Interpolating wind pressure time-histories around a tall building - A deep learning-based approach

Abstract

Machine learning research on estimating wind pressure on tall buildings has primarily focused on mean pressure predictions with limited studies on time history interpolations. In this study, a Deep Neural Network model (DNN) and Extreme Gradient Boost (XGB) were employed to interpolate the wind pressure time histories around the Commonwealth Aeronautical Advisory Research Council (CAARC) standard tall building for four wind directions. The results of a wind tunnel experiment conducted on a CAARC tall building model (1:300) were used to validate the Computational Fluid Dynamics (CFD) models. The pressure data extracted from the CFD model was used to train the DNN and XGB models. The results demonstrated that both XGB (R² = 93%) and DNN (R² = 96%) accurately modelled the wind pressure time histories around the CAARC building. Both models implicitly reconstructed flow features (e.g. pressure gradients, flow separation and conical vortex formations) on the building and compared well with the CFD results. Furthermore, the time-averaged pressure quantities obtained from machine learning models, and CFD models presented good agreement with wind tunnel results. The study shows that the DNN approach is a time-efficient and accurate complementary tool for interpolating wind pressure time histories on isolated tall buildings.