EVOLUTION CHARACTERISTICS OF ENSEMBLE FORECASTS THROUGH CLUSTERING

Open Access
- Author:
- Kuruppumullage Don, Prabhani
- Graduate Program:
- Statistics
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- None
- Committee Members:
- Francesca Chiaromonte, Thesis Advisor/Co-Advisor
Francesca Chiaromonte, Thesis Advisor/Co-Advisor - Keywords:
- ensemble forecasts
path clustering
Model-based clustering
cyclones - Abstract:
- Tropical and extra-tropical cyclones pose considerable societal risks, especially to the increasing populations living near the coasts. As a result, they have been studied since at least the time of Kublai Khan. The use of operational numerical models and of ensemble techniques has greatly improved accuracy in forecasting the evolution of these storms. In this study, we apply model based point and path clustering methodologies to the European Centre for Medium-Range Weather Forecasts (ECMWF) 51-member ensemble forecasts of Typhoon Sinlaku (2008) with two aims: (i) to investigate the structural evolution of the storm, and (ii) to explore any structure in the variation of the ensemble forecasts that might aid in prediction. For aim (i), we cluster points representing Typhoon Sinlakus progression in the 3-dimensional CPS space (spanned by lower-tropospheric thermal winds, upper-tropospheric thermal winds, and lower-tropospheric thermal asymmetry), which is used to define the evolving structure of a storm. Traditional and re-sampling based diagnostics indicate that a mixture of 5 spherical Gaussian components provides a good fit to the data. Comparison with well-established storm structures suggests that the resulting clustering solution successfully captures Typhoon Sinlakus structural evolution of over time. For aim (ii) we cluster entire ensemble forecast paths for Typhoon Sinlaku, both in physical space, i.e. the 2-dimensional space spanned by latitude and longitude, and in the 2-dimensional subspace of the CPS spanned by lower-tropospheric thermal winds and lower- tropospheric thermal asymmetry. A regression mixture framework is used for path clustering, with traditional and re-sampling based diagnostics pointing to a mixture of 5 cubic polynomial components. Path clustering in physical space creates a meaningful partition of ensemble members related to the subsequent evolution of the storm, and appears to identify subsets of ensemble members that best approximate the actual track of the storm. Interpretations are less straightforward but still informative for path clustering in the CPS space.