Wind pressures on roofs of nonrectangular buildings: Experimental and machine learning approaches

Wind effects on buildings with rectangular plans have been investigated widely by wind engineering researchers either in wind tunnels or through CFD simulations. These studies provided comprehensive overviews and detailed descriptions of wind pressures on rectangular buildings and created the basic source required to formulate the wind provisions in the national and international codes and standards. However, buildings with irregular (i.e., non-rectangular) plans have not received adequate attention from wind tunnel investigations. Therefore, wind loads on irregular buildings are described shortly and shyly, if at all, in the current building codes and standards. This paper describes the experimental investigations into the flat-roof pressures of buildings with four non-rectangular shapes —L, U, T, and X— in an atmospheric boundary layer wind tunnel. The results reveal that the distribution of wind loads on the outer roof corners and edges of buildings with non-rectangular plans resembles that experienced by a rectangular building. However, the wind loads on the inner perimeter area, particularly the inner edge of the U shaped building, were observed to be generally higher than those recorded on the edges of a typical rectangular roof. Furthermore, the wind tunnel measurements not only provided valuable data but also served as a dataset when applying Machine Learning (ML) as a tool to predict wind loads on irregular buildings. This involved the utilization of a Gradient Boosting Regressor (GBR) and Artificial Neural Networks (ANN), using two data split approaches: random and structured splits. The ML models exhibit significant predictive accuracy, achieving minimal Mean Squared Error (MSE) and coefficients of determination (R-squared) of about 0.97 for wind pressure coefficients. Further, the study demonstrated that a structured split of the dataset reflects a more realistic assessment of the ML models.