Surrogate optimisation holds a big promise for building energy optimisation studies due to its goal to replace the use of lengthy building energy simulations within an optimisation step with expendable local surrogate models that can quickly predict simulation results. To be useful for such purpose, it should be possible to quickly train precise surrogate models from a small number of simulation results (10-100) obtained from appropriately sampled points in the desired part of the design space. Two sampling methods and two machine learning models are compared here. Latin hypercube sampling (LHS), widely accepted in building energy community, is compared to an exploratory Monte Carlo-based sequential design method mc-intersite-proj-th (MIPT). Artificial neural networks (ANN), also widely accepted in building energy community, are compared to gradient-boosted tree ensembles (XGBoost), model of choice in many machine learning competitions. In order to get a better understanding of the behaviour of these two sampling methods and two machine learning models, we compare their predictions against a large set of generated synthetic data. For this purpose, a simple case study of an office cell model with a single window and a fixed overhang, whose main input parameters are overhang depth and height, while climate type, presence of obstacles, orientation and heating and cooling set points are additional input parameters, was extensively simulated with EnergyPlus, to form a large underlying dataset of 729,000 simulation results. Expendable local surrogate models for predicting simulated heating, cooling and lighting loads and equivalent primary energy needs of the office cell were trained using both LHS and MIPT and both ANN and XGBoost for several main hyperparameter choices. Results show that XGBoost models are more precise than ANN models, and that for both machine learning models, the use of MIPT sampling leads to more precise surrogates than LHS.
Read full abstract