OAK RIDGE, Tenn. — As many as 3,000 parameters can be specified when modeling a building’s energy use. Jibonananda Sanyal of the U.S. Department of Energy’s (DOE’s) Oak Ridge National Laboratory (ORNL) Building Technologies Research and Integration Center and principal investigator Joshua New have developed automated, or “Autotune,” calibration software, which reduces the amount of time and expertise needed to optimize building parameters for cost and energy savings.
Over the last 20 years, DOE has invested in EnergyPlus, its flagship whole-building energy simulation software. The tool estimates energy usage based on weather data and the thousands of input parameters related to HVAC, water heating, lighting, weather interaction, occupancy schedules, and more.
Use of EnergyPlus or other building energy modeling (BEM) tools for energy-efficient building design is growing. But before using BEM to identify energy improvements to existing buildings, BEM parameters must first be collected, entered into the tool, and adjusted so outputs reasonably match past energy usage. This can be a time-consuming chore, but it’s often required to receive tax rebates and utility incentives.
“Currently, the biggest barrier is the cost of getting an accurate model of the pre-retrofit building because it requires hiring an expert,” Sanyal said.
Sanyal and New’s Autotune software will reduce the time and expertise needed to achieve an accurate model, and they’re collaborating with universities and industry to make their approach accessible to more building professionals and owners.
“The Autotune methodology uses multiparameter optimization techniques, in combination with big data mining-informed artificial intelligence agents, to automatically modify software inputs so simulation output matches measured data,” New said.
“Instead of having a human change the knobs, so to speak, the Autotune methodology does that for you,” Sanyal said.
To develop their Autotune software, Sanyal and New used DOE supercomputing and computational resources — including ORNL’s 27-petaflop Titan supercomputer and the National Institute for Computational Sciences’ Nautilus system — to perform millions of EnergyPlus simulations for a range of standard building types with generated data totaling hundreds of terabytes. On Titan, the team has been able to run annual energy simulations for more than half a million buildings in less than one hour using over a third of Titan’s nearly 300,000 CPU cores in parallel.
Additionally, they worked with building technology experts to identify about 150 of the most important parameters. By focusing on those, they can reduce computational load while still ensuring highly accurate results. The software uses machine learning algorithms to “learn” successful versus unsuccessful paths to optimization. In this way, if similar building input parameters are introduced later, the software optimizes the results more quickly by cutting out what didn’t work before.
While programmatic guidelines for tax rebates and utility incentives often require an error rate below 30 percent when calibrating building models to monthly utility bills, Autotune’s fully-automated process has routinely calibrated models to an error rate below 1 percent on all building types tested. With such precision, an overnight Autotune process is far less costly than the time it would take an expert to manually calibrate a model.
The team is currently making Autotune capabilities available to a limited set of beta testers through a web service and anticipates making it publicly available in September 2015.
“We had to use supercomputing resources to create all the metadata used to train the software, but the Autotune software that will be available to the public doesn’t require all these high-performance resources,” Sanyal said. “We commonly run the software on a laptop.”
For more information, visit www.ornl.gov.
Publication date: 10/13/2014