Multi-model Ensemble Track Clustering to Improve Tropical Cyclone Forecasting

Regression mixture-model clustering is explored as a method to improve tropical cyclone track forecasts. Track forecasts are obtained out to 120 hours for 327 initialization times from the Atlantic, East Pacific, Central Pacific, and West Pacific during 2014 and 2015. Multi-model ensembles comprising combinations of three global ensemble prediction systems (ECMWF, GEFS, and UKMO) are clustered using regression mixture models with 2-7 clusters and 3rd-5th order polynomials. Error statistics of the most populous cluster of each clustering solution are compared to error statistics of the ECMWF + GEFS ensemble mean, which was found to have the smallest average error. Cluster track errors at 96-120-hour lead times are also analyzed a function of errors at 6-12 hours to determine whether clusters with smaller errors at small lead times also have smaller errors at large lead times. Clustering the ECMWF + GEFS ensemble combination with five clusters and a fifth-order polynomial produces a most populous cluster with the smallest mean error across all forecasts. This mean error is smaller than the ensemble mean error for the most poorly-performing ensemble combinations; however, it has a larger mean error than the best-performing ensemble combinations. Clusters with smaller errors at 6-12 hours are found to produce smaller errors at 96-120 hours. These results indicate that regression mixture-model clustering may prove a useful tool for interpreting ensemble track forecasts; however, further refinement is needed for clustering to be utilized as an operational forecast product.