Cooperative Institute for Research in the Atmosphere

Use of mixture-model track clustering to interpret tropical cyclone ensemble forecasts

Ensemble forecasts of tropical cyclones provide a wealth of data for operational forecasters and researchers, but fully utilizing the data remains a challenge. Regression mixture-model clustering is demonstrated as a method to partition tropical cyclone ensemble forecasts into a small number of groups based on track. Clustering facilitates the exploration of storm evolution and hazards (e.g. storm surge) associated with each track grouping.

Mixture-model clustering is first applied to partition an ensemble of 72 simulations of Hurricane Sandy (2012) into six clusters based on storm track. After clustering, the structural evolution of Sandy is examined in the four most populous clusters. Sandy undergoes a warm seclusion extratropical transition in each analyzed cluster, with extratropical transition timing the clearest difference between clusters. Inter-cluster differences are smaller, but still relevant, when extratropical transition is analyzed relative to the landfall time of each simulation.

Track clustering is next applied to simulations of Hurricane Irma (2017) to study storm surge hazard. Each of 51 WRF ensemble members, initialized five and two days before Irma’s Florida landfalls, is used to drive a corresponding ADCIRC ocean simulation.  Irma’s tracks in the WRF simulations are then partitioned into clusters; inundation volume and inundation probability from ADCIRC are examined for each cluster. The inundation results among clusters show how track clustering can augment probabilistic hazard forecasts by elucidating hazard scenarios and variability across a dynamical ensemble.