Analysis of building wind pressure using proper orthogonal decomposition, autoregressive moving average and neural networks

The proper orthogonal decomposition (POD), autoregressive moving average (ARMA) models, and neural networks are used in the analysis of wind-induced pressure on buildings.;The POD offers a unique way of representing a random field such as wind pressure. The pressure space covariance is employed to calculate the modes and the principal coordinates of the considered data. These quantities are then applied in a selective reconstruction of the pressure.;A procedure for the POD of pressure specified at non-uniformly spaced locations is described in detail. The numerical integration called for by the POD is performed using the rectangular rule and Lagrange's formulation. The POD results, the eigenvalues and eigenfunctions (modes), for the pressure at uniformly distributed taps are compared with those for the pressure at non-uniformly distributed taps.;The ARMA is used to model both the wind-induced pressure and the POD principal coordinates. Several model selection criteria are employed in the optimization of the model order. The model order is used in comparison of complexity of the pressure and the POD principal coordinates.;The wind-induced pressure are used to train neural networks and to evaluate the performance of the developed neural network. Several configurations of the backpropagation neural network are tested by varying the number of nodes in the input layer and the number of neurons in the hidden layer. The effects of the number of hidden layers as well as the sampling frequency of the time series on the performance of the neural network are also investigated.;The POD is found to be useful in compression of large sets of wind pressure data, while preserving pertinent features of the considered pressure field. The AR models established in this study can be used to generate wind-induced pressure on buildings and structures. The accuracy of the one-step prediction of the wind-induced pressure using the neural network is found to be compatible with that of the AR model.