Constructing large scale surrogate models from big data and artificial intelligence. (15th September 2017)

• Record Type: Journal Article
• Authors: Edwards, Richard E.; New, Joshua; Parker, Lynne E.; Cui, Borui; Dong, Jin
• Is Part Of: Applied energy. Volume 202(2017)
• Page Start: 685
• Page End: 699
• Publication Date: 2017-09-15
• Subjects: Machine learning -- EnergyPlus -- Building simulation -- Energy modeling -- Surrogate model; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2017.05.155
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: World's fastest supercomputer used for construction of benchmark datasets. Lasso regression and feed forward neural networks are scaled. Machine learning instances are evaluated for prediction vs. runtime performance. 3 datasets with 7–156 software inputs are used to predict over 3, 000, 000 outputs. Surrogate building energy model of EnergyPlus runs 60× faster. Abstract: EnergyPlus is the U.S. Department of Energy's flagship whole-building energy simulation engine and provides extensive simulation capabilities. However, the computational cost of these capabilities has resulted in annual building simulations that typically requires 2–3 min of wall-clock time to complete. While EnergyPlus's overall speed is improving (EnergyPlus 7.0 is 25–40% faster than EnergyPlus 6.0), the overall computational burden still remains and is the top user complaint. In other engineering domains, researchers substitute surrogate or approximate models for the computationally expensive simulations to improve simulation and reduce calibration time. Previous work has successfully demonstrated small-scale EnergyPlus surrogate models that use 10–16 input variables to estimate a single output variable. This work leverages feed forward neural networks and Lasso regression to construct robust large-scale EnergyPlus surrogate models based on 3 benchmark datasets that have 7–156 inputs. These models were able to predict 15-min values for most of the 80–90 simulation outputs deemed most important by domain experts within 5% (whole building energy within 0.07%) and calculate those results within 3 s, greatly reducing the required simulation runtime for relatively close results. The techniques shown here allow any software to be approximated by machine learning in a way that allows one to quantify the trade-off of accuracy for execution time.