CORE-2012: Autotune

You are here

Lead Performer:

  • Oak Ridge National Lab (ORNL) – Oak Ridge, TN


  • Jacksonville State University – Jacksonville, AL
  • University of Tennessee, Knoxville – Knoxville, TN
  • Karpay Associations – Potomac, MD


DOE Funding: $175,000 in FY15; $1,014,000 to date

Cost Share: Over 2 million hours of simulation time on ORNL leadership supercomputers

Project Term: 2012 – 2014

Funding Opportunity: Emerging Technologies and Commercial Buildings Integration core funding


A significant challenge in creating and effectively using energy models for existing buildings — for retrofit planning, retro-commissioning, or measurement-and-verification (M&V) of efficiency-measures — is deriving difficult to obtain model inputs like air infiltration rates, actual occupancies and plug loads, degraded equipment efficiencies, internal thermal mass, etc. This is done by “calibration” against measured data like monthly utility bills, interval meter or sub-meter data (e.g. GreenButton), zone temperatures, or other sensor data streams. During calibration, uncertain inputs are iteratively varied until simulation outputs are deemed sufficiently close to measurements. Traditional manual calibration is extremely labor intensive and more art and experience than science and process. Although graphical techniques help — OpenStudio uses a parallel dimension plot widget to facilitate manual calibration — the human element places practical limits on the amount of data used to calibrate and the “accuracy” that manual calibration can achieve. This is one of the reasons that traditional calibration standards like ASHRAE Guideline 14 are relatively lax.

The Autotune project aims to replace art with science and expensive human time with cheap computing time. Autotune uses evolutionary computation to calibrate model inputs using any sources of measured data which can map to simulation engine output. An initial population of candidate buildings seeds the first generation by intelligently sampling the parametric space defined for all inputs to be tuned while obeying any mathematical constraints placed on combinations of those parameters (e.g. cooling setpoint should be higher than heating setpoint by 1 degree Celsius). Machine learning agents — trained on large simulation datasets to learn correlations between simulation inputs and outputs — serve as hinting mechanisms to intelligently manipulate the parameters of the most promising members of the population to generate the next generation of candidate input models. This process proceeds iteratively to reduce error until a tuning accuracy or a time limit criterion is reached.

An important aspect of the Autotune project is a Trinity Test framework and web service for quantitatively evaluating any calibration algorithm. The service randomly generates a “secret” building and then uses simulation to generate virtual measured data for that building. Candidate calibration algorithms (manual, semi-automatic, or fully automated) can receive this simulated measured data, an example input file, and information on the tuning characteristics (potentially including range, distribution, and mathematical constraints for any parameter in the input file); when complete, a tuned model is submitted to the system. Algorithms are judged not only by how closely they match the simulated outputs, but by how closely the calibrated model’s tuned inputs match the secret model.

The Autotune algorithms have been tested using both this Trinity Test mechanism for all DOE reference buildings, in addition to calibration of real-world facilities including heavily instrumented residential buildings, small commercial building test facilities, and traditional office buildings. Autotune will be available via the OpenStudio platform in mid-2015.


Retrofitting existing buildings is the primary way to achieve significant short-term energy reduction in the building sector and having calibrated models of the buildings is necessary in order to plan and optimize retrofits for cost-effective energy-savings. Unfortunately, energy modeling is employed only in the largest projects due to the significant overhead required to create a calibrated model and even calibrated models often diverge from measurement by 30% and more. Developed and tested using DOE supercomputing resources and highly submetered and instrumented demonstration facilities, Autotune has already demonstrated the ability to outperform manual calibration and exceed ASHRAE Guideline 14 calibration requirements in several minutes.

A generalized automated building energy model calibration methodology would significantly reduce the costs to develop building energy retrofit projects, enhance the cost-effectiveness of retrofit projects, and expand their reach into smaller buildings. It would also significantly improve the state of the art in energy savings measurement and verification (M&V) for performance contracting and other purposes.  Assuming Autotune is successful enough to enhance cost-effectiveness of current retrofit initiatives while expanding reach into smaller buildings such that there is a 1% uptake in the retrofit market, this would amount to a cumulative energy savings of 27.4TBtus/yr according to a Rockefeller Foundation study on energy-efficiency building retrofits. 

Recent Publications and Presentations

  • Sanyal, Jibonananda, New, Joshua R., Edwards, Richard E., and Parker, Lynne E. (2014). "Calibrating Building Energy Models Using Supercomputer Trained Machine Learning Agents." In Journal on Concurrency and Computation: Practice and Experience, March, 2014.
  • Ostrouchov, George, New, Joshua R., Sanyal, Jibonananda, and Patel, Pragnesh (2014). "Uncertainty Analysis of a Heavily Instrumented Building at Different Scales of Simulation." In Proceedings of the 3rd International High Performance Buildings Conference, Purdue, West Lafayette, IN, July 14-17, 2014.
  • Edwards, Richard E. (2013). “Automating Large-Scale Simulation Calibration to Real-World Sensor Data”. A Dissertation presented for the Doctor of Philosophy Degree in the Archives of The University of Tennessee, Knoxville, TN, May 2013. 178 pages [PDF] [PPT]
  • New, Joshua R., Sanyal, Jibonananda, Bhandari, Mahabir S., Shrestha, Som S. (2012). "Autotune E+ Building Energy Models." In Proceedings of the 5th National SimBuild of IBPSA-USA, International Building Performance Simulation Association (IBPSA), August 1-3, 2012. [PDF] [PPT] [IBPSA]
  • Sanyal, Jibonananda, Al-Wadei, Yusof H., Bhandari, Mahabir S., Shrestha, Som S., Karpay, Buzz, Garret, Aaron L., Edwards, Richard E., Parker, Lynne E., and New, Joshua R. (2012). "Autotune: Building Energy Model Calibration using EnergyPlus, Machine Learning, and Supercomputing." In Proceedings of the 5th National SimBuild of IBPSA-USA, International Building Performance Simulation Association (IBPSA), August 1-3, 2012. [Poster]
  • Edwards, Richard E., New, Joshua R., and Parker, Lynne E. (2012). "Predicting Future Hourly Residential Electrical Consumption: A Machine Learning Case Study." In Journal of Energy and Buildings, volume 49, issue 0, pp. 591-603, June 2012. [PDF]
  • Garrett, Aaron, New, Joshua R., and Chandler, Theodore (2013). “Evolutionary Tuning of Building Models to Monthly Electrical Consumption.” Technical paper DE-13-008. In Proceedings of the ASHRAE Annual Conference, Denver, CO, June 22-26, 2013. [PDF] [PPT]
  • Garret, Aaron and New, Joshua R. (2013). "Trinity Test: Effectiveness of Automatic Tuning for Commercial Building Models." ORNL internal report ORNL/TM-2013/130, March 7, 2013, 24 pages.
  • New, Joshua R. (2013). "Autotune Building Energy Models." DOE Building Technology Office (BTO) Peer Review, Washington DC, April 2, 2013. [PPT]