Component-based machine learning modeling for predicting building energy performance – An overview

In addition to the use of building simulation, in recent years data-driven methods using ma-chine learning (ML) have been developed that have the ability to act as surrogate for building energy prediction (Amasyali and El-Gohary, 2018; Westermann and Evins, 2019). Such data-driven models have high potential by re-ducing modelling and computation effort drasti-cally and thus allowing extension of examined configurations by two to three magnitudes to consider variety of designs as well as uncer-tainties. However, there are two major draw-backs of such models: (1) First, they form a black box, i.e. they do not use physics when predicting which excludes the possibility to check and understand the calculations; (2) The prediction relies on the data that has been used to train the models, which causes problems for cases that “extrapolate” outside the training data.For this context, we have developed a method aligned with systems engineering that we call component-based machine learning (CBML) (Geyer and Singaravel, 2018). This method does not train a model for the whole building, as it is the case for most state-of-the-art energy predictions, which we call monolithic. CBML creates data-driven models for the components of a building, i.e. walls, windows, etc. Due to its object-oriented modelling paradigm as typical feature of building information model (BIM) (Borrmann et al., 2018), this approach allows for a close integration with the model structures of digital design and planning in BIM. One the one hand, the close integration with the struc-tures of design and planning is superior gener-alization and reusability of the data-driven mod-els created by machine learning; On the other hand, CBML uses exclusively engineering quantities to connect data-driven components leading to built-in interior interpretability and explainability (Geyer et al., 2023).This paper provides an overview of the current development of CBML. First, it introduces the method of CBML; experiments on the build en-velope and building systems show its accuracy and the generalization of CBML predictions compared to predictions of monolithic models. The next section extends static predictions of one value to time series predictions. A further step in the combination of data-driven compo-nents with building simulation; this leaves strict limitation to building components and includes process components. Finally, we will discuss interpretability and explainability of CBML.