Detection of thunderstorms is important to the wind hazard community to better understand extreme winds field characteristics and associated wind induced load effects on structures. This paper contributes to this effort by proposing a new course of research that uses machine learning techniques, independent of wind statistics based parameters, to autonomously identify and separate thunderstorms from large databases containing high frequency sampled continuous wind speed measurements. In this context, the use of Shapelet transform is proposed to identify key individual attributes distinctive to extreme wind events based on similarity of shape of their time series. This novel shape based representation when combined with machine learning algorithms yields a practical event detection procedure with minimal domain expertise. In this paper, the shapelet transform along with Random Forest classifier is employed for the identification of thunderstorms from 1 year of data from 14 ultrasonic anemometers that are a part of an extensive in situ wind monitoring network in the Northern Mediterranean ports. A collective total of 235 non-stationary records associated with thunderstorms were identified using this method. The results lead to enhancing the pool of thunderstorm data for more comprehensive understanding of a wide variety of thunderstorms that have not been previously detected using conventional gust factor-based methods.