Abstract/Summary

With the development of ensemble forecasting, operational meteorologists are faced with large amounts of constantly updating complex information which they must quickly interpret to issue forecasts and warnings. In this thesis a novel clustering technique is introduced that reduces ensemble forecasts to a few representative forecast trajectories. Clustering is performed using k-medoids with the distance metric defined by the Fractions Skill Score (FSS) of the gradient in 850hPa wet-bulb potential temperature to group ensemble members with similar frontal features. The number of clusters is selected using lead-time-coherence of the clusters over a window of interest when clustering is most distinct. Members nearest to the centre of each cluster during this window of interest are chosen as representative members to be viewed by forecasters. Clustering is found to be more coherent during low predictability events when ensemble spread is large. The clustering method was compared to an alternative that uses the FSS of large-scale rain rate and it was found that while similar, results are not interchangeable. The gradient of wet-bulb potential temperature had higher time-coherence and therefore was judged preferable. The method was evaluated during the Met Office winter testbed of 2021-22, and representative members found were found to correspond well to forecasters judgement of the distinct scenarios in the ensemble, hence providing a useful reduction in the data that needs to be considered in issuing forecasts. The method draws attention to low predictability events that appear across several forecasts. While this method has been created to fill a need with ensemble forecasting, it is anticipated that it can be used in many other areas of research such as identifying circulation patterns, seasonal and climate forecast trajectories, and exploring different meteorological phenomena by modifying variable choice and other parameters. The method is also planned for use at the Met Office.