Surrogate modelling of building energy use

替代模型作为一种革命性的估算建筑设计能源使用的方法正在兴起，这对于提供有助于防止灾难性气候变化所需的低能耗建筑至关重要。新建筑的设计和旧建筑的改造必须达到非常低的能源使用水平，以减少目前由建筑部门造成的约1/3的碳排放。这需要仔细平衡不同方面的性能，例如尽量减少冬季供暖需求与夏季制冷需求或过热风险。能量模拟程序，数值计算热传递在拟议的建筑设计，从而其能源需求目前广泛应用于工业。然而，所需的计算时间（每个设计选项几分钟）不利于直观地探索许多不同设计的性能。代理建模在可能的建筑设计空间中执行许多模拟，然后将机器学习模型与生成的合成数据相匹配，将设计变量映射到性能指标。这些代理模型是近似的，但已被证明可以非常准确地表示底层系统。替代模型几乎可以即时评估，允许设计师使用交互式应用程序以更自然的方式探索建筑性能权衡，其中视觉性能输出立即响应输入的变化。改进代理模型所需的科学方法跨越了传统的建筑能源模拟和机器学习方法。前者涵盖了控制建筑物中能量流动的基本物理，模拟它们的计算方法，以及自动改变关键参数的软件方法。代理建模所需的机器学习技术包括神经网络的基础知识、捕捉底层系统复杂性所需的高级方法，以及评估性能所需的统计技术。目标是发展和完善代理建模技术，将推动这一新兴领域大胆的新方向。这将涉及用于此类模型的机器学习技术及其在特定建模领域的应用。研究结果在短期效益和潜在的建筑能源模拟领域变革方面都具有重要意义。正如已经应用初步结果的行业合作所证明的那样，所产生的模型和方法立即适用。从长远来看，替代模型可以取代早期设计探索中的建筑能量模拟，在早期设计探索中，即时评估比绝对准确更有用。培训具有跨越现有模拟方法和新兴机器学习技术的独特技能的高素质人才也将受益。

Surrogate modelling is emerging as a revolutionary means of estimating energy use for building designs, which will be essential in delivering the low-energy buildings required to help prevent catastrophic climate change. New building designs and retrofits to old buildings must achieve very low levels of energy use to reduce the ~1/3 of carbon emissions currently caused by the building sector. This requires careful balancing of different aspects of performance, e.g. minimising winter heating demand versus summer cooling demand or overheating risk. Energy simulation programs that numerically calculate heat transfer in a proposed building design and thus its energy requirements are currently widely used in industry. However, the computational time required (several minutes per design option) is not conducive to the intuitive exploration of the performance of many different designs. Surrogate modelling performs many simulations spanning the design space of possible building designs, then fits a machine learning model to the resulting synthetic data that maps design variables to performance metrics. These surrogate models are approximate, but have been shown to represent the underlying system with great accuracy. Surrogate models are almost instant to evaluate, allowing designers to explore building performance trade-offs in a more natural manner using interactive apps in which visual performance outputs respond instantly to changes in inputs. The scientific approaches required to improve surrogate modelling span traditional building energy simulations and machine learning methods. The former covers the fundamental physics that governs energy flows in buildings, the computational means of simulating them, and the software methods to vary key parameters automatically. The machine learning techniques necessary for surrogate modelling include the fundamentals of neural networks, the advanced methods needed to capture the complexities of the underlying system, and the statistical techniques required to assess performance. The objective is to develop and refine surrogate modelling techniques that will advance this nascent field in bold new directions. This will involve both the machine learning techniques to be used for such models and their application to specific fields of modelling. The outcomes will have great significance in both short-term benefits and in the potential to revolutionize the field of building energy simulation. The models and methods produced are immediately applicable, as demonstrated by the industry collaborations already applying initial results. In the longer term, surrogate modelling could replace building energy simulation in early-stage design exploration, where instant evaluation is more useful than absolute accuracy. There will also be benefits in the training of highly-qualified personnel who possess a unique skillset spanning existing simulation approaches and emerging machine learning techniques.