Experiments in using nonlinear regression for business activity normalization in the Energy Star benchmarking method
Abstract
Building energy benchmarking provides information about a building’s energy performance compared with its peers. This information is important in pursuing more energy efficient building operations. In the United States, the... [ view full abstract ]
Building energy benchmarking provides information about a building’s energy performance compared with its peers. This information is important in pursuing more energy efficient building operations. In the United States, the Environmental Protection Agency (EPA) Energy Star is the most widely used building benchmarking tool, and more than 30 state and local governments have implemented regulations asking for its adoption. One important procedure in Energy Star’s benchmarking methodology is establishing a regression model to normalize building business activity features, as well as to compute the reference energy use intensity (EUI) values that serve as a reference for peer facilities. Energy Star currently uses a linear model for this regression task, but its lack of prediction power (in terms of low coefficient of determination) suggests a nonlinear model could more effectively normalize these features. We applied two kernel-based regression algorithms to establish flexible nonlinear models: kernel ridge regression and support vector regression, with 10-fold cross validation to avoid overfitting. By conducting experiments on the 2003 and 2012 Commercial Building Energy Consumption Survey (CBECS) office building datasets, we show that there is not enough data to capture the nonlinear relationships, if they exist. The experiments also show that the modeled linear and nonlinear relationships between considered business activities and the EUI changed considerably between 2003 and 2012. These results suggest that the relationship between business activities and the EUI of office buildings in the United States cannot be accurately captured by neither the nonlinear models implemented in this study, nor the linear model used by EPA, and that this relationship may be changing with time.
Authors
-
Xuechen Lei
(Carnegie Mellon University)
-
Mario Bergés
(Carnegie Mellon University)
-
Burcu Akinci
(Carnegie Mellon University)
-
Aarti Singh
(Carnegie Mellon University)
Topic Areas
Facility management and BEMS/HEMS , Big Data, data mining and machine learning
Session
O2 » Facility management (10:45 - Tuesday, 5th June, Sonaatti 2)
Paper
ICCCBE_0131Submit0415.pdf
Presentation Files
The presenter has not uploaded any presentation files.