SOLA, 2018, Vol. 14, 203−209, doi:10.2151/sola.2018-036 203 Analysis of Uncertainties in Forecasts of Typhoon Soudelor (2015) from Ensemble Prediction Models Deqiang Liu1, 2, 3, Xubin Zhang4, Yerong Feng4, Ning Pan1, and Chuanrong Huang5 1Fujian Meteorological Observatory, Fuzhou, China 2Laboratory of Straits Meteorology, Xiamen Meteorological Bureau, Xiamen, China 3Wuyishan National Park Meteorological Observatory, Wuyishan, China 4Institute of Tropical and Marine Meteorology, Guangdong Provincial Key Laboratory of Regional Numerical Weather Prediction, CMA, Guangzhou, China 5Fujian Meteorological Service, Fuzhou, China by Bassill (2014), who demonstrated that the choice of cumulus Abstract parameterization led to TCT differences between two operational models of the European Center for Medium-Range Weather Fore- Using data from nine ensemble prediction systems (EPSs), casting (ECMWF). Through an assessment of TIGGE data for we analyze uncertainties in forecasted tropical cyclone TC track hurricane Sandy (2012), Magnusson et al. (20140 demonstrated (TCT), TC intensity (TCI) and relevant heavy rainfall (TCHR) that higher-resolution global models performed better than coarse for Typhoon Soudelor (2015) as it affected the Taiwan Strait and global models in their prediction of tropical cyclone intensity surrounding regions. The largest uncertainties in track predictions (TCI). occurred when Soudelor traversed Taiwan and when it recurved Uncertainty in tropical cyclone prediction can also be northeastward after making landfall in mainland China. These attributed to the shortcomings of forecast models, e.g., a lack of large uncertainties seem to be ascribed to the topography of sufficient observations over the ocean, differences in the method- Taiwan and the spread of the perturbed steering flows, respectively. ology of the tracking algorithm, the ineffective data assimilation TCI spread was stronger before rather than after the Soudelor techniques, inadequate model representation of air-sea interaction, made landfall, with regional EPSs having stronger spread than and inproper description of the topography of Taiwan in numerical global EPSs. This TCI spread showed high correlation with the models (Braun 2002; Buizza et al. 2005; Ding and Li 2008; Zhang evolution of the spread of vertical wind shear at the location of and Luo 2010; Ding and Li 2012; Kunii and Miyoshi 2012; Duan TC center. Large spread in 24-h TCHR during Soudelor’s landfall et al. 2013; Li et al. 2013; Liu et al. 2015; Mu et al. 2015). correlated with low-level jets and convergences in most EPSs, and Large uncertainties in the forecasts of TCs that struck the TC track variation had played important role in TCHR uncertainty. Taiwan region are always associated with the factors mentioned At last, the spread–skill relationships among different groups are above. Therefore, challenges still exist in the enhanced use of explored. ensemble forecasts over the region. An evaluation of uncertainty (Citation: Liu, D., X. Zhang, Y. Feng, N. Pan, and C. Huang, in the TC forecasts for the region would improve our understand- 2018: Analysis of uncertainties in forecasts of Typhoon Soudelor ing of TC uncertainties and provide reference information for the (2015) from ensemble prediction models. SOLA, 14, 203−209, enhanced use of ensembles. doi:10.2151/sola.2018-036.) In this study, we analyze the common uncertainties in fore- casts of Typhoon Soudelor (2015) in nine ensemble prediction systems (EPSs) from several major operational centers. The 1. Introduction possible reasons of these uncertainties are also investigated. As TCT, TCI, and TC-related heavy rainfall (TCHR) are the most Tropical cyclones (TCs) are one of the major natural disasters important forecast parameters in TC prediction (Fourrié et al. affecting the Taiwan Strait, bringing great threat to human lives 2006), these parameters form the basis of our investigation. and property (Mu et al. 2009; Kim et al. 2010; Lang et al. 2012; Wu et al. 2013; Chan et al. 2014; Chen and Wu 2015). Although significant improvement has been achieved during the past decade 2. Typhoon, datasets, and methodology in predicting TC track (TCT), there remain large uncertainties in operational forecasts for TCs over Taiwan and the surrounding Overviews of the typhoon, datasets, and methodology used region, especially for TC intensity (TCI) and TC-related precipi- in this study are given in Text S1. We analyze uncertainties in tation (Harnisch and Weissmann 2010; Kim et al. 2010; Wu et al. Soudelor (2015) for the period when it caused greatest damage to 2013). the region (1200 UTC 7 to 0600 UTC 11 August 2015), by aver- Many studies have focused on the uncertainty in TC ensem- aging ensemble forecasts initialized at 1200 UTC 6, 0000 UTC 7, ble forecasts. Yamaguchi and Majumdar (2010) compared the and 1200 UTC 7 August 2015. The position of TC center was ensemble spread of TC track (TCT) predictions using data of identified as the location of minimum sea-level pressure. The Observing System Research and Predictability Experiment (THORPEX) Interactive Grand Global Ensemble (TIGGE, Bougeault et al. 2010). The results show that the spread of TCT 3. Results forecasts, which varied among ensemble models, was strongly correlated with initial perturbations. Yamaguchi et al. (2012), who 3.1 Uncertainty in typhoon track forecasts also focused on global ensemble models, found that uncertainties Most EPSs provided good forecasts of the general movement in TCT forecasts were not only associated with initial conditions of Soudelor (2015) in the vicinity of the Taiwan Strait. TCT fore- but also model uncertainties. Similar conclusions were reached cast error varied among the nine models, ranging from 20 to 100 km at 1200 UTC 7 August 2015 with an average of 53 km, and up to 180 km four days later (Table 1). Corresponding author: Yerong Feng, Institute of Tropical and Marine Meteorology, Guangdong Provincial Key Laboratory of Regional Numer- Figures 1 and S1a show the spatiotemporal characteristics of ical Weather Prediction, CMA, Guangzhou 510641, China. E-mail: the spread in TCT predictions. In general, the spread tended to [email protected]. ©The Author(s) 2018. This is an open access article published by the Meteorological Society of Japan under a Creative Commons Attribution 4.0 International (CC BY 4.0) license (http://creativecommons.org/license/by/4.0). 204 Liu et al., Uncertainties of Soudelor (2015) through Ensemble Models Table 1. The mean of the grand ensemble spread and forecast bias for TCT and TCI from nine EPSs at T + ∆T, where T is defined as 1200 UTC on 7 August 2015. ∆T (h) 0 6 12 18 24 30 36 42 48 54 60 66 72 78 84 90 Mean of grand ensemble 22.6 33.4 39.3 37.6 30.6 34.8 50.4 54.6 51.5 53.5 55.4 68.7 74.0 60.8 60.8 61.8 TCT spread (km) Mean of grand ensemble 52.9 81.8 76.8 77.1 43.1 71.8 100.6 127.7 110.4 85.5 136.0 171.9 132.9 157.2 187.6 180.7 TCT forecast bias (km) Mean of grand ensemble 2.7 3.4 3.3 2.8 2.9 3.0 2.7 2.3 2.0 1.9 1.8 1.8 1.9 2.0 2.0 2.1 TCI spread (hPa) Mean of grand ensemble 18.5 24.2 14.4 6.3 0.4 −4.0 −7.0 −7.5 −6.3 −3.9 −5.7 −4.9 −3.7 −2.9 −0.1 −0.5 TCI forecast bias (hPa) Fig. 1. Spatial distribution of the spread (km) in tropical cyclone track predictions from nine ensemble prediction systems, averaged by the ensemble fore- casts initialized at 1200 UTC 6, 0000 UTC 7, and 1200 UTC 7 August 2015, for the period when Soudelor traversed the Taiwan Strait (1200 UTC 7 to 0600 UTC 11 August 2015). The black line represents the best track, and the spread is denoted by colored dots, with dot size indicating the magnitude of the spread. increase as Soudelor moved from ocean to land, reaching its max- PSF is known to have an important role in TCT recurvature. PKE imum when the typhoon was deflected northward after making can be used as the crucial predictor of TC recurvature, for the high landfall, indicating that large uncertainty in TCT predictions often correlation between PKE and PSF. The increasing of PKE would appears during TC recurvature. This certainty seems to be closely have greatly influenced the prediction of TC motions and resulted related to perturbed kinetic energy (PKE) in the upper troposphere in large TCT uncertainty. (Fig. 2). It is clear that PKE in all models tended to increase with Most of the models showed large spread in track prediction lead time, implying the growth of the magnitude of wind pertur- after the typhoon passed over Taiwan’s Central Mountain range bations. The evolution of the spread of perturbed steering flows (CMR) (Figs. 1 and S1a; Table 1). TCs similar to Soudelor have (PSF) in the troposphere (200−850 hPa) also increased markedly been observed to deflect slightly southwest, (relative to their with lead time and showed high relationship with PKE (Fig. S2). northwestward movement along the steering flows southwest SOLA, 2018, Vol. 14, 203−209, doi:10.2151/sola.2018-036 205 Fig. 2. Temporal evolution of averaged 250-hPa perturbed kinetic energy over the area of 15°N−40°N, 110°E−130°E for nine models. The results are calcu- lated by averaging the ensemble forecasts initialized at 1200 UTC 6, 0000 UTC 7, and 1200 UTC 7 August 2015. of the subtropical high) after they pass over the CMR. This is (Buizza et al. 2005; Fortin et al. 2014; Feng et al. 2016). For TCT followed by a northwest path prior to landfall, in China, resulting grand ensemble forecasts in all EPSs, the mean of grand ensemble in a “V-like” track in the Taiwan Strait (Wu et al.