混合时，确定实际IDC问题的最佳解决方案是很困难的。我们没有积极简化问题以提高易处理性，而是使用高质量决策规则开发基于模拟的优化程序，将可在线测量的信息映射到最佳控制操作。特别是，我们依赖贝叶斯优化（BO）框架，该框架已被证明在嘈杂且评估成本高昂的目标函数上表现良好。我们还讨论了如何扩展 BO 以利用计算成本更低的高保真 IDC 成本函数的低保真近似。这项工作描述了三种主要的低保真度近似策略，它们与系统模拟器的简化、决策规则求解方法和时间网格有关。最后，我们展示了在不确定的天气和需求条件下，在一年的规划范围内每小时变化的情况下，多保真 BO 在太阳能建筑供暖/制冷系统（具有电池和电网支持）设计中的优势。
We present a multi‐fidelity black‐box optimization approach for integrated design and control (IDC) of constrained nonlinear systems in the presence of uncertainty. The IDC framework is becoming increasingly important for the systematic design of next‐generation (flexible) manufacturing and energy systems. However, identifying optimal solutions to realistic IDC problems is intractable when (i) the dynamics occur on much shorter timescales than the system lifetime, (ii) the uncertainties are described by continuous random variables with high variance, and (iii) operational decisions involve a mixture of discrete and continuous variables. Instead of aggressively simplifying the problem to improve tractability, we develop a simulation‐based optimization procedure using high‐quality decision rules that map information that can be measured online to optimal control actions. In particular, we rely on the Bayesian optimization (BO) framework that has been shown to perform very well on noisy and expensive‐to‐evaluate objective functions. We also discuss how BO can be extended to take advantage of computationally cheaper low‐fidelity approximations to the high‐fidelity IDC cost function. Three major low‐fidelity approximation strategies are described in this work, which are related to the simplification of the system simulator, decision rule solution method, and time grid. Lastly, we demonstrate the advantages of multi‐fidelity BO on the design of a solar‐powered building heating/cooling system (with battery and grid support) under uncertain weather and demand conditions with hourly variation over a year‐long planning horizon.