Regression mixture model clustering of multimodel ensemble forecasts of hurricane sandy: Partition characteristics

Track and cyclone phase space (CPS) forecasts of Hurricane Sandy from four global ensemble prediction systems are clustered using regression mixture models. Bayesian information criterion, cluster assignment strength, and mean-squared forecast error are used to select optimal model specifications. Fourth-order (third order) polynomials for 168-h forecasts (60-h forecast segments) and 5 (6) clusters for track (CPS) forecasts are selected. Mean cluster paths from eight initialization times show that track and CPS clustering meaningfully partition potential tracks and structural evolutions, distilling a large number of ensemble members into several representative and distinct solutions. Rand index and adjusted Rand index calculations demonstrate a relationship between track and CPS cluster membership for both 168-h forecasts and 60-h forecast segments, indicating that certain tracks are preferentially associated with certain structural evolutions. These relationships are explained in greater detail using forecasts initialized at 0000 UTC 25 October. Storm-centered cluster composite maps of 500-hPa geopotential height and 850-hPa equivalent potential temperature for the 120-h forecast valid at 0000 UTC 30 October (initialized at 0000 UTC 25 October) indicate that both track and CPS clustering successfully capture variations in the Sandy-trough interaction and the strength of the lower-troposphere warm core of Sandy at the time of observed landfall. Together, these results illustrate the relationship between the track and structural evolution of Sandy and suggest the potential of multiensemble mixture-model path clustering for tropical cyclone forecasting.