A Gaussian Process-Based emulator for modeling pedestrian-level wind field
A.U. Weerasuriya, Xuelin Zhang, Bin Lu, K.T. Tse, C.H. Liu
Abstract
Wind tunnel tests and computational fluid dynamics (CFD) simulations remain the main modeling techniques in wind engineering despite being expensive, time-consuming, and requiring special facilities and expert knowledge. There is a clear need for a fast, accurate, but, at the same time, computationally economical substitute. This study proposes a Gaussian Process-based (GP-based) emulator to predict the pedestrian-level wind environment near a lift-up building – an isolated, unconventionally configured building. The proposed GP-based emulator transcends the limitations of previous emulators as it can handle many inputs (8) and output parameters (384) and a large dataset (150 CFD simulations). To increase computational efficiency, the current study proposes a data reduction method based on Principal Component Analysis (PCA) and a technique to estimate hyper-parameters based on optimization. The latter can efficiently compute 250 hyper-parameters and requires no prior knowledge of their probability distributions. The emulator is faster, by a factor of 107 than CFD simulations in predicting wind speeds, and its accuracy is substantiated using both qualitative and quantitative analyses, which reveal that the emulator's predictions of all-prominent flow features near a building have no systematic bias, are highly accurate, and have great reproductivity.