Neural Network Model for Predicting Wind Loads on Isolated Buildings

This study introduces a novel Localized Wind Pressure (LWP) prediction model, a neural network approach for predicting wind pressure distribution on building facades. We address the challenge of accurately predicting wind-induced pressure coefficients across building surfaces by leveraging a limited experimental wind tunnel dataset. The model employs a normalized facade-based coordinate system to process pressure measurements from a cubic building, combining Gaussian process regression for high-resolution pressure mapping with a multilayer perceptron neural network that captures nonlinear relationships between wind angles and pressure patterns. Through rigorous cross-validation, we demonstrate the model's ability to accurately reproduce key aerodynamic phenomena including flow collision, parallel flow, transition, and separation, while maintaining smooth reconstructions. Our approach generates physically consistent pressure distributions even for wind directions absent from training data, offering an efficient alternative to extensive wind tunnel testing or computational fluid dynamics. The results show significant improvements in prediction accuracy and generalization capability, advancing building design practices in civil engineering.