Assessment of approaching wind field for high-rise buildings based on wind pressure records via machine learning techniques

Many accidents of buildings and bridges in service suggest it is essentially important for wind-sensitive structures to timely clarify approaching wind field so as to better understand associated wind effects and then adopt appropriate wind-resistant measures. However, difficulties usually exist for directly measuring undisturbed wind flows on or around large-scale structures. This paper presents a study on the assessment of approaching wind field for a supertall building via Machine Learning (ML) methods based on wind pressure records. A total of 8 specific ML models, i.e., AdaBoost, Bagging, Decision Tree, Extra Tree, Gradient Boosting, KNN, Random Forest, and SVR, are exploited to deduce associated wind information (in terms of mean values of speed and direction and turbulence intensity of speed) from the time series or statistical parameters of wind pressure. The models are trained and validated by results from wind tunnel experiments, and tested by field measurements from the prototype building. It is shown that most of the models driven by "statistical parameters" rather than "time series" are able to map wind pressure information to that for approaching wind field accurately, which reflects the importance of selecting appropriate input variable to driven ML models. The robustness of these models is further examined through a series of tests with reduced training/validation data. The presented results are helpful to further understand the correlations between the statistical features of wind pressure on high-rise buildings and corresponding wind field. The documented ML techniques can also be adopted to determine the approaching wind field for other wind-sensitive structures on the basis of various kinds of wind-effect records.