Gaussian-Process-Based Emulators for Building Performance Simulation

In this paper, we present an emulator of a buildingenergy performance simulator. Previous work on emulators for this application has largely focused on linear models. Since the simulator itself is a collection of differential equations, we expect non-linear models to be better emulators than linear models. The emulator we present in this paper is based on Gaussian-process (GP) regression models. We show that the proposed non-linear model is 3-4 times more accurate than linear models in predicting the energy outputs of the simulator. For energy outputs in the range 10-800 kWh/m, our model achieves an average error of 1025 kWh/m compared to an average error of 30-100 kWh/m from using linear models. In addition to being very accurate, our emulator also heavily reduces the computational burden for building designers who rely on simulators. By providing performance feedback for building designs very quickly (in just a few milliseconds), we expect our approach to be particularly useful for exercises that involve a large number of simulations, e.g., Uncertainty Analysis (UA), Sensitivity Analysis (SA), robust design, and optimisation.