NMME-based hybrid prediction of Atlantic hurricane season activity

Abstract

A hybrid dynamical–statistical model is pursued for prediction of Atlantic seasonal hurricane activity driven by output of the North American Multimodel Ensemble (NMME). This is an updated version of a proven multiple linear regression method conditioned on forecast vertical wind shear from the Climate Forecast System and observed sea surface temperatures (SSTs). The method pursued for prediction utilizes August–October (ASO) Main Development Region (MDR; 10–20°N, 20–80°W) vertical wind shear and observed North Atlantic (NATL; 55–65°N, 30–60°W) SST averaged over the 3 months preceding the forecast in conjunction with the full hurricane climatology. NMME forecasts improve upon representations relative to individual members. The NMME multi-model mean better reproduces vertical wind shear distributions over the MDR and captures the observed relationships between SST and vertical wind shear with hurricane trend and interannual variability despite occasionally poor reproductions by individual members. Cross-validation reveals the multi-model average of the hybrid model outputs from the individual NMME members yields forecast errors 10–30% less than the individual members, while correlations with observed hurricane-related activity typically improve. The NMME methodology is shown to be competitive with official outlooks from Colorado State University and the National Oceanic and Atmospheric Administration over recent years.

Appendix 1: Initial performance of the hybrid NMME forecast for CPC operations

The NMME hybrid model was incorporated as a tool for the Atlantic Hurricane Seasonal Outlook issued by the Climate Prediction Center preceding the 2015 season. This year proved to be an interesting gauge, given building El Niño conditions early in the year, culminating with El Niño declaration by NOAA during March.⁷ Observed JFM SST anomalies for the NATL (Fig. 7) were −0.14 °C relative to 1982–2014, the 16th lowest ranked value relative to the 1982–2014 period. Lead 3 forecasts for ASO wind shear in 2015 (Fig. 12) projected above normal wind shear across the MDR throughout the NMME members (Fig. 12a–c), with the strongest shear on the periphery of the MDR with the exception of CanCM34. CanCM34 forecasted the strongest MDR shear (4.91 m/s) with enhanced shear across much of the MDR, followed by the CCSM4 (2.06 m/s) and CFSv2 (0.46 m/s) that instead portrayed positive shear anomalies for the southwestern MDR and negative shear anomalies in the eastern MDR. The NMME mean shear forecast (Fig. 12d) for the MDR lies in between the CanCM34 and other members at 2.45 m/s, with modest positive anomalies throughout the MDR.

Fig. 12

Forecast anomalous vertical wind shear for ASO 2015 at lead 3 (April initial conditions) relative to 1982–2010 for CanCM34 (a), CCSM4 (b), CFSv2 (c), and multi-model mean (d). Contour intervals are 2 m/s, with the zero contour omitted. MDR averages are listed in the title of eachpanel

Climate conditions further suggested reduced 2015 hurricane activity at lead 0 during 2015. Observed NATL SST for AMJ (Fig. 7) was −0.58 °C relative to 1982–2014 for the 7th lowest ranked year. By early July, many ENSO predictions were indicating the potential for a major El Niño event to be established by ASO,⁸ implying potential severe reduction in hurricane activity. NMME forecasts of lead 0 ASO vertical wind shear are shown in Fig. 13. Apparent are the extreme shear forecasts for the tropical Atlantic by CanCM34 (Fig. 13a), a value exceeding four standard deviations above normal and far larger than any shear magnitude forecast at this lead in the hindcast period. These strong shear projections were present in both CanCM3 and CanCM4 as they forecast +4 and +5 standard deviation events relative to their individual climatologies. While the extent of the CanCM34 positive shear anomalies extend across the entirety of the MDR, the greatest values generally are found in the southern Caribbean Sea. The remaining NMME members (Fig. 13b, c) are less extreme in their ASO shear projections, with CCSM4 exhibiting above normal shear from the northwest to southeast across the MDR with a mean magnitude of 3.40 m/s, while the CFSv2 forecasts near normal MDR shear (0.02 m/s) with negative values off the African coastline through 40°W. Strong positive shear anomalies are seemingly associated with the developing El Niño by CFSv2, however they are predominantly constrained to the East Pacific (peaking >18 m/s) and south of the MDR across the Atlantic basin. Averaging the three members yields the NMME mean in Fig. 13d with 4.00 m/s of shear across the MDR with this number strongly influenced by the CanCM34 projections both in magnitude and spatial distribution with positive anomalies extending across the full entirety of the MDR despite the weakness in the eastern Atlantic from the CCSM4 and CFSv2.

Fig. 13

As in Fig. 12, but at lead 0 (July initial conditions)

NMME forecasts for the 2015 season in Table 3 for hurricane, tropical storm, and ACE activity were included as part of guidance towards development of the 2015 NOAA HSO. The 2015 season yielded 4 hurricanes, 11 tropical storms, and an ACE value of 59.3 × 10⁴ kt² (64.2% of median). NMME predicted ranges capture the observed hurricane activity (Table 3) at both leads for hurricanes, with the best performing member being CCSM4, which captured the observed activity within its forecast range at both leads. Tropical storm activity was also reasonably well captured by NMME, with forecast ranges only failing to match observations for lead 0, but outperformed by the CFSv2 which saw observed activity fall within its predicted ranges at each lead. The failure by the NMME at lead 0 is largely attributable to the CanCM34 which projected an average of two tropical storms for the season, despite three tropical storms developing prior to the HSO update release in August 2015. NMME best captured ACE during 2015, with forecast ranges aligning with observations for both leads. Also noteworthy with ACE forecasts is that the CFSv2 failed to reproduce the observed ACE range for any lead, underscoring the original hybrid method conditioned solely on the CFS would have failed in its ACE projections for 2015. The 2015 hurricane season was somewhat unusual in that despite the near-normal tropical storm activity there were reduced hurricane and ACE activity. Nevertheless, the NMME prediction was able to reproduce the observed distributions for five of the six analyzed forecast combinations of forecast lead and predictands. This underscores the ability of the multi-model mean to aggregate diverse forecasts into an improved product that is not restricted to any singular model with its individual strengths, weaknesses, and biases.

Table 3 Expected value and range (in parentheses) of predictions for 2015 North Atlantic hurricane, tropical storm, and ACE (as a percent of median) activity from the NMME models and ensemble mean