Surrogate Modelling on BESOS

Surrogate modelling or metamodelling is a way to emulate physics-based building simulation models [1]. BESOS integrates building simulation and surrogate modelling using: + EnergyPlus model editing and execution via EPPy using Evaluators + sampling toolboxes + machine learning toolboxes like ScikitLearn and TensorFlow + the besos module to simplify the interconnection of the above items.

Surrogate Modelling workflow

The figure below shows a BESOS workflow that covers all the elements needed to derive surrogate models. This is supported by access to powerful hardware for conducting sampling and model training quickly in parallel.

BESOS workflow for Surrogate Modelling [`1 <>`__].

The derivation of surrogate models demonstrated in the following notebooks:

  • Interactive Surrogate is a quick tour of a lots of BESOS: make an EnergyPlus model with two parameters, generate samples that span the design space, use these to train a surrogate model, then explore the design space using an interactive plot that queries the surrogate model.
  • Fit GP Model makes a Gaussian Process surrogate model using latin hypercube sampling and the ScikitLearn syntax.
  • Fit feedforward Neural Network makes a Neural Network surrogate model using latin hypercube sampling and the ScikitLearn syntax.
  • Fit NN Tensorflow makes a Tensorflow graph and trains it on building simulation data.
  • Fit GP Adaptive adaptiveley selects simulations instead of picking them all in one advance. This aims to reduce the number of samples required to derive a surrogate model.


  • Parameter sets is a .py file which defines some sets of besos.parameters to avoid redifining design parameters in each of the notebooks above.