An automated classification method of thunderstorm and non-thunderstorm wind data based on a convolutional neural network.

Historical wind data analysis is a key part of estimating design wind loads. Current design standards do not separately consider the wind loading effects by different wind hazard types. One reason for this lack of consideration is that the separation between thunderstorm and non-thunderstorm wind data is still an issue. A previous study about the Automated Surface Observing System (ASOS) provided a classification method of wind data as thunderstorm or non-thunderstorm based on thunderstorm 'flags' (Lombardo et al., 2009). However, this method relies mainly on manual or automated weather observations which are limited to a subset of stations worldwide. This paper first develops a revised wind hazard type recognition method based on a neural network. In this method, the historical wind data recorded is segmented in different time domains to be applied in a one-dimensional convolutional neural network (1D-CNN) for an automated thunderstorm (T) or non-thunderstorm (NT) classification. Also, based on the trained 1D-CNN, a more comprehensive wind database can be extracted. The classification result from ASOS can automatically provide different peak wind speed for different wind hazard types. • One-minute wind time history data from ASOS DSI-6405 can be extracted as the data source for a deep learning algorithm. • The K-fold validation on the 1D-CNN with the ASOS database for thunderstorm classification is reliable. • The trained ​1D-CNN can be applied to other wind datasets without the requirement for a specific signal sample format. [ABSTRACT FROM AUTHOR]