A novel nested BO-LM-BPNN method for wind pressure field prediction of non-isolated low-rise buildings

Abstract

The prediction of wind pressure field of non-isolated low-rise building is a prerequisite for the wind-resistant design of building structures. Due to the powerful performance of physical wind tunnel tests and computational fluid dynamic simulations, they have been widely used to directly obtain and predict the wind pressure field of non-isolated low-rise buildings. However, these two methods are susceptible to the economy and time cost constraints respectively. Moreover, because non-isolated low-rise buildings usually have complex surrounding environments, it is difficult for these two methods to obtain the wind pressure fields under all complex working conditions. In this study, a nested Bayesian optimization (BO)-Levenberg Marquardt (LM)-backpropagation neural network (BPNN) method is proposed. Firstly, based on the measured wind pressure field data under some representative working conditions, the conversion relationship between multi-input and single-output of wind pressure field using BPNN model is derived. Then, the hyperparameters and trainable parameters in BPNN model are optimized by BO and LM algorithms respectively. Finally, a nested input structure is established to further improve the accuracy of high-order moments and peak values of wind pressure field. Numerical results show the proposed method has high accuracy in the prediction of the wind pressure field of non-isolated low-rise buildings. The nested input structure can further improve the prediction accuracy of high-order moments and peak values in hazardous areas of roof. Therefore, the research results can serve as a reference for studies on wind pressure fields and wind loads of non-isolated low-rise buildings.