Prediction of wind load power spectrum on high-rise buildings by various machine learning algorithms

To evaluate the suitability of machine learning algorithms in predicting wind effects on high-rise buildings, this study investigates the prediction of wind load power spectrum for the CAARC standard model of high-rise buildings using various machine learning algorithms. The inputs of the machine learning model include turbulence intensity, wind direction and reduced frequency. The outputs consist of the power spectrum for the along-wind base moment coefficients, across-wind base moment coefficients, and torsional base moment coefficients. 53,200 sets of wind load power spectrum data are obtained from wind tunnel tests for training and verifying four machine learning algorithms, including Gradient Boosted Regression Tree, Histogram Gradient Boosted Regression Tree, XGBoost, and Neural Network. The hyperparameters of the machine learning algorithms are then optimized using Tree-structured Parzen Estimator and cross-validation. By comparing the mean square errors of the four machine algorithms on the test set, it is determined that the Gradient Boosted Regression Tree algorithm performs well in predicting the power spectrum of the base moment coefficient for the standard model of high-rise buildings in the along-wind, across-wind, and torsional directions. The correlation coefficients between the predicted values and the experimental values are found to be larger than 0.97. This study demonstrates the feasibility of using machine learning to predict the wind load power spectrum on high-rise buildings. It provides valuable insights for the application of machine learning in the wind-resistant design of high-rise buildings.