Early global model forecasts of hurricane Sandy differed substantially in their track predictions. European Centre for Medium-Range Weather Forecasting (ECMWF) deterministic model and ensembles predicted the eventual left turn more accurately than the NOAA/NWS Global Forecast System (GFS) and its ensembles. However, both ensemble prediction systems (EPSs) contained members in which Sandy turned westward, and members in which Sandy tracked eastward out to sea. These EPSs and others also contained forecasts in which Sandy moved farther westward and more rapidly northward than observed, and forecasts in which Sandy turned westward more slowly than observed.
To gain a better understanding of the differing meteorological conditions that led some ensemble members to correctly predict the system's turn to the left into the Mid-Atlantic coast, while other members forecasted Hurricane Sandy to move out to sea, we begin by partitioning the ensemble forecasts via a regression mixture-model path clustering of the track forecasts (Gaffney et al. 2007). This work builds on previous studies that analyzed causes of track differences in ensemble forecasts of hurricane Sandy (Magnusson et al. 2013; Bassill 2014; Munsell and Zhang 2014; Torn et al. 2015). However, this study differs by using an objective regression mixture model to partition forecast tracks and by including forecasts from multiple EPSs.
For this study, we analyze forecasts from 00 UTC October 25, when Sandy was located south of Cuba. A total of 117 forecasts (113 perturbed ensembles and four control forecasts) from four EPSs (ECMWF, GEFS, Canadian GEM, and United Kingdom UKMO) are clustered out to seven days. Development of the mixture model leads to selection of four clusters as the optimum partition.
Forecasts from 00 UTC 25 October produced by the 117 ensemble members show widely varying tracks (Fig. 1a). Clustering of track forecasts produces a coherent division of these tracks (Fig. 1b). In the westernmost (red) cluster (12 members) Sandy moves rapidly up the Atlantic seaboard, west of the observed track. The dominant blue cluster (61 members) follows a similar path to the observed track, but turns left slightly less sharply. The magenta cluster (31 members), takes a similar path as the blue cluster during the first 72 hours. However, Sandy moves farther eastward and turns westward more gradually, making landfall farther north. In the easternmost (green) cluster (13 members) Sandy moves east-northeastward out to sea.
Figure 1: (a) Forecasts from the 117-member multi-model ensemble and (b) the mean track of each cluster. Positions of Sandy at 00 UTC October 30 are shown as dots.
Among the four clusters, the red cluster diverges most from the other three early in the forecast. The red cluster is associated with higher 500 hPa and 300 hPa geopotential heights north of Sandy early in the forecast period (Fig 2, a, b). These higher geopotential heights generate an anomalously easterly perturbation in the steering flow, which steers Sandy farther westward during the first 24-36 hours compared to the other clusters. The blue cluster also shows amplified 500 hPa (but not 300 hPa) ridging compared to the magenta and green clusters within 12-18 hours of initialization, while the magenta cluster shows amplified 500 hPa and 300 hPa ridging compared to the green cluster. A stronger ridge is not observed north of the blue cluster compared to the magenta cluster, consistent with the two clusters' similar paths early in the forecast.
Greater upper-troposphere divergence northwest of Sandy is found in western clusters compared to eastern clusters. Greater lower troposphere moisture north of Sandy and mid-troposphere moisture northwest of Sandy is also observed in the red cluster compared to other clusters within 0-6 hours of initialization. This provides evidence that larger quantities of moisture and upper-troposphere divergence in the farther-west clusters contributed to greater negative potential vorticity (PV) advection and ridge amplification.
Figure 2: 12-hour cluster mean forecast differences between the red and the blue cluster for (a) 300 hPa and (b) 500 hPa geopotential height. Units are in meters.
Torn et al. (2015) identified that western members of an experimental GEFS ensemble had greater negative PV advection north of Sandy compared to the easternmost members. This negative PV advection enhanced the ridge north of Sandy, causing Sandy move more westward. They attributed the greater negative PV advection to greater upper-troposphere divergence and mid and lower-troposphere moisture north of Sandy. Here, we find similar qualitative results using a multi-model ensemble and four clusters. The 250 hPa PV difference field between each cluster pair show that the ridge north of Sandy built farther to the west in clusters in which Sandy moved more westward (Fig. 3). Within 12-18 hours of initialization, a PV dipole showing ridge extent is observed over the eastern Gulf of Mexico between each pair of clusters except blue vs. magenta. Thus, path clustering out to seven days captures differences in ridge amplification that affected atmospheric steering during the first 24-36 hours of the forecast.
Figure 3: 18-hour cluster-mean forecast differences in 250 hPa potential vorticity between clusters: (a) red and blue, (b) red and magenta, (c) red and green, (d) blue and magenta, (e) blue and green, and (f) magenta and green. Units are in s-1x106.
In summary, regression mixture-model clustering of hurricane Sandy track forecasts from a multi-model ensemble highlights substantial synoptic differences among ensemble members, even early in the forecast period. Ensemble members in which Hurricane Sandy moved rapidly up the Atlantic coast show the greatest differences from other clusters during the first twenty-four hours. The smallest synoptic differences are observed between forecasts in which Sandy made landfall in the Mid-Atlantic coast and those in which Sandy made landfall farther north, as these clusters did not diverge substantially until later in the forecast. These results suggest the potential operational utility of regression mixture-model path clustering for developing physically meaningful partitions of EPSs and for diagnosing the underlying meteorology leading to different tropical cyclone track forecasts. These meteorological insights may be useful to the forecaster in determining which scenario is more likely.