supercomputer assisted generation of machine learning...
TRANSCRIPT
![Page 1: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/1.jpg)
Supercomputer Assisted Generation of Machine Learning Agents for the Calibration of Building Energy Models
Jibonananda Sanyal Joshua New Richard Edwards
Oak Ridge National Laboratory
University of Tennessee, Knoxville
![Page 2: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/2.jpg)
2 Managed by UT-Battelle for the U.S. Department of Energy
Sustainability is the Defining Challenge of Our Time
• Buildings in China
– 60% of urban building floor space in 2030 has yet to be built
• Buildings in India
– 67% of all building floor space in 2030 has yet to be built
• Buildings in the U.S. consume:
• 73% of all electricity
• 55% of all natural gas
Buildings, 41% Industry 31%
Transportation, 28%
U.S. Primary Energy Consumption in 2010
![Page 3: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/3.jpg)
3 Managed by UT-Battelle for the U.S. Department of Energy
Building Energy Modeling
![Page 4: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/4.jpg)
4 Managed by UT-Battelle for the U.S. Department of Energy
The biggest hurdle is the
cost-effective calibration of
Building Energy Models
![Page 5: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/5.jpg)
5 Managed by UT-Battelle for the U.S. Department of Energy
EnergyPlus takes about 3 minutes for a simulation… already too long for industry
![Page 6: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/6.jpg)
6 Managed by UT-Battelle for the U.S. Department of Energy
.
.
.
E+ Input
Model
Autotune
![Page 7: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/7.jpg)
7 Managed by UT-Battelle for the U.S. Department of Energy
Each building is unique
• Buildings must conform to code
• DOE has standard reference buildings
– Representative of U.S. building stock
– Starting point of BEM experts
Surrogate
Modeling
EnergyPlus
Modeling
Time Accuracy
![Page 8: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/8.jpg)
8 Managed by UT-Battelle for the U.S. Department of Energy
Types of buildings
• Residential
– 5 million simulations
• Medium Office
– 1 million
• Stand-alone retail
– 1 million
• Warehouse
– 1 million
• 8 million = 270+TBs of data
• 16 ASHRAE climate zones
• Vintage: Old, Recent, New construction
![Page 9: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/9.jpg)
9 Managed by UT-Battelle for the U.S. Department of Energy
High Performance Computing Resources
One of the largest users of NICS Nautilus, over 300,000 SUs SDSC Gordon
NICS Kraken
ORNL Titan
![Page 10: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/10.jpg)
10 Managed by UT-Battelle for the U.S. Department of Energy
MLSuite: Machine Learning on
Supercomputers
• Support Vector Machines
• Genetic Algorithms
• FF and Recurrent Neural Networks
• (Non-)Linear Regression
• Self-Organizing Maps
• C/K-Means
• Ensemble Learning Shared memory Nautilus and
other distributed memory
supercomputers
![Page 11: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/11.jpg)
11 Managed by UT-Battelle for the U.S. Department of Energy
So, how well can we calibrate?
![Page 12: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/12.jpg)
12 Managed by UT-Battelle for the U.S. Department of Energy
Using metrics such as CVRMSE and MAPE
Below 0.5% error
ASHRAE requires only within 30%
![Page 13: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/13.jpg)
13 Managed by UT-Battelle for the U.S. Department of Energy
Key technical merits
• EnergyPlus is desktop software
– Parametric simulations traditionally scale poorly
– Scalability on Leadership Class Infrastructure: up to 131,072 cores
• Surrogate modeling, Machine learning agents
– Runtime: from 3 minutes down to 3 seconds
• Big-data
– 45 TB in 68 minutes for ½ million E+ runs
– Total: 270+ TB raw
• Data analysis
– Move computation to data
– Parallel Big-Data R
![Page 14: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/14.jpg)
14 Managed by UT-Battelle for the U.S. Department of Energy
Autotune is
simulation engine agnostic
![Page 15: Supercomputer Assisted Generation of Machine Learning ...web.eecs.utk.edu/.../2013_XSEDE_LightningTalk.pdf · 2 Managed by UT-Battelle for the U.S. Department of Energy Sustainability](https://reader033.vdocuments.net/reader033/viewer/2022050211/5f5d8b9e4609d860030a5deb/html5/thumbnails/15.jpg)
Jibonananda Sanyal Joshua New
Richard Edwards
Oak Ridge National Laboratory
University of Tennessee, Knoxville
Supercomputer Assisted Generation of Machine Learning Agents for the Calibration of Building Energy Models