Application of machine learning methods to predict aerodynamic pressure coefficients on rectangular buildings and structures

Introduction. Wind effects are one of the key factors in the design of buildings and structures. Normative calculations, physical and numerical modelling, as well as in-situ measurements have a number of limitations in application. The use of machine learning (ML) technologies opens up new opportunities for rapid and accurate prediction of wind loads. The application of ML models to assess the distribution of aerodynamic pressure coefficients on rectangular buildings is considered, which allows not only to calculate integral characteristics (forces, moments), but also to analyze in detail the distribution of loads on facades.Materials and methods. For model training, the Tokyo Polytechnic University database was used, which presents the results of wind tunnel tests on building models of various heights and widths. Data augmentation was performed, which increased the original example size and increased the ability of ML models to generalize various geometric configurations. During feature processing, different angles of wind attack were taken into account, and the correlation of features was analyzed in order to eliminate multicollinearity. Linear regression, decision tree and gradient boosting (CatBoost) were the main prediction methods.Results. The calculations showed that the best balance between the accuracy of predictions and maintaining physical interpretability was provided by gradient boosting over the decision tree (CatBoost), reducing the average weighted error to 16–18 %. In addition, a comparison was made with the results of aerodynamic tests, which confirmed the adequacy of the proposed approach.Conclusions. The application of machine learning methods, in particular gradient boosting, makes it possible to reliably predict aerodynamic pressure coefficients on various dimensional shapes of buildings at a wide range of wind attack angles. The obtained results demonstrate the promising use of ML models to accelerate and reduce the cost of wind impact assessment stages.

wind load, machine learning (ML), data augmentation, decision tree, gradient boosting, aerodynamic coefficient, buildings and structures