Component-based machine learning for performance prediction in building design. (15th October 2018)

• Record Type: Journal Article
• Authors: Geyer, Philipp; Singaravel, Sundaravelpandian
• Is Part Of: Applied energy. Volume 228(2018)
• Page Start: 1439
• Page End: 1453
• Publication Date: 2018-10-15
• Subjects: Component-based machine learning -- Systems engineering -- Parametric systems modeling -- Building performance prediction -- Building simulation; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2018.07.011
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: A component-based method of machine learning for performance prediction in engineering. Components instead of one monolithic model extend reusability and generalization. Flexible design support by machine learning for early design phases. Internal parameters between components allow insights in the "black box" of machine learning. Good prediction accuracies (error < 3.9%) for test cases different from the training model. Abstract: Machine learning is increasingly being used to predict building performance. It replaces building performance simulation, and is used for data analytics. Major benefits include the simplification of prediction models and a dramatic reduction in computation times. However, the monolithic whole-building models suffer from a limited transfer of models and their data to other contexts. This imposes a vital limitation on the application of machine learning in building design. In this paper, we present a component-based approach that develops machine learning models not only for a parameterized whole building design, but for parameterized components of the design as well. Two decomposition levels, namely construction level components (wall, windows, floors, roof, etc.), and zone-level components, are examined. Results in test cases show that, depending on how far the cases deviate from the training case and its data, high prediction quality may be achieved with errors as low as 3.7% for cooling and 3.9% for heating.