Component-based machine learning for predicting representative time-series of energy performance in building design
Abstract:
The building industry has benefited by building performance simulation (BPS) for design assistance. Machine learning (ML) has been widely used for quick performance prediction; however, it lacks the flexibility to scale for new designs. By spatially and semantically decomposing the building design into components, this article links the ML approach with the system engineering paradigm of BPS to develop component-based machine learning (CBML). While previous use of CMBL focused on point predictions, this study proves that the CBML is able to predict dynamic time-series energy performance for new design cases by deriving a set of reusable model components. We trained and tested the ML model on a dataset of 1000 examples. The objective is to ascertain the ability of the ML model to generalize via different decomposition levels. Hourly energy predictions during the design phase are useful for equipment sizing, controlling peak energy demands, and leveling the load in the networks.
Citation: Chen, X., Singh, M.M. & Geyer, P. (2021). Component-based machine learning for predicting representative time-series of energy performance in building design. In 28th International Workshop on Intelligent Computing in Engineering, EG-ICE 2021. Berlin, Germany.