Comparative study of metamodelling techniques in building energy simulation: Guidelines for practitioners

Computer simulation of real system behaviour is increasingly used in research and development. As simulation models become more reliable, they also often become more complex to capture the progressive complexity of the real system. Calculation time can be a limiting factor for using simulation models in optimisation studies, for example, which generally require multiple simulations. Instead of using these timeconsuming simulation models, the use of metamodels can be considered. A metamodel approximates the original simulation model with high confidence via a simplified mathematical model. A series of simulations then only takes a fraction of the original simulation time, hence allowing significant computational savings. In this paper, a strategy that is both reliable and time-efficient is provided in order to guide users in their metamodelling problems. Furthermore, polynomial regression (PR), multivariate adaptive regression splines (MARS), kriging (KR), radial basis function networks (RBF), and neural networks (NN) are compared on a building energy simulation problem. We find that for the outputs of this example and based on Root Mean Squared Error (RMSE), coefficient of determination (R), and Maximal Absolute Error (MAE), KR and NN are the overall best techniques. Although MARS perform slightly worse than KR and NN, it is preferred because of its simplicity. For different applications, other techniques might be optimal.