Clustering and Selection of Hurricane Wind Records Using Autoencoder and k-Means Algorithm

AbstractIn wind engineering, to accurately estimate the nonlinear dynamic response of structures while considering uncertainties of hurricanes, a suite of wind records representing the hurricane hazards of a given location is of great interest. Such a suite generally consists of a large number of hurricane wind records, which may lead to highly computational cost for structural analysis. To reduce the computational demand while still preserving the accuracy of the uncertainty quantification process, this paper proposes a machine learning approach to select a representative subset of all collected hurricane wind records for a location. First, hurricane wind records, which are expressed as time series with information that includes both wind speed and direction, are collected from a synthetic hurricane catalog. The high dimensional hurricane wind records are then compressed into a set of low dimensional latent feature vectors using an artificial neural network, designated as an autoencoder. The latent feature vectors represent the important patterns of wind records such as duration, magnitude, and the changing of wind speeds and directions over time. The wind records are then clustered by applying the k-means algorithm on the latent features, and a subset of records is selected from each cluster. The wind records selected from each cluster are those whose latent feature points are closest to the centroid of all latent feature points in that cluster. In order to do regional analysis while taking into account that the hurricane wind records are site-specific, this paper suggests that a region can be discretized into a set of grids, with the proposed hurricane selection approach applied to each grid. This procedure is demonstrated using Massachusetts as a testbed.