Understanding Predictability of Tropical Cyclones over the North Atlantic Ocean

Tropical cyclones (TCs) are extreme weather events capable of causing major loss of life and property. Despite continued improvements in our understanding of TC dynamics and the advancement of numerical models over the past few decades, accurate and precise forecasting of TCs beyond the short-term (<=72 hours) window remains an operational challenge. In other words, there seems to be a ‘predictability barrier’ for TCs at several days in lead time. For example, according to the US National Hurricane Centre (NHC), the average error for a 5-day track forecast of Atlantic tropical storms exceeds 150 nmi for seven out of the last ten years, with arguably little change in trend. The predictability of TCs is thus an interesting and practically relevant topic for investigation.

We evaluate TC-relevant ensemble statistics with lead time of up to 8 days before genesis for the Atlantic basin, starting from 2017. ECMWF operational ensemble forecast data are used on NHC best-tracked tropical and subtropical systems. TC-like vortices are first tracked using the feature-tracking algorithm TRACK based on lower-tropospheric relative vorticity. They are then matched against the respective TCs from the operational analysis data by a technique known as Dynamic Time Warping, which allows for some spatio-temporal discrepancies. Two main groups of TCs, i.e., those that form in the Main Development Region (MDR) and those that undergo tropical transition (TT) are selected for analysis due to their distinctly different genesis characteristics. Special attention is drawn to the phenomenon of ‘forecast jumps’ (i.e., sudden and substantial shifts in the ensemble probabilities with lead time) as identified in previous work on this topic. The evaluation parameters used include minimum central MSLP, maximum 10m mean wind, storm location and various parameters on tropospheric thermal structures (for storms with TT).

Preliminary work on two case studies has produced some interesting results and offers early insights to studying TC-ensemble behaviour. Ensemble plots for Ophelia, a TT storm, reveal a notable instance of forecast jump around 3 to 4 days before its TT occurrence when multiple parameters display large statistical changes, but only after track-matching ensemble members double in number. For the other case storm, Harvey, the temporal evolution of its ensemble forecast is examined by successive plots with respect to its MDR genesis. A prevalent signal is found in the ensemble sets where some members predict a much stronger system at TC genesis, related to a particular model initialisation 3 days prior. After the actual genesis event, however, ensemble spread is significantly reduced. When looking at the seasons 2017 and 2018, TC genesis is generally well predicted by ensemble members within 4 days. While some of the ‘classical’ TCs that form in the deep tropics offer a relatively long window of predictability, perhaps to more than a week, others are poorly forecast up until genesis, suggesting high variability in their predictive skill. In the subtropical and mid-latitudes, storms are sometimes only captured by the model a few days prior to their TT event. Ensemble members can struggle to distinguish between a warm or cold-core system, before converging on a solution rather abruptly. Additional results will be presented as further work is done to seek physical understanding of forecast jumps from analysis of environmental fields, as well as to establish behavioural statistics while enlarging the TC sample set.