Content area
Full Text
ABSTRACT
Inaccuracy in classical forward energy modeling often occurs. In this study, an inverse data-driven approach is followed by applying machine learning to building energy prediction at the national level. The objectives are to: 1) develop a data-driven model for post-occupancy energy use prediction, 2) inform future efforts on targeting effective data inputs and machine learning techniques. Commercial Buildings Energy Consumption Survey (CBECS) dataset from Energy Information Administration (EIA) was used to build the model. Machine learning algorithms such as Gradient Boosting, Linear Regression, and Random Forest Tree, etc. were applied and evaluated. Benchmarking datasets from New York City (NYC) and Chicago were used to validate the model performance on local buildings and future energy use. The result shows in general Gradient Boosting has better performance compared to other algorithms. A data-driven model with 54 building features was established, by targeting the most important features of the envelope, building system, internal use, energy resource, and post-occupancy maintenance. The accuracy of raw NYC benchmarking dataset is then assessed. Finally, the potentials and limitations of applying data-driven approach with machine learning to building energy use prediction are discussed.
(ProQuest: ... denotes formulae omitted.)
INTRODUCTION
Energy modeling is widely used in building and energy infrastructure design, building code development, and energy policy formulation. Accurate and timely energy predictions are required to inform decision making. Currently, most of the energy predictions are based on the forward method, to build thermodynamic models using engineering principles. Computer programs such as EnergyPlus and eQuest are often used and energy code including AHSRAE 90.1 and California Title 24 are followed. Intensive workforce input is required on detailed building information collection, manual data input, and quality assurance of model. However, Inaccuracy usually occurs, due to factors including incomplete building information, unpredictable occupancy behavior, inaccurate assumptions, limited workforce input, and changing operations strategies in post-occupancy phase. New Building Institute (NBI) and U.S. Green Building Council (USGBC) studied post-occupancy energy performance for Leadership in Energy and Environmental Design (LEED) projects (Turner and Frankel 2008). It showed inconsistency exists between modeled energy use intensity and actual energy use intensity and there's a normalized root-mean-square-error of 0.18 (Turner and Frankel 2008; Yan et al. 2015).
With the development of computing, machine learning has diverted researchers' attention from...