Sparse-measurement-based peak wind pressure evaluation by super-resolution convolutional neural networks

Abstract

An accurate estimation of the extreme wind pressure acting on glazed panels is essential for the wind-resistant design of buildings. The traditional time-length-velocity (TVL) approach is severely dependent on the determination of the TVL factors and the selection of single-point measurement tap. To overcome the difficulties, this study aims to generate a generalized model based on convolutional neural network (CNN) to reconstruct the super-resolution pressure distributions from low-resolution pressure measurements at different wind directions. The constraint represented by the pressure gradient are embedded in the loss function of the CNN to enable the model to generate real pressure distribution characteristics. By spatially averaging the super-resolution distributions, the peak space-averaged pressure on the glazed panels could be finally predicted. On the basis of a rectangle-section high-rise building, the regions near the top corner and near the middle-height under different wind directions are focused on. Results show that the pressure peaks computed from the super-resolution distributions can provide more accurate and robust representations for the true area-averaged peaks, compared with the traditional TVL method. The generalization ability of the pressure gradient guided SRCNN is systematically investigated among different pressure modes and flow patterns influenced by the incident angles.