Use of Mixture-Model Clustering to Inform Tropical Cyclone Track Forecasts

Regression mixture-model clustering provides an objective method to partition tropical cyclone track simulations from one or more ensemble prediction systems. We explore the utility of this clustering method to inform track forecasts by discriminating between high- and low-skill clusters based on cluster population and composition. We also demonstrate how “pruning” the ensemble by excluding all members of clusters below a threshold size yields smaller ensemble-mean errors across the set of forecasts.

Track forecasts out to 144 hours are obtained from the ECMWF IFS (51 members), NCEP GEFS (21 members), UKMET MOGREPS-G (36 members) and Environment Canada GEPS (21 members), forming a multi-model ensemble from 153 initialization times during 2017-2018. Forecasts used in this study include storms in the North Atlantic, Eastern Pacific, Western North Pacific, South Pacific, and South Indian basins. Results are primarily evaluated using 5-cluster solutions, with the latitude and longitude time-evolution of each cluster independently described by 2nd order polynomials.

A modest inverse linear correlation (R²=0.27) is observed between cluster size and mean 96-144-hour cluster error across all forecasts. This correlation increases to 0.43 when an exponential fit for cluster size and error is used, highlighting the contribution of very large errors produced by small clusters. For 58% of forecasts, the smallest cluster produces the largest error, whereas for 38% of forecasts, the largest cluster produces the smallest error. The mean 96-144-hour error of the largest cluster outperforms the 4-EPS ensemble mean for 48% of forecasts, increasing to 54% for forecasts in which the largest cluster equals or exceeds 30% of the total ensemble population. Cluster membership also provides a useful discriminator between high- and low-skill clusters, with clusters that contain substantial membership from each of the ECMWF, UKMET, and NCEP ensembles (but not GEPS) greatly outperforming all other clusters.

Pruning ensembles by cluster population shows promise in reducing ensemble-mean forecast error. For 5-cluster solutions, excluding members of clusters with populations less than 12.5% of the total ensemble impacts 76% of forecasts; of those forecasts, pruning reduces mean 96-144 hour forecast error by 6.2% (17.0 km). For 3-cluster solutions, excluding members of clusters with populations less than 20% of the total ensemble impacts 38% of forecasts, reducing mean error by 9.8% (27.7 km). Meaningful improvements are also observed when clustering and pruning are performed using only the three more skillful ensembles (ECMWF, UKMET, and NCEP). These results indicate that mixture-model clustering is a useful technique for improving multi-ensemble-mean tropical cyclone track forecasts.