Question-and-Answer Resource for the Building Energy Modeling Community
Get s tarted with the Help page
Ask Your Question

Revision history [back]

OpenStudio has developed a cloud based platform to solve calibration problems. While Autotune is not included (nor are there plans to) there are several optimization algorithms available thru the OpenStudio Analysis Spreadsheet and will soon be available in PAT. Currently the list is:

NSGA2 (Non-dominated Sorting Genetic Algorithm)

  • Multi-objective
  • Parallel F evaluations
  • Mixed Continuous or Discrete variables

SPEA2 (Strength Pareto Evolutionary Algorithm)

  • Multi-objective
  • Parallel F evaluations
  • Continuous variables only

Rgenoud (GENetic Optimized Using Derivatives)

  • Single-objective
  • Parallel F evaluations for genetic search
  • Parallel gradient calculation for continuous variables

Optim (quasi-Newton method with bounds)

  • Single-objective
  • Parallel gradient calculation
  • Continuous variables only

PSO (Particle Swarm)

  • Single-objective
  • Continuous variables only

The calibration workflow is Measure based, meaning almost any OpenStudio method is available for calibration purposes (ex, swapping out full HVAC systems, changing geometry, or simple idf parameter changes). A good Measure resource is https://bcl.nrel.gov/

The OpenStudio calibration / optimization workflow accepts timeseries and monthly utility data. Objective functions can also be put into groups for multi-objective problems. The suggested workflow is to first sample the variable space with one of the sampling algorithms available on the server (ex, LHS). Remove any non-impactful variables and choose appropriate bounds for the impactful variables. Run a calibration using a gradient based method (ex, Optim) for all continuous variables, a hybrid genetic / gradient based method (Rgenound) or choose a multi-objective algorithm such as NSGA2. Once that is completed the robustness of the solution can be checked by using one of the sensitivity methods.

Computing time is a function of base simulation time and the number of variables that are selected for calibration. All the available algorithms are parallelized meaning gradient calcs are done in parallel for Optim and Rgenoud and all population calcs for NSGA2 are run utilizing all available cores. Costs are always changing but right now you can get a 32 core box for $1.68/hr with the capability to add more boxes for larger problems.

Several papers were presented at ASHRAE:

  • Ball, 2015, Calibration example with OpenStudio.
  • Macumber, 2014, A graphical tool for cloud-based building energy simulation.
  • Long, 2014, Scaling building energy modeling horizontally in the cloud with openstudio.

There is also a calibration example in the github site for the spreadsheet which can be found here