Latent space learning of energy consumption and indoor environmental quality data in context of building technology and construction informatics

近年来，与室内环境质量(IEQ)、居住者行为(OB)和建筑能耗相关的传感器数据的可用性显著增加。这种类型的数据在以居住者为中心的健康建筑运营中或在建筑系统内部署先进的数字技术方面正变得越来越重要。与建筑物相关的数据量不断增加，通常采用数据驱动建模来处理这些数据。在这里，机器学习(ML)被广泛应用于处理大量数据。数据处理的典型目标是故障检测和随机或复杂影响(如OB)的预测ML建模，以及包括模型预测控制(MPC)在内的数据驱动控制策略。总而言之，所有这些技术在经验上都比用于实现建模目标的传统的基于规则的模型或统计模型表现出更好的性能。除了经验证明的更高的建模精度外，这些建模范例不需要对与目标建筑相关的领域知识进行显式编码。这种对特定领域信息的显式编码的独立性为开发通用的模型和框架提供了一个假设的机会，这些模型和框架可以用于高度不同的建模。遵循这种方法，在建筑能耗的背景下，单个模型可以用于重建室内空气温度和OB数据的时间序列，并且可以进一步应用于显著不同的建筑以及地区甚至城市尺度的能耗时间序列。通常，可以使用类似的方法对这些高度不同的数据流进行建模。然而，由于每个目标域的潜在空间不同，不能保证相同模型的适用性。为了更科学地使用与建筑、IEQ和能耗建模相关的数据进行更通用的建模，本项目旨在实现以下目标：1)创建一个语义空间来映射不同目标变量和不同建筑相关的数据流的相似性和比对；2)确定允许针对特定目标开发的模型在应用于新的建模目标(领域适应)时适应领域特定数据的方法，以及3)使用所获得的关于这些数据源的潜在空间的知识来生成合成IEQ、OB和能源消耗数据集。作为领域自适应技术，我们在本项目中考虑了迁移学习、基于自动编码神经网络的潜在空间学习(AENN)、对抗性学习，最后是物理信息学习。

In recent years, the availability of sensor data related to the indoor environmental quality (IEQ), occupant behavior (OB), and energy consumption in buildings increased significantly. This type of data is gaining more importance within occupant-centric healthy building operation or in the deployment of advanced digital technologies within the building system. Related to the constantly increasing volume of data associated with buildings, the data-driven modeling is commonly applied to process these data. Here, machine learning (ML) is widely adopted for processing large amount of data. Typical objectives of data processing are fault detection and predictive ML modeling of stochastic or complex effects such as OB, as well as data-driven control strategies including model predictive control (MPC). In sum, all these technologies demonstrated empirically better performance than the conventional rule-based or statistical models used to achieve the modeling objectives. In addition to the empirically shown higher modeling accuracy, these modeling paradigms do not require explicit encoding of domain knowledge related to the target building. This independence from explicit encoding of domain-specific information offers a hypothetical opportunity to develop generic models and frameworks, that could be used for highly variant modeling.Following this approach in the context of energy consumption in buildings, a single model could then be used to reconstruct the time-series of indoor air temperatures and OB data, and could furthermore be applied to significantly different buildings as well as energy consumption time-series at district or even urban scale granularity. In general, these highly variant data streams might be modeled using similar approaches. However, the applicability of identical models is not warranted due to the differences in the latent space of each target domain. In order to make a scientific progress towards more generic modeling using data related to buildings, IEQ, and energy consumption modeling, this project aims to fulfill the following objectives: 1) create a semantic space to map the similarities and comparing the data streams related to different target variables and different buildings; 2) identify approaches that allow models developed for a particular objective to accommodate for domain specific data when applying them for new modeling objectives (domain adaptation), and 3) use the knowledge gained about the latent space of these data sources to generate synthetic IEQ, OB, and energy consumption data sets.The methodological focus of the proposed project lies on machine learning methods for latent space learning and on methodologies for ensuring the models’ applicability in different target domains. As domain adaptation techniques, we consider transfer learning, latent space learning using autoencoder neural networks (AENN), adversarial learning, and lastly physics-informed learning in this project.