Using surrogate models in the integrated design process for high-performance buildings

2016年12月，泛加拿大清洁增长和气候变化框架承诺履行其国际承诺，即到2030年将温室气体(GHG)排放量在2005年水平的基础上减少30%，其中2000万吨预计来自建筑业。制造低碳建筑的技术和施工过程众所周知；然而，建筑业本质上需要对建筑存量产生长期影响的前期决策。现在建造的建筑预计到2030年甚至2050年仍然存在。传统的建筑能耗分析、碳核算和生命周期成本分析都是通过详细的模拟来进行的，既耗时又昂贵。为了实现温室气体减排目标，必须开发有效集成到早期概念筛选和设计决策过程中的工具，允许快速、可靠地评估选项和措施。该项目将把一种称为代理建模的机器学习方法应用于建筑模拟领域，重点是将较大的问题分解为子模型(如将围护结构负荷与暖通空调交付分离)，以便更好地与多学科、集成的设计过程保持一致，并纳入行业知识和生产供工程师实际使用的工具。代理建模依赖于拥有大量经过验证的基线模型来有效地探索新设计的设计空间(应用机器学习优化技术来确定新设计在一系列预渲染模型中的位置，而不必创建新的高保真模型)。因此，基础基线模型的质量至关重要，该项目将制定一种部分基于行业工作流程进行质量控制的验证方法。这将反过来支持未来的工作，为学术研究产生高质量的基线，并在工业中找到有效的应用。这个项目最终将采用上述模块和方法，并开发供行业专业人员使用的互动工具。这将包括研究有效的可视化和界面设计，并将在整个项目时间表内通过需求评估调查、讲习班和实地测试的各个阶段纳入迭代反馈过程。

In December 2016, the Pan-Canadian Framework on Clean Growth and Climate Change committed to meeting its international commitment of a 30% reduction of greenhouse gas (GHG) emissions below 2005 levels by 2030, of which 20 million tonnes are expected to come from the buildings sector.The technology and construction processes to make low carbon buildings are well understood; however, the buildings industry inherently requires front-loaded decision-making with long-term implications on building stock. The buildings constructed now are expected to still be present in 2030 and even 2050. Conventional analysis of building energy performance, carbon accounting and life-cycle costs through detailed simulation is time-consuming and expensive. In order to achieve the GHG reduction targets, tools must be developed that integrate effectively into the early concept screening and design decision-making processes, allowing rapid, reliable evaluation of options and measures.This project will adapt a machine learning method termed surrogate modeling to the building simulation domain, with an emphasis on breaking up the larger problem into sub-models (such as separating enclosure loads from HVAC delivery) that better align with a multi-discipline, integrated design process, as well as incorporating industry knowledge and producing tools for practical use by engineers, architects and planners during early concept design phases.Surrogate modeling relies on having a large number of validated baseline models to efficiently explore the design space for a new design (applying machine learning optimization techniques to identify where the new design fits among a constellation of pre-rendered models, without having to create a new high-fidelity model). The quality of the underlying baseline models is therefore vital, and this project will develop a validation methodology partly based on industry workflows for quality control. This will in turn support future work generating high-quality baselines for academic research, and in finding effective application in industry.This project will finally take the modules and methods described above and develop interactive tools for use by industry professionals. This will include research into effective visualizations and interface design, and will incorporate an iterative feedback process through phases of needs assessment surveys, workshops and field testing throughout the project timeline.