Prediction of nonlinear structural response under wind loads using deep learning techniques

Abstract

The wind actions on buildings are continuous in nature and could range from a few minutes to more than hours. Such long duration high intensity winds can push the structure to enter post elastic structural range and cause nonlinear behavior. Therefore, unlike performance-based earthquake engineering, where the structural simulations require seismic loads acting for a few seconds, the simulations for performance-based wind engineering (PBWE) requires wind load models acting for much longer durations. However, the nonlinear 3D model of a tall building will contain thousands of connections and structural members making it a complex finite element model. Dynamic time history analysis of such a model under loads lasting for hours can be tedious and often fail to achieve convergence. Using data-driven techniques that utilizes limited numerical and field data to obtain accurate structural responses under long duration loads is an exciting alternative. Deep learning techniques have been extensively used in the studies for structural health monitoring and earthquake engineering. However, the implementation of such data-driven techniques is very limited and has the potential for exploration in problems related to wind dynamics on tall buildings. This paper aims to predict the nonlinear structural response of tall buildings under sustained durations of wind loads using deep learning models. A Long Short-term Memory (LSTM) architecture is used to assess the efficiency of data-driven methods to replace computationally intensive 3D finite element analyses. The architecture will be tested on a 150 m tall building for response predictions under long duration wind loads. The robustness of the architecture will be further evaluated with predicting the acceleration response history of a scaled aeroelastic model based on experimental studies conducted at the Wind Simulation and Testing Laboratory (WiST) at Iowa State University.